Vous êtes sur la page 1sur 229

Preface to Perspectives

In 2004 the AIChE Journal celebrated its 50th anniversary. In connection with this milestone, a symposium on The Future of Chemical
Engineering Research was organized at the Annual AIChE Meeting in Austin. The symposium consisted of three sessions entitled:
Biological Engineering
Complex Systems
In each session seven talks by leading chemical engineers were followed by a three-member panel discussion. A one-person
overview that integrated the talks and discussions within the context of our disciplines research tradition closed the proceedings.
The symposium proved an excellent opportunity for taking a broad look at the current landscape of chemical engineering
research, and for thinking about its future. It is clearly impossible to organize three sessions that can provide a comprehensive
representation of every important aspect of chemical engineering research. Nevertheless, we believe that the scope and quality of
the talks were such that it is of value to our profession to preserve a written record of the symposium. We are delighted that the
AIChE Journal will be the medium for its publication.
This issue of the AIChE Journal features the third collection of Perspective articles dedicated to the 2004 Symposium The
Future of Chemical Engineering Research. The three articles in this issue correspond to the session on Biological Engineering.
The session on Complex Systems appeared in July, and the series of Fundamentals Perspectives appeared in the September issue.
We are grateful to the many speakers who were able to contribute to this collection by putting their talks and thoughts on paper.
We feel honored and privileged to have organized the symposium and to edit this collection of papers. The program of the session
on Biological Engineering is reproduced below.

Arup K. Chakraborty, Pablo G. Debenedetti, Julio M. Ottino

The Future of Chemical Engineering Research
November 8, 2004
AIChE Annual Meeting, AUSTIN, TX
Arup Chakraborty, Chair; University of California-Berkeley
Pablo Debenedetti, co-Chair; Princeton University
Materials For Drug Delivery
Mark E. Davis, California Institute of Technology
Metabolic Engineering of Bacteria for Drug Production
Jay D. Keasling, University of California-Berkeley
Entropy, Disease, and New Opportunities for Chemical Engineering Research
Michael W. Deem, Rice University
Protein Engineering and Biopharmaceutical Design
K. Dane Wittrup, M.I.T.
Biomaterials in Regenerative Medicine: Future Thrusts
Kristi Anseth, University of Colorado-Boulder
Developing Pharmaceutical Products
Mauricio Futran, Bristol, Myers, Squibb
Intercellular Communication in the Adaptive Immune System: Plenty of Room at the Bottom
Arup Chakraborty, University of California-Berkeley
Summary and Discussion
George Georgiou (University of Texas, Austin)
Daniel A. Hammer (University of Pennsylvania)
Gregory Stephanopoulos (MIT)

AIChE Journal, Vol. 51, 3082 (2005)

2005 American Institute of Chemical Engineers
DOI 10.1002/aic.10728
Published online October 28, 2005 in Wiley InterScience (www.interscience.


December 2005

Vol. 51, No. 12

AIChE Journal



Directed Evolution in Chemical Engineering
K. Dane Wittrup
Dept. of Chemical Engineering and Division of Biological Engineering, Massachusetts Institute of Technology,
MIT 66-552, Cambridge, MA 02139

Entropy, Disease, and New Opportunities for

Chemical Engineering Research
Michael W. Deem
Rice University, 6100 Main Street-MS 142, Houston, TX 77005

Metabolic Engineering: Developing New

Products and Processes by Constructing
Functioning Biosynthetic Pathways in vivo
Gregory Stephanopoulos and Kyle L. Jensen
Massachusetts Institute of Technology, Dept. of Chemical Engineering 77 Massachusetts Avenue, Cambridge, MA 02137

AIChE Journal

December 2005

Vol. 51, No. 12



Directed Evolution in Chemical Engineering

K. Dane Wittrup
Dept. of Chemical Engineering and Division of Biological Engineering, Massachusetts Institute of Technology,
MIT 66-552, Cambridge, MA 02139
DOI 10.1002/aic.10706
Published online October 4, 2005 in Wiley InterScience (www.interscience.wiley.com).

Keywords: directed evolution, protein engineering, molecular bioengineering

Predictions are hard to make, especially about the future. Niels Bohr, and Yogi Berra

A role for prediction?

op-down exercises in research area prognostication often miss the mark. The most general predictions can be
platitudinously unfalsiable; the most specic are often
amusingly wrong in retrospect. Given the primacy of investigator-initiated efforts in the successful denition and exploitation of new research areas, it is fair to question the utility of
the future of. . . pieces, such as this one.
I well recall sitting through keynote presentations by eminent
senior colleagues in the biochemical engineering eld at national
meetings in the late 1980s and early 1990s, the tenor of which was
that the essential premise of my research program was an enormous mistake, straying too far from the hard ChE core. A particularly memorable assertion was that It is better to be a rst-rate
chemical engineer than a second-rate biologist. (Fortunately
these do not appear to be the only two available options.) Editorializing is an ineffective means of squelching initiative, by comparison to the ruthless efciency of the existing marketplace of
ideas. Successful research directions are determined, over a sufcient time span, by peer-reviewed funding and publication, the
interest of new students, and the willingness of academia and
industry to hire students trained to perform such research. If
exercises such as this one are capable of serving a useful purpose,
perhaps it is to constructively recognize incipient grass-roots
movements and highlight exciting challenges.

Synthesis in Chemical Engineering

A shift from process to product engineering within ChE has
been noted previously, and has been embraced systemically in
U.S. academic departments. Synthetic capabilities are an essential tool for invention, as evidenced by numerous accomplishments of ChE researchers applying and developing synthetic tools in the elds of electronic materials, polymers,
MEMS devices, and drug delivery. A new opportunity for ChE
contributions beckons with the explosive growth in the development of protein biopharmaceuticals. Proteins play a central
K. D. Wittrups e-mail address is wittrup@mit.edu.

2005 American Institute of Chemical Engineers


December 2005

role in biological function and many pathologies. The therapeutic efcacy of a protein drug is inextricably linked to its
binding properties, and consequently tools for engineering protein binding are indispensable, as well as analytical tools to
properly determine biophysical design criteria.1
At present the most powerful and robust approach to engineering protein properties is Directed Evolution, and the eld
of chemical engineering has quietly gone about becoming the
predominant academic home for Directed Evolution research.
A cursory examination of departmental web sites reveals over
20 ChE faculty in the U.S. (most of whom began their careers
in the past 15 years) whose research programs are extensively
dedicated to developing or using Directed Evolution. Combinatorial polypeptide library screening consists of a series of
equilibrium and/or transient biochemical transformations that
are well suited to analysis with classical chemical reaction
engineering tools.2-5 Widely-practiced innovations in protein
screening methodology have originated from ChE research
groups.6-8 The fundamental structure of the tness landscape,
and improved search methods for it, has been a fruitful source
for interesting research problems.9-14 A growing direction is the
use of chemical engineering analyses to guide the development
of optimized protein biopharmaceuticals.1, 15-19 In terms of
analytical contributions, cellular signaling pathways,20, 21 metabolism,22, 23 and immunology24, 25 hold tremendous promise
for chemical engineers, and these approaches are likely to
contribute to rationalization of protein drug pharmacology.

Curricular development in chemical and biological

There has been a signicant level of angst in the ChE community with respect to the Whitaker-driven expansion of academic
biomedical engineering programs, with concomitant reductions in
ChE undergraduate enrollments. However, allowing the ChE curriculum to be driven by the actions of external communities is
existentially unsatisfying and unlikely to be a recipe for innovation.
At the risk of Pollyanna optimism, the ChE discipline is in a
very strong position on the ground with respect to biomolecular
engineering. If the ChE analytical toolkit did not exist, it would
be necessary to invent it in order to solve many of the problems
in modern bioengineering. Critical foundations of biomolecular
processes are the rates of biochemical conversions and their
Vol. 51, No. 12

AIChE Journal

equilibria, and the rates of biomolecular movement by diffusion and convection. The ChE triumvirate core of kinetics,
thermo, and transport are the required tools for such analyses.
Although BME curricula generally include transport coursework, BME training programs in general do not incorporate
kinetics and chemistry to the extent necessary for modern
biomolecular engineering.
ChE should not rest on its laurels, however; incorporating
biology into the ChE curriculum is not so simple as adding a
few new examples or homework problems, because there are
new intellectual principles from biochemistry, biophysics, and
cell biology that must be integrated throughout the toolkit. The
most effective means for syllabus evolution will be to staff
courses with champions dedicated to change; given the large
numbers of young ChE faculty with biological interests now
disseminated throughout the U.S., this is already occurring.
These efforts will be amplied by the emergence of new
textbooks incorporating biological principles in an integrated
fashion rather than grafted onto existing outlines.

Precis and outlook

There are an extraordinary number of signicant opportunities in biomedicine to engineer proteins by directed evolution.1,
26-28 Many chemical engineers have recognized this opportunity and are exploiting it with vigor, making chemical engineering the academic center of mass for the burgeoning eld of
Directed Evolution.

Literature Cited
1. Rao BM, Lauffenburger DA, Wittrup KD. Integrating celllevel kinetic modeling into the design of engineered protein therapeutics. Nat Biotechnol. 2005;23(2):191-4.
2. Maheshri N, Schaffer DV. Computational and experimental analysis of DNA shufing. Proc Natl Acad Sci U S A.
3. Daugherty PS, Olsen MJ, Iverson BL, Georgiou G. Development of an optimized expression system for the screening of antibody libraries displayed on the Escherichia coli
surface. Protein Eng. 1999;12(7):613-21.
4. Boder ET, Wittrup KD. Optimal screening of surface-displayed polypeptide libraries. Biotechnol Prog. 1998;14(1):5562.
5. Moore GL, Maranas CD. Modeling DNA mutation and
recombination for directed evolution experiments. J Theor
Biol. 2000;205(3):483-503.
6. Zhao H, Giver L, Shao Z, Affholter JA, Arnold FH. Molecular evolution by staggered extension process (StEP) in vitro
recombination. Nat Biotechnol. 1998;16(3):258-61.
7. Francisco JA, Earhart CF, Georgiou G. Transport and anchoring of beta-lactamase to the external surface of Escherichia
coli. Proc Natl Acad Sci U S A. 1992;89(7):2713-7.
8. Boder ET, Wittrup KD. Yeast surface display for screening combinatorial polypeptide libraries. Nat Biotechnol.
9. Voigt CA, Mayo SL, Arnold FH, Wang ZG. Computational
method to reduce the search space for directed protein evolution. Proc Natl Acad Sci U S A. 2001;98(7):3778-83.
10. Moore GL, Maranas CD. Identifying residue-residue clashes
in protein hybrids by using a second-order mean-eld approach. Proc Natl Acad Sci U S A. 2003;100(9):5091-6.
AIChE Journal

December 2005

11. Bogarad LD, Deem MW. A hierarchical approach to protein molecular evolution. Proc Natl Acad Sci U S A.
12. Earl DJ, Deem MW. Evolvability is a selectable trait. Proc
Natl Acad Sci U S A. 2004;101(32):11531-6.
13. Daugherty PS, Chen G, Iverson BL, Georgiou G. Quantitative analysis of the effect of the mutation frequency on
the afnity maturation of single chain Fv antibodies. Proc
Natl Acad Sci U S A. 2000;97(5):2029-34.
14. Drummond DA, Iverson BL, Georgiou G, Arnold FH.
Why high-error-rate random mutagenesis libraries are enriched in functional and improved proteins. J Mol Biol.
15. Graff CP, Wittrup KD. Theoretical analysis of antibody
targeting of tumor spheroids: importance of dosage for
penetration, and afnity for retention. Cancer Res. 2003;
16. Sarkar CA, Lauffenburger DA. Cell-level pharmacokinetic
model of granulocyte colony-stimulating factor: implications for ligand lifetime and potency in vivo. Mol Pharmacol. 2003;63(1):147-58.
17. Rao BM, Driver I, Lauffenburger DA, Wittrup KD. HighAfnity CD25-Binding IL-2 Mutants Potently Stimulate Persistent T Cell Growth. Biochemistry. 2005;44(31):10696-701.
18. Haugh JM. Mathematical model of human growth hormone (hGH)-stimulated cell proliferation explains the efcacy of hGH variants as receptor agonists or antagonists.
Biotechnol Prog. 2004;20(5): 1337-44.
19. Kim JR. Gibson TJ, Murphy RM. Targeted control of kinetics
of beta-amyloid self-association by surface tension-modifying peptides. J Biol Chem. 2003;278(42):40730-5.
20. Kao KC, Yang YL, Boscolo R, Sabatti C, Roychowdhury
V, Liao JC. Transcriptome-based determination of multiple transcription regulator activities in Escherichia coli by
using network component analysis. Proc Natl Acad Sci U
S A. 2004;101(2):641-6.
21. Sachs K, Perez O, Peer D, Lauffenburger DA, Nolan GP.
Causal protein-signaling networks derived from multiparameter single-cell data. Science. 2005;308(5721):523-9.
22. Price ND, Reed JL, Palsson BO. Genome-scale models of
microbial cells: evaluating the consequences of constraints. Nat Rev Microbiol. 2004;2(11):886-97.
23. Alper H, Miyaoku K, Stephanopoulos G. Construction of
lycopene-overproducing E. coli strains by combining systematic and combinatorial gene knockout targets. Nat Biotechnol. 2005;23(5):612-6.
24. Chakraborty AK. Decoding communications between cells
in the immune system using principles of chemical engineering. AIChE J. 2003;49(7):1614-1620.
25. Munoz ET, Deem MW. Epitope analysis for inuenza
vaccine design. Vaccine. 2005;23(9):1144-8.
26. Brekke OH, Sandlie I. Therapeutic antibodies for human
diseases at the dawn of the twenty-rst century. Nat Rev
Drug Discov. 2003;2(1):52-62.
27. Vasserot AP, Dickinson CD, Tang Y, Huse WD, Manchester
KS, Watkins JD. Optimization of protein therapeutics by
directed evolution. Drug Discov Today. 2003;8(3):118-26.
28. Lazar GA, Marshall SA, Plecs JJ, Mayo SL, Desjarlais JR.
Designing proteins for therapeutic applications. Curr Opin
Struct Biol. 2003;13(4):513-8.
Vol. 51, No. 12



Entropy, Disease, and New Opportunities for

Chemical Engineering Research
Michael W. Deem
Rice University, 6100 Main Street-MS 142, Houston, TX 77005
DOI 10.1002/aic.10718
Published online October 24, 2005 in Wiley InterScience (www.interscience.wiley.com).

Keywords: inuenza, bird u, cancer, autoimmune disease

Executive Summary

andomness, diversity, uctuations, and correlations

play a signicant role in disease, disease treatment, and
the immune response to disease. I suggest that statistical
mechanics, the physical theory of randomness that uses a
systems physical behavior at the molecular or atomic scale to
synthesize a picture of the behavior at a larger level, can
address some of these issues. Theory and mathematical modeling can help design or redesign treatment strategies. Theory
can also help determine what makes redesign necessary. I
describe three lines of research in my group seeking this new
mathematics of biology.

Efcacy of the Inuenza Vaccine The

Response of the Immune System to a
Inuenza epidemics are annually responsible for the deaths
of 25,0000 to 50,0000 people in the world and cause illness in
5 to 15% of the total population each year (World Health
Organization, 2003). The total cost associated with inuenza in
the U.S. is roughly $10 billion (Lave et al., 1999), and the
economic cost of an inuenza pandemic has been estimated to
be $71167 billion (Meltzer et al., 1999) in the U.S. alone. The
primary method employed to prevent infection by inuenza
and its associated complications is vaccination. Mutation and
antigenic change, combined with the high rate of transmission
of inuenza strains, means that the vaccine must be redesigned
each year. This is currently done with phylogenetic, animal
model, and epidemiological analysis.
The effectiveness of the inuenza vaccine varies each year
due to changes in the molecular structure of the inuenza
strains that are circulating. Three strains are customarily included in the annual inuenza vaccine, with these three strains
chosen to be as similar as possible to those estimated to be the
M. W. Deems e-mail address is mwdeem@rice.edu.

2005 American Institute of Chemical Engineers


December 2005

most widespread circulating strains in the upcoming u season.

Currently, the vaccine contains a H3N2 and a H1N1 inuenza
A component and an inuenza B component. Due to mutation
of the inuenza virus, vaccine efcacy is rarely 100%, and is
more typically 30 60%, against inuenza-like illness. The
estimated worldwide mortality rises by another 160%260% if
inuenza-induced complications in patients with other conditions are included (Neuzil et al., 1999; Sprenger et al., 1993).
It is believed that the inuenza vaccine, on average, signicantly reduces such excess mortality (Hak et al., 2002). Vaccine efcacy can be negative, however, due to original antigenic sin (Davenport et al., 1953; Fazekas de St. Groth and
Webster, 1966; Deem and Lee, 2003), which is the tendency
for antibodies produced in response to exposure to one inuenza vaccine antigen to suppress the creation of new, different
antibodies in response to exposure to a new version of the
inuenza virus. The efcacy of the annual inuenza vaccine,
and whether original antigenic sin occurs, depends delicately
on the similarity between the vaccine and circulating viral
strains. The current standard of practice is to measure antigenic
distance by ferret antisera hemagglutinin inhibition assays
(Smith et al., 2004; Smith et al., 1999; Lee and Chen, 2004),
and these distances have been assumed to correlate well with
vaccine efcacies in humans. However, to my knowledge no
such signicant correlation has ever been shown for an experimental, animal model, or theoretical measure of antigenic
distance. Besides being useful for the annual u shot design, a
reliable measure of antigenic distance would help to stem the
spread of a newly emerged inuenza strain by allowing for
streamlined decision making if preparation and rush production
of a modied vaccine is necessary (Ault, 2003).
So, what is the best order parameter to describe antigenic
distance, and which also correlates well with vaccine efcacy?
Using the tools of statistical mechanics, we provided a quantitative denition of the difference between the dominant
epitope regions in the vaccine and circulating strain, ppeptide
(Deem and Lee, 2003; Gupta et al., 2005). We showed that this
denition of antigenic distance correlates well with human
inuenza vaccine efcacy over the past 35 years (see Figure 1)
(Munoz and Deem, 2005; Gupta et al., 2005).
Vol. 51, No. 12

AIChE Journal

Figure 1. Vaccine efcacy for inuenza-like illness as a

function of pepitope as observed in epidemiological studies and as predicted by theory.
Also shown is a linear least squares t to the data (long
dashed, R2 0:81). From (Gupta et al., 2005).

A different and highly lethal strain of inuenza, the H5N1 avian

inuenza strain, was rst detected in humans in Hong Kong in
1997 (Saw et al., 1998; Claas et al., 1998). Since then, it has
spread to at least eight other Asian countries (Normile and
Enserink, 2005; Cyranoski, 2005a), and Russia (Allakhverdov and
Enserink, 2005), and it is widely expected to enter the rest of
Europe through migrating birds. To date there have been roughly
60 reported deaths due to the H5N1 strain. The initial mortality of
70% decreased to roughly 20% (Normile, 2005b), which suggests
the bird u is evolving to become less fatal but concurrently more
able to persist and, thus, to create an epidemic. Avian inuenza
has also been observed in pigs, a classic mixing vessel for inuenza (Cyranoski, 2005b). Person-to-person transmission has been
suggested (Ungchusak et al., 2005). Avian inuenza is, thus,
evolving (Normile, 2005a; Hulse-Post et al., 2005). Vaccines,
antivirals, animal culling, and public health measures (Longini et
al., 2005; Ferguson et al., 2005) are the main weapons against
spread of the bird u. Various countries are stockpiling bird u
vaccines, and a vaccine produced from one test strain of the bird
u has produced an immune response in healthy adults (World
Health Organization, 2005). However, since the bird u is mutating, which strains should be stockpiled? The National Institute of
Allergy and Infectious Disease is making sequences of the H5
strains available (Kaiser and Vogel, 2004). The U.S. Centers for
Disease Control and Prevention are, controversially, investigating
the potential reassortments of the bird u that might create an
epidemic (Khamsi, 2005). However, the key question remains:
What is the required or optimal diversity of vaccine stockpile? In
particular: How cross-protective will a bird u vaccine be against
other strains that exist or will evolve into existence? The extent of
cross-protection is needed to determine the optimal vaccines to
stockpile and how to administer them (Patel et al., 2005).
To protect against the strains observed to date, how many
vaccine components are needed? We used results from data
(Gupta et al., 2005) and theory (Deem and Lee, 2003) of the
inuenza vaccine effectiveness to estimate how cross-protective
the bird u vaccine will be (Zhou and Deem, 2005). We used the
AIChE Journal

December 2005

50 strains of H5N1 observed to date (Macken et al., 2001) to

characterize the diversity of a typical pandemic. While vaccines
against typical inuenza A strains contain only the dominant
strain, as subdominant strains typically last only a few seasons
(Fitch et al., 1997), a bird u pandemic may well be over in one
season. There may be signicant mortality from multiple strains,
and so protection against all strains may be important. We considered that the vaccine will contain only those strains that have
been observed in the wild, due to the ethical questions arising from
developing and vaccinating with mutant strains. We examined all
combinations of the wild-type strains, searching for the combination with the least number of components. To cover all these
strains with at least 15% vaccine efcacy, we predicted that 3
vaccine strains would be needed.
Evolution of the H5N1 strain may well render existing vaccine
stockpiles ineffective, and public health authorities may have to
depend almost completely on the production of new vaccine
strains, in addition to antivirals, animal culling, and quarantine.
The value of pepitope can be used to estimate the expected
efcacy of vaccine stockpiles and, thus, to decide whether to
deploy them. Our results suggest that reverse genetics or DNAbased vaccines, both able to rapidly deploy a new vaccine, will be
useful in the prevention of wide-spread infection during an H5N1
epidemic given the slow and unreliable nature of the traditional
vaccine production in hens eggs. Thus, a chemical engineering
approach may contribute not only to the vaccine design, but also
to the production (Khosla, 2002).

Evolution in Antibody Sequence Space and

Autoimmune Disease
The immune system normally protects the human host against
death by infection. The immune system tends to produce antibodies with binding constants of at most 106 - 107 l/mol. Experimentally, however, it is possible to nd binding constants between
antibodies and substrates on the order of 1011 1013 l/mol
(Schier et al., 1996), and the laboratory techniques to nd these
antibodies (Maynard and Georgiou, 2000; Swers et al., 2004)
mimic mechanisms that exist within the natural hierarchy of
evolutionary events (Kidwell and Lisch, 2001; Earl and Deem,
2004). The method that the immune system uses to search sequence space is rather slow the same mechanisms that can nd
antibodies with higher afnity can also nd them more quickly.
Thus, one would think that these more powerful evolutionary
mechanisms would give an immune system that responds faster
and more effectively against disease. So, why didnt we evolve
that kind of adaptive response?
To answer this question, we rst sought to understand the
evolutionary rules that govern the way the immune system responds to an infection. With that framework in place, we identied
a biologically-plausible strategy that would allow the immune
system to react more quickly and with more effective antibodies.
Our analysis revealed that such a system would be about 1,000
times more likely to produce antibodies that attack healthy tissues
(See Figure 2) (Sun et al., 2005). Such cross reactivity due to
increased afnity has recently been observed (Holler et al., 2003).
Antibodies that bind with a molecule other than the antigen they
evolved to attack are called cross-reactive, and cross-reactivity can
cause autoimmune disease. For example, chronic infection has
been found to be correlated with increased probability of autoimVol. 51, No. 12


Figure 2. Afnity of memory antibody sequences after a

primary immune response for the two different
immune system strategies (PM and GSSPM)
to altered antigens.
The binding constant is K, and the antigenic distance of the
new altered antigen from the original antigen is p. Crossreactivity ceases at larger distances in the GSSPM case (no
cross-reactivity for p 0:472) than in the PM only case (no
cross-reactivity for p 0:368). Theory shows that these
results imply the antibodies evolved by the GSSPM dynamics will recognize on average 103 more epitopes than the
antibodies evolved by the PM dynamics alone. From (Sun et
al., 2005).

mune disease (Leirisalo-Repo, 2005; Kaplan et al., 1997). However, the strength and signicance of this correlation is controversial (Carty et al., 2003). Our model suggests a broad distribution
for the time of onset of autoimmune disease due to chronic
infection. Researchers have been looking for a clear, signicant
correlation in time, but a long distribution of onset times would
lead to weaker statistical correlations, particularly in those cases
where the infection persisted the longest. Searching for this distribution could elucidate this immunological puzzle and settle the
scientic controversy.
We found that the human immune system evolved to minimize the risk of cross-reactivity. For example, each cell in our
bodies contains about 100,000 proteins with an average of 500
amino acids apiece. Consequently, there are about 1012 potential docking sites, or epitopes, where antibodies could mistakenly attach themselves to proteins in a healthy cell. The mutation response method employed by our adaptive immune
system seems keyed to this number, producing antibodies that
are statistically likely to mistakenly bond with healthy proteins
slightly less than one in 1012 times, meaning that on average,
they recognize only invading pathogens.

Randomness of Cancer
The percentage of Americans dying from cancer is the
same as what it was in 1970. . . and the same as what it was
in 1950 (Leaf, 2004). Although some progress has been
made, especially for childhood cancers, cancer remains a
largely unsolved purge of modern society. Mouse models
remain largely unpredictive, and cancer seems a tremen3088

December 2005

dously complicated disease. Heterogeneity and randomness

are two factors that distinguish cancer from other, more
treatable diseases. New fundamental concepts are desperately needed in the ght against cancer.
I suggest that the tools of statistical mechanics, the physical
theory of randomness, may provide a conceptual framework
for cancer drug and vaccine discovery. The design of therapeutic cancer vaccines to effectively besiege a cancer epitope is
a pivotal piece of the current war on cancer. Indeed, a molecular-level understanding of cancer is vital to develop much
needed diagnostic and therapeutic tools. Thus, the immune
response to cancer vaccines must be dissected at the molecular
biophysics level.
Several features of cancer and the immune system limit the
effectiveness of traditional vaccination techniques (Dunn et al.,
2002; Schreiber et al., 2002; Whelan et al., 2003). In light of
the need to deliver multiple, related vaccines to eradicate a
cancer (Stuge et al., 2004; Whelan et al., 2003; Markiewicz and
Kast, 2004), a multisite vaccination strategy appears promising
(Schreiber et al., 2002). In my group, we developed a quantitative theory that explains the success of the new multisite
approach (see Figure 3), and we are using this theory use to
guide the arduous process of vaccine design (Yang et al.,
2005). By inducing a T cell response to each cancer-associated
epitope in a distinct lymph node, vaccine efcacy is increased,
and immunodominance is reduced. The approach captures the
recognition characteristics between the T-cell receptors
(TCRs), and tumor, the primary dynamics due to TCR resource
competition (Kedl et al., 2003), and the secondary dynamics
due to competition between escape of tumor cells by epitope
mutation and allele loss, and elimination of tumor cells by
TCRs. This approach may be applied to both solid tumors and
post-surgical micrometastases.
Sculpting the diversity of the TCR repertoire is a means to
reduce immune evasion of tumor cells due to epitope mutation or
MHC I allele loss. The TCR diversity provides a recognition
reserve to target the unmutated, subdominant epitopes (NikolichZugich et al., 2004). The importance of diversity, the biological
analog of entropy in this problem, is naturally appreciated within
the context of statistical mechanics.
The solid nature of tumors presents some challenges to immune
control, but also some opportunities for engineering contributions.
It may be difcult, for example, for T cells to enter the solid
tumor. Conversely, high intensity focused ultrasound can disrupt
solid tumors and can enhance systemic antitumor cellular immunity (Wu et al., 2004). Although the exact mechanism of this
enhancement is unknown, one possibility is that the fragments of
tumor after destruction can travel to different lymph nodes and,
thus, induce a diverse TCR repertoire, in similar fashion to polytopic vaccination. More prosaically, the physical disruption of the
tumor can allow easier entry of the T cells. Another feature of
solid tumors is the enhanced probability of uptake of large proteins, due to the highly porous capillaries in tumors (Raucher and
Chilkoti, 2001). By using this property, stimulants of T cell
activity may be localized to the tumors. The concentration of such
stimulants could be increased further by conjugation with elastinlike polypeptides that undergo a thermally triggered phase transition in heated tumors, causing selective aggregation in the tumor
(Raucher and Chilkoti, 2001).
Vol. 51, No. 12

AIChE Journal

Figure 3. Value of the parameter mixing round for vaccination to different lymph nodes at different
distances from the heart.
Humans have several hundred lymph nodes. For effective
polytopic vaccination, well-separated sites on different limbs
are used.

By a discussion of three examples, I hope to have convinced the
reader that signicant unsolved theoretical problems exist in medicine and that statistical mechanics has a pivotal role to play in
their solution. The importance of randomness to the proper functioning of the immune system seems an especially ripe topic for
statistical mechanical analysis. Many, if not most, problems in
immunological diversity remain open, and identication of the
models and theories to tackle these problems is just starting.
Interaction with immunologists and pathologists has proved helpful to my group as we explore these issues. We are also fortunate
as a eld and as a profession that graduate students are keen to
contribute to the new mathematics of biology.

It is a pleasure to acknowledge the group members who
contributed to this work: Enrique T. Munoz, David J. Earl,
Vishal Gupta, Hao Zhou, Jun Sun, Ming Yang, and Jeong-Man
Park. Support by the National Institutes of Health of research in
my group is gratefully acknowledged.

Literature Cited
Allakhverdov, A., and M. Enserink, Bird Flu Hits Russia,
Science, 309, 861 (2005).
Ault, A., Shifting Tactics in the Battle Against Inuenza,
Science, 303, 1280 (2003).
Carty, S. M., N. Snowden, and A. J. Silman, Should Infection
Still be Considered as the Most Likely Triggering Factor for
Rheumatoid Arthritis? J. Rheumatology, 30, 425 (2003).
AIChE Journal

December 2005

Claas, E. C. J., A. D. M. E. Osterhaus, R. van Beek, J. C. D.

Jong, G. F. Rimmelzwaan, D. A. Senne, S. Krauss, K. F.
Shortridge, and R. G. Webster, Human Inuenza A H5N1
Virus Related to a Highly Pathogenic Avian Inuenza Virus, Lancet, 351, 472 (1998).
Cyranoski, D., Bird Flu: Crossing Borders, Nature, 436, 310
Cyranoski, D., Bird Flu Spreads among Javas Pigs, Nature,
435, 390 (2005b).
Davenport, F. M., A. V. Hennessy, and T. Francis, Epidemiologic and Immunologic Signicance of Age Distribution of
Antibody to Antigenic Variants of Inuenza Virus, J. Exp.
Med., 98, 641 (1953).
Deem, M. W., and H.-Y. Lee, Sequence Space Localization in
the Immune System Response to Vaccination and Disease,
Phys. Rev. Lett., 91, 068101 (2003).
Dunn, G. P., A. T. Bruce, H. Ikeda, L. J. Old, and R. D.
Schreiber, Cancer Immunoediting: From Immunosurveillance to Tumor Escape, Nature Immunol., 3, 991 (2002).
Earl, D. J., and M. W. Deem, Evolvability is a Selectable
Trait, Proc. Natl. Acad. Sci. USA, 101, 11531 (2004).
Fazekas de St. Groth, S., and R. G. Webster, Disquisition on
Original Antigenic Sin: Evidence in Man, J. Exp. Med.,
124, 331 (1966).
Ferguson, N. M., D. A. Cummings, S. Cauchemez, C. Fraser,
S. Riley, A. Meeyai, S. Iamsirithaworn, and D. S. Burke,
Strategies for Containing an Emerging Inuenza Pandemic
in Southeast Asia, Nature, 436 (2005), doi: 10.1038/nature04017.
Fitch, W. M., R. M. Bush, C. A. Bender, and N. J. Cox, Long
Term Trends in the Evolution of H(3)HA1 Human Inuenza
Type A, Proc. Natl. Acad. Sci. USA, 94, 7712 (1997).
Gupta, V., D. J. Earl, and M. W. Deem, Quantifying Inuenza
Vaccine Efcacy and Antigenic Distance, (2005),
Hak, E., J. Nordin, F. Wei, J. Mullooly, S. Poblete, R. Strikas,
and K. L. Nichol, Inuence of High-Risk Medical Conditions on the Effectiveness of Inuenza Vaccination among
Elderly Members of 3 Large Managed-Care Organizations,
CID, 35, 370 (2002).
Holler, P. D., L. K. Chlewicki, and D. M. Kranz, TCRs with
High Afnity for Foreign pMHC Show Self-Reactivity,
Nature Im., 4, 55 (2003).
Hulse-Post, D. J., K. M. Sturm-Ramirez, J. Humberd, P. Seiler,
E. A. Govorkova, S. Krauss, C. Scholtissek, P. Puthavathana, C. Buranathai, T. D. Nguyen, H. T. Long, T. S. P.
Naipospos, H. Chen, T. M. Ellis, Y. Guan, J. S. M. Peiris,
and R. G. Webster, Role of Domestic Ducks in the Propagation and Biological Evolution of Highly Pathogenic H5N1
Inuenza Viruses in Asia, Proc. Natl. Acad. Sci. USA, 102,
10682 (2005).
Kaiser, J., and G. Vogel, NIAID Tackles Flu Genomes,
Science, 306, 1275 (2004).
Kaplan, D., I. Ferrari, P. L. Bergami, E. Mahler, G. Levitus,
P. Chiale, J. Hoebeke, M. H. V. Van Regenmortel, and
M. J. Levin, Antibodies to Ribosomal P Proteins of
Trypanosoma cruzi in Chagas Disease Possess Functional
Autoreactivity with Heart Tissue and Differ from anti-P
Autoantibodies in Lupus, Proc. Natl. Acad. Sci. USA, 94,
10301 (1997).
Kedl, R. M., J.W. Kappler, and P. Marrack, Epitope DomiVol. 51, No. 12


nance, Competition and T Cell Afnity Maturation, Curr.

Opin. Immunol., 15, 120 (2003).
Khamsi, R., Deadly Combinations, Nature, 435, 406 (2005).
Khosla, C., Biochemistry-Engineering Interface in Biochemical Engineering, AIChE J., 48, 1366 (2002).
Kidwell, M. G., and D. R. Lisch, Perspective: Transposable
Elements, Parasitic DNA, and Genome Evolution, Evolution, 55, 1 (2001).
Lave, J. R., C. J. Lin, M. J. Fine, and P. Hughes-Cromick, The
Cost of Treating Patients with Community-Acquired Pneumonia, Sem. Respir. Crit. Care. Med, 20, 189 (1999).
Leaf, C., Why Were Losing the War on Cancer [And How to
Win It], Fortune, 76-91 (2004).
Lee, M. S., and J. S. E. Chen, Predicting Antigenic Variants
of Inuenza A/H3N2 Viruses, Emerg. Infect. Dis., 10, 1385
Leirisalo-Repo, M., Early Arthritis and Infection, Curr.
Opin. in Rheumatology, 17, 433 (2005).
Longini, I. M., Jr, A. Nizam, S. Xu, K. Ungchusak,W. Hanshaoworakul, D. A. T. Cummings, and M. E. Halloran,
Containing Pandemic Inuenza at the Source, Science,
309, 1083 (2005).
Macken, C., H. Lu, J. Goodman, and L. Boykin, The Value of
a Database in Surveillance and Vaccine Selection, in Options for the Control of Inuenza IV, A. D. M. E. Osterhaus, N. Cox, and A.W. Hampson, eds., Elsevier Science
(2001), hemagglutinin H3 epitope structural mapping. http://
Markiewicz, M. A., and W. M. Kast, Progess in the Development of Immunotherapy of Cancer Using Ex Vivo-Generated Dendritic Cells Expressing Multiple Tumor Antigen
Epitopes, Canc. Investi., 22, 417 (2004).
Maynard, J., and G. Georgiou, Antibody Engineering, Ann.
Rev. Biomed. Eng., 2, 399 (2000).
Meltzer, M. I., N. J. Cox, and K. Fukuda, The Economic
Impact of Pandemic Inuenza in the United States: Priorities
for Intervention, Emerg. Infect. Dis., 5, 659 (1999).
Munoz, E. T., and M. W. Deem, Epitope Analysis for Inuenza Vaccine Design, Vaccine, 23, 1144 (2005).
Neuzil, K. M., G. W. Reed, J. E. F. Mitchel, and M. R. Grifn,
Inuenza-Associated Morbidity and Mortality in Young
and Middle-Aged Women, JAMA, 281, 901 (1999).
Nikolich-Zugich, J., M. K. Slifka, and I. Messaoudi, The
Many Important Facets of T-cell Repertoire Diversity, Nat.
Rev. Immunol., 4, 123 (2004).
Normile, D., Genetic Analyses Suggest Bird Flu Virus Is
Evolving, Science, 308, 1234 (2005a).
Normile, D., Outbreak in Northern Vietnam Bafes Experts,
Science, 308, 477 (2005b).
Normile, D., and M. Enserink, Avian Inuenza Makes a
Comeback, Reviving Pandemic Worries, Science, 305, 321
Patel, R., I. M. Longine, and M. E. Halloran, Finding Optimal
Vaccination Strategies for Pandemic Inuenza using Genetic
Algorithms, J. Theor. Biol., 234, 201 (2005).
Raucher, D., and A. Chilkoti, Enhance Uptake of a Thermally
Responsive Polypeptide by Tumor Cells in Response to Its


December 2005

Hyperthermia-mediated Phase Transition, Cancer Res., 61,

7163 (2001).
Saw, T. A., W. Lim, K. Shortridge, J. Tam, K. K. Liu, K. H.
Mak, T. Tsang, Y. Y. Ho, F. Y. Lee, and H. Kwong,
Isolation of Avian Inuenza A (H5N1) Viruses from Humans-Hong Kong, May-December 1997, J. Am. Med. Assoc., 279, 263 (1998).
Schier, R., et al., Isolation of Picomolar Afnity Anti-cerbB-2 Single-Chain Fv by Molecular Evolution of the Complementary Determining Regions in the Center of the Antibody Binding Site, J. Mol. Bio., 263, 551 (1996).
Schreiber, H., T. H. Wu, J. Nachman, and W. M. Kast, Immunodominance and Tumor Escape, Cancer Biology, 12, 25
Smith, D. J., S. Forrest, D. H. Ackley, and A. S. Perelson,
Variable Efcacy of Repeated Annual Inuenza Vaccination, Proc. Natl. Acad. Sci. USA, 96, 14001 (1999).
Smith, D. J., A. S. Lapedes, J. C. de Jong, T. M. Bestebroer,
G. F. Rimmelzwaan, A. D. M. E. Osterhaus, and R. A. M.
Fouchier, Mapping the Antigenic and Genetic Evolution of
Inuenza Virus, Science, 305, 371 (2004).
Sprenger, M. J. W., P. G. Mulder, W. E. P. Beyer, R. VanStrik,
and N. Masurel, Impact of Inuenza on Mortality in Relation to Age and Underlying Disease, Int. J. Epidemiol., 22,
334 (1993).
Stuge, T. B., S. P. Holmes, S. Saharan, A. Tuettenberg, M.
Roederer, and J. S. Weber, Diversity and Recognition Efciency of T Cell Responses to Cancer, PLOS Medicine, 1,
149 (2004).
Sun, J., D. J. Earl, and M. W. Deem, Glassy Dynamics in the
Adaptive Immune Response Prevents Autoimmune Disease, Phys. Rev. Lett., 95, 148104 (2005).
Swers, J. S., B. A. Kellogg, and K. D. Wittrup, Shufed
Antibody Libraries Created by in vivo Homologous Recombination and Yeast Surface Display, Nucleic Acids Res., 32,
e36 (2004).
Ungchusak, K., P. Auewarakul, S. F. Dowell, R. Kitphati, W.
Auwanit, P. Puthavathana, M. Uiprasertkul, K. Boonnak, C.
Pittayawonganon, N. J. Cox, S. R. Zaki, P. Thawatsupha, M.
Chittaganpitch, R. Khontong, J. M. Simmerman, and S. Chunsutthiwat, Probable Person-to-Person Transmission of Avian
Inuenza A (H5N1), New Eng. J. Med, 352, 333 (2005).
Whelan, M., J. Whelan, N. Russell, and A. Dalgleish, Cancer
Immunotherapy: An Embarrassment of Riches? Drug Discov. Today, 8, 253 (2003).
World Health Organization (2003), Media Centre Inuenza
Fact Sheet.
World Health Organization, H5N1 Avian Inuenza: First
Steps Towards Development of a Human Vaccine, Weekly
Epidemiological Record, 80, 277 (2005).
Wu, F., Z.-B. Wang, P. Lu, Z.-L. Xu, W.-Z. Chen, H. Zhu, and
C.-B. Jin, Activated Anti-tumor Immunity in Cancer Patients after High Intensity Focused Ultrasound Ablation,
Ultrasound Med. & Biol., 30, 1217 (2004).
Yang, M., J.-M. Park, and M. W. Deem, A Theory of Cancer
Mutation, Escape, and Control by the Immune System, (2005).
Zhou, H., and M. W. Deem, A Cross-Protective Bird Flu
Vaccine will Likely Be Multicomponent (2005).

Vol. 51, No. 12

AIChE Journal


Metabolic Engineering: Developing New

Products and Processes by Constructing
Functioning Biosynthetic Pathways in vivo
Gregory Stephanopoulos and Kyle L. Jensen
Massachusetts Institute of Technology, Dept. of Chemical Engineering 77 Massachusetts Avenue, Cambridge, MA 02137
DOI 10.1002/aic.10725
Published online October 26, 2005 in Wiley InterScience (www.interscience.wiley.com).

Keywords: metabolic engineering, pathway engineering, cell engineering

n a recent AIChE Journal perspective on Metabolic Engineering1 we asked rhetorically whether the microbial world
is so diverse as to allow one to isolate some microbe
capable of producing any desired molecule. It turns out that this
is very likely the case, a manifestation of the enormous diversity of molecules and reaction processes resident in a microbe.
These actually constitute the mechanisms by which cellular
functions are being carried out. The problem with most organisms is that they may make only traces of any single desired
molecule and under conditions that may be difcult to implement on an industrial scale. These microbes must be improved
before their potential is realized. Furthermore, after demonstrating that it is possible to cross the species barrier, pathways
that are incomplete for the production of a molecule in one
organism can now be completed by transferring the missing
pieces from another microbe. Thus, the enormous diversity
present in an array of species can now be commandeered for
accomplishing a specic purpose, such as the production of
entirely new products or the construction of new synthesis
routes for existing products. This is the goal and essence of
metabolic engineering.
Metabolic engineering was developed in the previous decade
to improve industrial strains using modern genetic tools. While
microorganisms were modied before by random mutation and
selection methods, the development of recombinant technologies in early 80s allowed directed strain modication by introducing specic genes conferring desirable properties to cells
for industrial, medical and environmental applications. Metabolic engineering, thus, emerged as the scientic discipline
occupied with the improvement of cellular properties through
introduction to cells of specic transport, enzymatic or regulatory reactions using, primarily, recombinant technologies.2,3
In less than a decade an impressive number of metabolic
engineering applications have appeared in diverse areas includ*Correspondence concerning this article should be addressed to G. Stephanopoulos
at gregstep@mit.edu.

2005 American Institute of Chemical Engineers

AIChE Journal

December 2005

ing aminoacid fermentations,4 polyketide and novel antibiotic

synthesis,5 indigo and aromatic aminoacid synthesis in Escherichia coli,6 golden rice,7 ethanologenic E. coli,8 isoprenoid
(lycopene) overproduction,9 indene biocatalysis for the synthesis of chiral pharmaceuticals,10 tricistronic gene expression in
Chinese Hamster Ovary cells for foreign protein overproduction under no growth and high viability conditions,11 manipulation of the glycosylation pathway in mammalian cells,12
1,3-propanediol and succinate production in E. coli as monomers for polymer production from renewable resources, and
many, many others along with numerous applications in the
medical and environmental areas.13,14
Metabolic engineering makes extensive use of applied molecular biological methods in order to introduce pathway modications and controls at the genetic level. As such, its experimental implementation has a strong molecular orientation.
However, metabolic engineering is much more than just an
industrial variant of genetic engineering. Since the goal is the
overproduction of a product, we must be concerned with the
function of the entire pathway, as well as its optimal conguration in terms of adequate precursor supply and kinetic controls. This means that one needs to examine a broader bioreaction network that extends beyond the strict collection of
those reactions just necessary for product synthesis. For example, many biosynthetic pathways are net producers or consumers of energy (ATP) and reducing equivalents (NADH,
NADPH). These resources are produced and consumed by
many other reactions and for many functions in a cell. The
introduction of a new pathway or the amplication of its rate
should be done in a way that does not disturb the cellular
molecules too far away from their normal (physiological)
steady state. This means that proper consideration must be
given to the entire bioreaction network and this is the main
differentiating characteristic of metabolic engineering: It concerns itself with a biosynthetic route in its entirety instead of
isolated cellular reactions.15 As mentioned, it is often necessary
to transfer some reactions from a different organism in order to
complete a pathway. This opens many possibilities as to the
possible biosynthetic routes that may be used for product
Vol. 51, No. 12


synthesis.16 Some may be thermodynamically infeasible and

some others may incur particularly high costs in terms of
precursor use or energetics. Finally, the ultimate carbon and
energy source(s) (glucose, oxygen, minerals, vitamins) must be
transferred into the cell from the medium. The rate of reactant
import and, similarly product export, are important steps in the
overall process.
It is clear from the above simple outline of a cell factory that
its design, optimization and control make heavy use of the
principles and methods of chemical engineering developed for
the design and operation of chemical plants. This is the basis
for the critical role of the discipline in the genesis and continued growth of metabolic engineering.
While attention is often focused on pathway optimization,
the importance of using the diversity of microorganisms for
new product synthesis should not be underestimated. This can
be done either by introducing specic new enzymes (through
the expression of their genes) in a cell, or by the introduction
of enzyme libraries creating large diversities of entirely new
molecules from which those with desired properties must be
selected. Efcient methods for creating large libraries along
with creative ideas for the selection of desired products are
critical elements in this approach.
Rational, or model-based, and combinatorial methods can be
used in the design of optimal pathways.17 Rational approaches
make use primarily of stoichiometric approaches as stoichiometric models are the only type of models that can be used with
reasonable reliability on a cell-wide basis. Kinetic models are
not as useful because models of enzymatic kinetics that are
valid under intracellular conditions are rare. In addition, the
regulation of these reactions at the transcriptional and enzymatic levels is largely unknown. Combinatorial approaches,
whereby a cell is transformed with random genomic libraries
and well-dened mutants are selected, compliment the rational
approaches. A key concept here is that of inverse metabolic
engineering,18 whereby a desired mutant is selected from a
library and then its specic genetic modication, typically gene
deletion or overexpression, are well characterized. This is a
powerful approach to identifying specic genes that materially
impact the product phenotype either by their stoichiometric,
kinetic or (primarily) regulatory effects.
Biotechnological routes are presently preferable to chemical
ones in the production of chiral pharmaceuticals and complex
ne chemicals, in precision chemistry and addition of new
functionalities to existing molecules, and in the utilization of
renewable resources, all areas with high expected growth rate.
Furthermore, various problems that impaired the development
of biological processes in the past (such as low titer, product
inhibition, and slow rates) are rapidly being solved by a variety
of newly developed methods. For example, product inhibition
is minimized or altogether avoided by engineering enzyme
mutants resistant to high product levels, while osmo-tolerant
strains allow accumulation of continuously higher product concentrations. Continued levels of strong R&D support by the
U.S. Federal Government and industry will be creating an
increasing number of opportunities for biotechnological applications. As a result of these developments, life sciences will
impact the chemical industry in a very profound way. Evidence
of this assertion is widespread: an approximate market of $60B
for chiral pharmaceuticals; a robust and growing biotechnological industry (with more than $40B in sales and hundreds of

December 2005

new products in the regulatory approval pipeline); and an

increasing number of biotechnologically produced products
such as 1,3-propanediol, polylactic acid and an array of new
biopolymers. This trend is very likely to continue in the future
and metabolic engineering will be providing the enabling technologies for harnessing the potential of microbes for an expanding portfolio of new applications.
So, why might chemical engineers be interested in metabolic
engineering? First, metabolic engineering combines the intellectual framework and implementation tools required to capture the enormous potential of biology for industrial and medical applications. Its concepts and tools should be familiar to
chemical engineers as metabolic engineering borrows heavily
from chemical reaction engineering. Second, the importance of
metabolic engineering in materials, fuels, and specialty chemicals (pharmaceuticals and chiral compounds) is undeniable as
evidenced by a growing number of applications in these areas.19 In the medical eld, the greatest impact of metabolic
engineering will be in the development of methods for the
rigorous assessment of the physiological state and determination of reasonable enzymatic targets for the treatment of disease. This will be implemented either by direct therapeutic
intervention or screening programs for the discovery of new
drugs. Finally, this is an excellent entry for chemical engineers
into a very rich eld of scientic inquiry.20 Biological systems
owe their exceptional properties to specic chemical reactions
catalyzed by enzymes that, in a growing number of cases, can
be uniquely prescribed from genomic information. In other
words, genomics provides the means to dene the specic steps
of the chemical reaction system that can be subsequently analyzed using the tools of metabolic engineering. This is a profound difference from typical chemical reacting systems where
dening the actual reaction steps is a major challenge.
Chemical reactions are, for the most part, responsible for the
wonders of biology. Metabolic engineering combines the tools
and concepts of reaction engineering and molecular biology for
the analysis and purposeful modication of bioreaction networks.21 It also provides a framework for integrating and
quantifying genomic information and cell-wide data generated
from modern technologies. As such, it is the natural vehicle for
capturing the enormous potential of biology and transforming
it into the enabling science of many new industrial and medical
applications. Chemical engineers are in a unique position to
extend their educational and research paradigm into the most
exciting eld of scientic inquiry. This will require that they
embrace biology as a foundational science equal to chemistry
and modify the curriculum to reect this fundamental change
into chemical and biological engineering.

NSF award number: BES-0 331364.

Literature Cited
1. Stephanopoulos G. Metabolic engineering: Perspective of
a chemical engineer. AIChE J. 2002:48;920-926.
2. Bailey JE.Toward a science of metabolic engineering. Science, 1991:252;1668.
3. Stephanopoulos G,Vallino JJ. Network rigidity and metaVol. 51, No. 12

AIChE Journal










bolic engineering in metabolite overproduction. Science

Koffas M, Jung GY, Stephanopoulos G. Engineering metabolism and product formation in Corynebacterium glutamicum by coordinated gene overexpression. Metabolic
Eng., 2003:5;32-41.
Pfeifer BA, Admiraal SJ, Gramajo H, Cane DE, Khosla C.
Biosynthesis of complex polyketides in a metabolically
engineered strain of E-coli. Science. 2001:291;1790.
Murdock D. Ensley BD, Serdar C, Thalen M. Construction
of metabolic operons catalyzing the de-novo biosynthesis
of indigo in escherichia-coli. Biotechnol. N.Y., 1993:11;
Ye X. Al-Babili S, Kloti A, Zhang J, Lucca P, Beyer P,
Potrykus I, Engineering the provitamin A (beta-carotene)
biosynthetic pathway into (carotenoid-free) rice endosperm. Science. 2000:287;303.
Ingram LO, Aldrich HC, Borges ACC, Causey TB,. Martinez A, Morales F, Saleh A, Underwood SA, Yomano LP,
York SW, Zaldivar J, Zhou SD. Enteric bacterial catalysis
for fuel ethanol production. Biotechnol Prog. 1999:15,855.
Farmer WR, Liao JC. Improving lycopene production in
escherichia coli by engineering metabolic control. Nat
Biotechnol. 2000:18;533.
Stafford DE, Yanagimachi KS, PA. Lessard, Rijhwani
SK,. Sinskey AJ, Stephanopoulos G. Optimizing bioconversion pathways through systems analysis and metabolic
engineering. Proc Natl Acad Sci USA. 2002:99;1801.
Schlatter S. Bailey JE, Fussenegger M. Novel surface
tagging technology for selection of complex proliferationcontrolled mammalian cell phenotypes. Biotechnol Bioeng. 2001:75;597.
Weikert S, Papac D, Briggs J, Cowfer D, Tom S,
Gawlitzek M, Lofgren J, Mehta S, Chisholm V, Modi N,

AIChE Journal

December 2005








Eppler S, Carroll K, Chamow S, Peers D, Berman P,

Krummen L. Engineering Chinese hamster ovary cells to
maximize sialic acid content of recombinant glycoproteins.
Nat Biotechnol, 1999:17;1116.
Yarmush DM, MacDonald AD, Foy BD, Berthiaume F,
Tompkins RG, Yarmush ML. Cutaneous burn injury alters
relative tricarboxylic acid cycle uxes in rat liver. J of
Burn Care & Rehab. 1999:20;292.
Keasling JD, Van Dien SJ, Trelstad P, Renninger N, McMahon K. Application of polyphosphate metabolism to
environmental and biotechnological problems. Biochem.Moscow. 2000:65;324.
Stephanopoulos G, Sinskey AJ. Metabolic engineering:
Issues and methodologies. Trends in Biotechnology, 1993;
Mavrovouniotis M, Stephanopoulos Geo, Stephanopoulos
Gr. Synthesis of biochemical production router. Comp in
Chem Eng. 1992:16;605.
Alper H, Miyaoku K, Stephanopoulos G. Construction of
overproducing E. coli strains by combining rational and
combinatorial gene knockout targets. Nature Biotechnology. 2005:23;612-616.
Yong-su Jin, H. Alper, Yang Y-T, Stephanopoulos G.
Improvement of xylose uptake and ethanol production in
recombinant Saccharomyces cerevisiae through inverse
metabolic engineering approach. Applied and Environ. Microbiology. (in press) 2005.
Stephanopoulos G, Kelleher J. How to make a superior
cell. Science, 2001:292:2024-2026.
Stephanopoulos G. Chemical and biological engineering.
Chem Eng Sci. 2003:58:3291-3293 and 58:4931-4933.
Stephanopoulos G, Aristidou A, Nielsen J. Metabolic Engineering: Principles and Methodologies.Academic Press,
Inc: San Diego, CA; 1998.

Vol. 51, No. 12



Dimensional Analysis for Planetary Mixer:

Modied Power and Reynolds Numbers
G. Delaplace, R. Guerin, and J. C. Leuliet
INRA-LGPTA, 59651 Villeneuve dAscq, France
DOI 10.1002/aic.10563
Published online August 29, 2005 in Wiley InterScience (www.interscience.wiley.com).

Mixing times, power consumption, heat transfer, and scale-up predictions in an agitated
vessel require the use of correlations between dimensionless groups such as Prandtl, Nusselt,
Power, and Reynolds numbers. These dimensionless numbers are now well established for an
agitated vessel equipped with a vertically and centrally mounted impeller in the tank for both
Newtonian and non-Newtonian uids. To our knowledge, there is more ambiguity as to the
denition of the characteristic speed and dimensions, which should be taken into account in
the dimensional analysis of planetary mixers. The aim of this paper is twofold: (1) to propose
modied Reynolds and Power numbers for planetary mixers and (2) to ascertain the reliability
of the modied dimensionless number proposed for a particular planetary mixer, The TRIAXE, which uses a combination of rotation and gyration of a pitched blade turbine to achieve
mixing. The modied Reynolds and Power numbers proposed involve the maximum tip speed
as characteristic velocity and are consistent with the denition of traditional Reynolds and
Power numbers when only a single revolution around the vertical axis of the mixing device
occurs in the vessel, as is the case for a standard mixing system. Experimental power
measurements carried out with a planetary mixer when mixing highly viscous Newtonian
uids show that the modied Reynolds and Power numbers proposed succeed in obtaining a
unique power curve for the mixing system independently of the speed ratio. This close
agreement proves that the modied Reynolds and Power numbers are well adapted for
engineering purposes and can be used to compare the power-consumption performances of
planetary mixers with well-established technologies. 2005 American Institute of Chemical
Engineers AIChE J, 51: 3094 3100, 2005

Keywords: mixing, planetary mixers, dimensional analysis, power consumption,

Newtonian viscous uids

In the second half of the twentieth century, the systematic
use of dimensional analysis to investigate mixing processes has
allowed this eld to evolve from arts into sciences. Today the
whole eld of classical stirring technology (here, the word
classical refers to impellers vertically and centrally mounted in
the tank) has been examined, so that the denition of signicant dimensionless groups has now become well established.
Correspondence concerning this article should be addressed to G. Delaplace at

2005 American Institute of Chemical Engineers


December 2005

Thus, for any stirring operation (heat transfer, blending, and so

on), depending on the ow regime and the mixing systems
investigated, numerous correlations involving various dimensionless numbers have been proposed in the open literature for
both design and/or scale-up. Consequently, a deeper process
understanding and/or better or reproducible products can be
achieved. This is not yet the case for the planetary mixers.
Indeed, for this kind of mixing equipment, the literature is
relatively scarce.1-6 In addition, some dimensionless numbers
still suffer from ambiguity with respect to the characteristic
speed and dimension, which should be taken into account.
This lack prevents the comparison of performances of the
planetary mixers with those of classical mixing systems. For
Vol. 51, No. 12

AIChE Journal

example, the Reynolds number Re and Power number Np are

now well established for an agitated vessel equipped with
various geometries of impeller vertically and centrally mounted
in the tank for both Newtonian and non-Newtonian uids. Np
and Re are traditionally used to characterize the power demands of a classical mixing system (Np vs. Re), when no vortex
formation occurs in the tank (the Froude number can be neglected). In these dimensionless numbers, the characteristic
length chosen is the diameter of the impeller that, in fact,
corresponds to an external dimension of the impeller perpendicular to the vertical revolution axis. The characteristic velocity chosen (ND) is proportional to the maximum linear velocity encountered in the vessel (ND). Note that the
characteristic velocity corresponds to the impeller tip speed
divided by .
However, for planetary mixers the maximum linear velocity
encountered in the vessel is inuenced by both the relative and
displacement velocities and, consequently, depends on the two
revolution speeds. In the same way, the characteristic dimension of the mixing system is much more complex to dene.
Therefore the classical Reynolds and Power numbers should be
adequately modied to take into account the complexity of the
combined motion required by the agitator.
The aim of this paper is twofold : (1) to propose modied
Reynolds and Power numbers for planetary mixers and (2) to
ascertain the reliability of the dimensionless numbers proposed
for a new planetary mixer, The TRIAXE system, which uses
a combination of rotation and gyration of a pitched blade
turbine to achieve mixing.

Mixing equipment
The mixing equipment used in this investigation is the TRIAXE system (HOGNON S.A., Mormant, France), which
allows the agitator to combine two motions: gyration and
rotation (Figure 1). This planetary mixer is characterized by
two revolutionary motions that are nearly perpendicular. Gyration is a revolution of the agitator around a vertical axis,
whereas rotation is a revolution of the agitator around a nearly
horizontal axis. These double motions allow the agitator to
periodically come in contact with the entire volume of the
vessel. In this work, the mixing tool for the TRIAXE system
is a four pitched blade turbine (Figure 2). The tank used is a
transparent glass cylinder with rounded bottom (Figure 2). The
diameter of the vessel is 0.4 m. In this work, experimental
measurements were carried out when the agitator was fully
immersed in the liquid. This liquid height corresponds to a
liquid volume of 38 L, corresponding to a liquid height of 0.39
The mixing equipment was driven by two variable-speed
motors. To obtain the total power consumption of the TRIAXE system, power draw measurements were carried out
alternately for the two variable-speed drive motors that control
the impeller revolutions. To do so, a torque meter (Scaime Inc.,
Annemasse Cedex, France) ranging from 0 to 5 Nm, was
mounted alternately on the two motor drive shafts (Figure 2)
and the torque was measured for various impeller speed ratios.
Total power draw was estimated as a simple summation of the
two motor drives. More exactly, the procedure to obtain power
requirement of each motor was the following:
AIChE Journal

December 2005

Figure 1. TRIAXE system.

For the planetary mixer studied, ds 0.14 m and D 0.38 m.

(1) The torquemeter is mounted on one of the two motor

drive shafts just before the reduction gearbox. The reduction
gearbox ratio RG for the gyration motor drive is equal to 145.
The reduction gearbox ratio RR for the rotation motor drive is
equal to 34.
(2) The revolution speed of the motor drive shaft that is not
equipped with the torque meter is maintained constant, whereas
torque measurements exerted on the second motor drive shaft
(MG motor or MR motor, respectively) are performed for various
revolution speeds of the second motor drive shaft (NG motor or
NR motor, respectively). This was done both when mixing a
Newtonian uid (MG motor or MR motor, respectively) or when air
(no loading) was contained in the vessel (MG0 motor or MR0 motor,
respectively), to determine the effective torque exerted on each
motor drive shaft [(MG motor MG0 motor) or (MR motor
MR0 motor), respectively].
(3) Torque measurements were carried out on the full range
of speeds available on the motor drive shafts. Because of the
gearbox reductions on the two motor drive shafts, the range of
revolution speeds available for the two agitator shafts (after the
gearbox) are: 0 to 16 rpm for NG and 0 to 90 rpm for NR.
(4) Then neglecting mechanical friction in the gear box, the
effective torque required by each agitator shaft was plotted:
(MG MG0) (MG motor MG0 motor)RG or (MR MR0)
(MR motor MR0 motor)RR, respectively, as a function of the
agitator speeds (respectively NG or NR). At this step, we have
observed that effective torque values exerted on rotation agitator shaft were not dependent on the gyrational speed of the
impeller (see Figure 3 as an example) and vice versa. Consequently, the effective torque values exerted on rotation agitator
shaft (respectively on gyration agitator shaft) can be predicted
by the only knowledge of viscosity and rotational impeller
speed: MR MR0 1.265NR (respectively MG MG0
1.595NG). Note that, for a xed revolution speed and the
Vol. 51, No. 12


Figure 4. Classical geometric parameters and notations

used for mixing vessel equipped with an impeller vertically and centrally mounted in the

Figure 2. Picture of the agitator and vessel of the TRIAXE system used in this investigation.
Red circle on the picture refers to the torquemeter device. The
vessel tank diameter T and liquid height HL were set equal to
0.4 and 0.39 m, respectively. For the planetary mixer studied,
ds 0.14 m and D 0.38 m.

TRIAXE system under investigation, the torque exerted on

gyration agitator shaft is the same order as that obtained for
rotation agitator shaft. The same observation was already made
and detailed in a previous paper7 when the agitator was not
fully immersed in the liquid. Then, the power input for each
motor can be deduced. Power attributed to gyration and rota-

tion are, respectively, PG (MG motor MG0 motor)2NG motor

and PR (MR motor MR0 motor)2NR motor, where NG motor and
NR motor are, respectively, the gyrational and rotational speeds
of the motor shaft. Finally, total power consumption P can be
estimated by adding the contribution of power draw required
by the two motors (PG PR).
(5) Using this procedure, the total power consumption of the
planetary mixer when mixing a Newtonian uid ( 22 Pas
and 1400 kg/m3) was plotted in Figures 8 and 9 (see

Agitated uid
The agitated uid is a highly viscous Newtonian uid. The
liquid consists of glucose syrup/water mixtures of various
viscosities ranging from 15 to 29 Pas, depending on the
temperature encountered in the vessel. The rheological properties and densities of the test uid were obtained at the same
temperatures as those encountered in the mixing equipment.
Density at 20C of the test uid is 1400 kg/m3. The rheological
properties of the test uid were measured in standard controlled
rotational speed concentric cylinders (Rheomat 30, Contraves
AG, Zurich, Switzerland). The shear rate range applied to the
controlled shear rate viscosimeter varied from 0 to 500 s1and
corresponds to the shear rate range encountered in the tank.

Dimensional analysis for an agitated vessel equipped
with an impeller vertically and centrally mounted in the

Figure 3. Effective torque exerted on rotation/ vs. impeller rotational speed.

Symbols refer to various impeller gyrational speeds tested.
Solid line is obtained by linear regression.


December 2005

When mixing Newtonian liquids in an agitated vessel without bafes, equipped with an impeller vertically and centrally
mounted in the tank (Figure 4), the power P of a given stirrer
type and given installation conditions (vessel diameter T, agitator height H, liquid height HL, and bottom clearance C) in a
homogeneous liquid depends on the agitator diameter d (as the
characteristic length), the material parameters of the liquid
Vol. 51, No. 12

AIChE Journal

and dynamic viscosity of the liquid. The installation conditions

refer here to the vessel diameter T, the liquid height HL, and the
bottom clearance C.
For xed installation conditions, this seven-parameter dimensional space leads to a power characteristic consisting of
four pi-numbers

F 5 N pG, ReG ,

NG ds


3 5
where N p G P/ N G
d s and ReG N G d s2 /.
For a given planetary system, the ratio D/ds is xed and the
power characteristic is reduced to a mutual dependency between each of the following three parameters constituting the

F 6 N pG, ReG ,

Figure 5. Sketch and symbols used for the TRIAXE system.

(density and dynamic viscosity ), and on the stirrer speed N.

When no vortex exists, the acceleration attributed to gravity g
is not relevant with respect to the power draw.
The above-mentioned dependency on dimensional parameters
F 1P, d, T, H, H L , C, , , N 0


leads, by way of dimensional analysis, to the following

dependency between dimensionless numbers8:
F 2N p, Re, T/d, H/d, HL/d, C/d 0


Np P/N3d5 is the so-called Newton number (often termed

the Power number) and Re Nd2/ is the Reynolds number.
For a given agitator and xed installation conditions, Eq. 2
is reduced to a dependency between Np and Re.
F 3N p, Re 0



Note that in Eqs. 5 and 6 the characteristic length and

velocity that appear in the modied power N p G and Reynolds
ReG numbers are ds and NGds, respectively. This was chosen at
rst approximation, given that when no rotational motion occurs (NR 0), the planetary mixer under investigatiom is
transformed to a classical mixing system (Figure 6). Using
symbols adopted for classical mixing systems, ds, D, and NG
(Figure 6) become, respectively, equal to d, H, and N (Figure
4). Thus, modied power N p G and Reynolds ReG numbers
transform to the well-known Np and Re traditionally used for
classical mixing systems.
A closer look at Eq. 4 facilitates a reduction in the number
of physical quantities in the list of relevant parameters. Indeed,
it is possible to introduce as a characteristic velocity uch a value
that is proportional to the impeller tip speed and reduce the list
of relevant physical variables describing the problem by three
parameters: NR, NG, and D:
F 7P, d s, , , u ch, installation conditions 0



Dimensional analysis of a planetary mixer

Assuming that the acceleration arising from gravity (as in the
previous case) does not inuence the mixing process of highly
viscous uids, the list of relevant dimensional parameters inuencing the power consumption, when mixing Newtonian
liquids with the TRIAXE system equipped with a pitched
blade turbine, is
F 4P, D, d s, , , N R, N G, installation conditions 0
In Eq. 4, NR and NG are, respectively, the rotational and the
gyrational impeller speeds; D and ds are the geometric parameters reported in Figure 5. and are, respectively, the density
AIChE Journal

December 2005

Figure 6. Analogies between the geometric parameters

of the classical mixing system and the TRIAXE planetary mixer when rotational speed is
set at zero.
Vol. 51, No. 12


u impeller tipt






D 2 2
cos 2NR t 22 NR NG ds D sin2NR t

1/ 2


Figure 7. Example of evolution with time of the absolute

impeller tip speed.
For the two symbols, D 0.38 m and ds 0.14 m. For the
black symbols, NR and NG were set at 0.286 and 0.1 rev/s,
respectively. For the gray symbols, NR and NG were set at
0.286 and 0.345 rev/s, respectively.

For xed installation conditions and a given planetary system

(D/ds is known), this ve-parameter dimensional space leads to
a power characteristic consisting of only two pi-numbers
F 8N pm, Rem 0


with N p m P/ u 3ch d s2 and Rem uchds/.

Note that for a planetary mixer, the instantaneous impeller
tip speed uimpeller tip(t) in an inert reference frame, is not a
constant value with time t. This will be shown and discussed
later. By analogy with a dimensional analysis for classical
mixing systems, the characteristic velocity chosen in Eq. 8 was
the maximum velocity encountered in the vessel divided by
u ch maxuimpeller tipt/


To sum up, the dimensional analysis of the power consumption of the TRIAXE system in a xed installation condition
when mixing highly viscous uids, leads either to a relationship between three pi-numbers (Eq. 6) or two pi-numbers (Eq.
8) if a characteristic speed proportional to the maximum impeller tip speed is introduced in the parametric dimensional
space. In addition, it has been shown that the pi-numbers can be
reduced to well-known Reynolds and Power numbers, when
the rotational speed is zero. This is quite logical because, in this
case, the TRIAXE system is transformed to a classical mixing
system equipped with an agitator vertically and centrally
mounted in the tank.
In the following section, the way to compute the characteristic speed for the TRIAXE system will be detailed. Then, the
reliability of the two pi-numbers proposed will be ascertained,
using power consumption measurements.

where NR and NG are, respectively, the rotational and the

gyrational impeller speeds; D and ds are the geometric parameters reported in Figure 1; and t represents the time.
Depending on whether the ratio NRds/NGd is 1 (see Appendix B), the function t x uimpeller tip(t) does not have the
same evolution with time and reaches two (NRds/NGd 1) or
four extrema (NRds/NGd 1). The various shapes of the
absolute impeller tip speed with time are illustrated in Figure 7.
The derivation of Eq. 10 allows us to obtain the instants for
which the magnitude of absolute impeller tip speed reaches
extrema (see Appendix B), and thus to compute the maximum
value for impeller tip speed max[uimpeller tip(t)]. Finally, using
Eq. 9, uch can be deduced.

Results and Discussion

Power consumption measurements obtained for the TRIAXE
system investigated when mixing a Newtonian uid are shown in
Figure 8. Results are presented in terms of the dimensionless
numbers (NpG vs. ReG) previously dened in Eq. 6.
Figure 8 clearly shows that the power characteristic of the
TRIAXE system cannot be reduced to a unique relation
between Power and Reynolds numbers when ds and NGds are
used as characteristic length and velocity, respectively. Indeed
in this case, it would appear that the power data are strongly
inuenced by the speed ratio values NR/NG.
In contrast, it is shown in Figure 9 that the use of modied
Reynolds and Power numbers, such as those suggested in Eq.
8, succeed in obtaining a unique power characteristic for the
TRIAXE system, independently of the speed ratios chosen.
This proves to a certain extent the reliability of using maximum
impeller tip speed as characteristic velocity and a dimension
perpendicular to the vertical axis of revolution as characteristic
length. Moreover, it can be observed from Figure 9 that the
product N p mRem is constant, as obtained with classical mixing
systems when mixing highly viscous uids. The value of the
product was found to be 343. As with classical mixing equip-

Determination of the characteristic speed of the

TRIAXE system
For the TRIAXE system, which combines the dual motions
reported in Figure 1, the magnitude of instantaneous impeller
tip speed in an inert reference frame is dened as follows (more
details are given in Appendix A)

December 2005

Figure 8. Power characteristics of the planetary mixer

using the three pi-numbers dened in Eq. 6.
Vol. 51, No. 12

AIChE Journal

acteristic velocity that takes into account the two different

revolutionary speeds.

The authors are grateful to F. Brisard and J. F. Dauphin for valuable
work in obtaining the experimental data and gures presented here.

Literature Cited

Figure 9. Power characteristic of the planetary mixer

using the two pi-numbers dened in Eq. 8.

ments, these results allow a Reynolds number to be dened and

the power consumption of the planetary mixer to be predicted,
on the basis of the two impeller revolution speeds (gyration and
rotation) and the liquid properties. It should be noted that a
product N p mRem of 343, does not mean that the power consumption of the planetary mixer is fairly similar to that of the
helical ribbon impeller (NpRe around 300 for standard helical
ribbon9). Indeed, for a given liquid under a laminar regime,
power consumption is obtained not only by multiplying the
power constant NpRe, but also by the characteristic length
exponent 3. In this case, to include the same tank volume, the
characteristic length required for the helical ribbon is around
twofold higher than that of the planetary mixer under investigation. Consequently, the power draw required by the TRIAXE system is lower than that by standard helical ribbon

In this paper, modied Reynolds and Power numbers for a
planetary mixerthe TRIAXE system, which combines dual
revolutionary motions has been developed.
It has been shown that the proposed modied Reynolds and
Power numbers, which involve the maximum impeller tip
speed as characteristic velocity and a dimension perpendicular
to the vertical axis of revolution as characteristic length, allow
one to obtain a unique power characteristic of the mixing
system, regardless of the speed ratios. Moreover, the characteristic length and velocity chosen ensure that the modied
dimensionless numbers proposed are consistent with the denition of traditional Reynolds and Power numbers, when the
impeller performed only a single revolution around the vertical
axis in the tank, as in the case of a classical mixing system. So,
it clearly appeared that the modied Reynolds and Power
numbers can be easily used to compare the power consumption
of planetary mixers with that of conventional mixing systems.
In our judgment, the suggested modied Power and Reynolds numbers can be used for other planetary mixers that
combine two revolutions around a vertical axis. Only the
expression of the characteristic velocity must be modied, to
allow for the differing operating conditions. In contrast to
double-armed planetary mixers or twin-blade arrangement, dimensional analysis dened in Eq. 6 should be preferred because there is no way to reduce the pi-numbers using a charAIChE Journal

December 2005

1. Tanguy PA, Bertrand F, Labrie R, Brito de la Fuente E. Numerical

modelling of the mixing of viscoplastic slurries in a twin blade
planetary mixer. Trans IChemE 1996;74:499-504.
2. Tanguy PA, Thibault F, Dubois C, At-Kadi A. Mixing hydrodynamics
in a double planetary mixer. Trans IChemE. 199;77:318-323.
3. Landin M, York P, Cliff MJ, Rowe RC. Scaleup of a pharmaceutical
granulation in planetary mixers. Pharm Dev Technol. 1999;4:145-150.
4. Zhou G, Tanguy PA, Dubois C. Power consumption in a double
planetary mixer with non-Newtonian and viscoelastic material. Trans
IChemE. 2000;78:445-453.
5. Jongen T. Characterization of batch mixers using numerical ow
simulations. AIChE J. 2000;46:2140-2150.
6. Delaplace G, Bouvier L, Moreau A, Guerin R, Leuliet J-C. Determination of mixing time by colourimetric diagnosisApplication to a
new mixing system. Exp Fluids. 2004;36:437-443.
7. Delaplace G, Guerin R, Bouvier L, Moreau A, Leuliet J-C. Performances of a novel mixing device: The TRIAXE systemMixing
times and power consumption for highly viscous uids. Proceedings of
the 11th European Conference on Mixing, Bamberg, Germany. Dusseldorf, Germany: VDIGVC; 2003:589-596.
8. Zlokarnik M. Stirring Theory and Practice. Weinheim, Germany:
WileyVCH Verlag; 2001.
9. Delaplace G, Leuliet J-C, Relandeau V. Circulation and mixing times
for helical ribbon impeller: Review and experiments. Exp Fluids.

Appendix A Scheme 1
Considering M to be a point located at the TRIAXE impeller tip, the absolute velocity of M in the rotating reference
frame (e x , e y , e z ) is given by

V a









Scheme 1.
Vol. 51, No. 12












V a

D/2cos2NR t

D/2sin2NR t

D/2cos2NR t
2NG D/2sin2NR t
2NG D/2cos2NR t

0 2NG ds /2


ds /2

D/22NR sin2NR t 2NG ds /2

D/22NR cos2NR t
2NG D/2cos2NR t

Consequently, the absolute velocity of M in the inert reference

frame (X , Y , Z ) is given by

D/22NR sin2NR t
D/22NR cos2NR t

D/22NR sin2NR t 2NG ds /2sin2NG t 2NG D/2cos2NR tcos2NG t

V a D/22NR sin2NR t 2NG ds /2cos2NG t 2NG D/2cos2NR tsin2NG t
D/22NR cos2NR t
Thus, the magnitude of absolute velocity of M is given by
V a u impeller tipt

D/22NR 2 2NG ds /22 2D/22NR 2NG ds /2sin2NR t

2NG D/22 cos2 2NR t

Appendix B
It can be shown that the function t x uimpeller tip(t), dened in
Eq. 10, reaches an extremum when the derivative of the function
t x w(t) is equal to zero with w(t) given by w(t) 2(D/
2)2NR2NG(ds/2)sin(2NRt) [2NG(D/2)]2cos2(2NRt).
The derivative of the function t x w(t) is dw(t)/dt
2NRcos(2NRt){[2(D/2)2NR2NG(ds/2)] 2[2NG(D/
Consequently dw(t)/dt 0 when the following conditions

4N R

4N R 4NR

4N R 4NR

or at



4N R

NR ds
NG d


NR ds
NG d

From these extrema, a maximum value for impeller tip speed

max[uimpeller tip(t)] can be selected.

NR ds
sin2NR t
NG d


So, depending on whether ratio NRds/NGd is 1, function t x

uimpeller tip(t) has two or four extrema that are respectively
reached at

Manuscript received July 7, 2004, and revision received Apr. 4, 2005.

December 2005

Vol. 51, No. 12

AIChE Journal


Experimental and Numerical Investigation of

Particle Transport in a Horizontal Pipe
Robert Rundqvist, Camilla Ljus, and Berend van Wachem
Dept. of Thermo and Fluid Dynamics, Chalmers University of Technology, S-412 96 Goteborg, Sweden
DOI 10.1002/aic.10571
Published online September 2, 2005 in Wiley InterScience (www.interscience.wiley.com).

Measurements were made of the concentration of particles in a horizontal transport

pipe using an optical probe. The distribution of particles over the pipe cross-section
was measured. Particle transport in a horizontal two-dimensional (2-D) channel was
simulated using an Eulerian two-uid model. A model is proposed for the turbulent
dispersion of particles and implemented in the two-uid model. The turbulence in the
gas phase is modeled using a modied k model that accounts for the inuence of
the particles on the gas-phase turbulence. Comparisons with measurements of airparticle ow in a horizontal pipe show qualitative agreement between measurements and
calculations. 2005 American Institute of Chemical Engineers AIChE J, 51: 31013108, 2005
Keywords: air-particle ow, two-uid model, particle dispersion, turbulence modication,
channel ow

Particle transport by uids is a common process in nature
and in technical applications. Examples of such two-phase
ows are mixing in bioreactors, the spreading of small particles
in the atmosphere, transport of particles through pipes and
channels, separation in cyclones and sediment transport in
rivers and oceans. In some cases, such as pneumatic conveying,
the transport of particles is strongly inuenced by turbulence.
Thus, turbulent dispersion must be accounted for when the
distribution of particles is calculated. Studies of pneumatic
conveying show that there is a minimum mean velocity at
which particles can be transported through a horizontal pipe or
a channel without being deposited on the pipe bottom. Predictions for this minimum velocity have been derived and discussed by, for example, 1 4. At lower velocities the turbulent
velocity will not be high enough to carry the particles and they
will accumulate at the lower pipe wall. This indicates that, to
avoid this accumulation, turbulent dispersion must be great
enough to counteract the inuence of gravity on the particles.
When numerical calculations are made of these kinds of ows,
Correspondence concerning this article should be addressed to R. Rundqvist at

2005 American Institute of Chemical Engineers

AIChE Journal

December 2005

it is necessary to include models for the turbulent dispersion

force to obtain a particle distribution over the pipe crosssection. If no such force is included, all particles will eventually accumulate at the lower wall of the pipe. Most studies of
particle transport in horizontal channels and pipes are experimental or analytical investigations. There are few studies in the
literature in which the ow elds and the particles distribution
have been calculated numerically. 5 Made numerical calculations of pneumatic conveying. Their study includes particlewall collisions but neglects turbulent dispersion of particles
because of the large particle size. 6 Studied particle transport in
horizontal pipes and included effects of turbulent dispersion,
particle-particle and wall-particle collisions. The two-way coupling between the particles and the gas-phase turbulence was
also modeled. A Lagrangian approach was used in both these
investigations for the particle phase, for example, each particle
is traced through the ow eld. 7 Claimed to have made
Eulerian calculations of gas-particle ow in a horizontal pipe.
The gravitational force is neglected, however, and the turbulent
dispersion is not included in the model. While the Eulerian
approach was also used in calculations in horizontal, vertical,
and inclined pipelines by 8, turbulent dispersion was not discussed.
The presence of particles in turbulent uid ow may alter the
turbulence intensity and the structure of turbulence. ExperiVol. 51, No. 12


mental work by, for example, 9 12 showed that the presence of

particles might either increase or decrease the turbulence levels
depending on the size of the particles. Small particles increase
the dissipation of turbulent energy, and large particles can
cause additional turbulent production. Intermediate sized particles can have either effect, depending on which part of the
ow eld is studied. The length scales of the turbulence may
also be affected, since a redistribution of energy between
different wave numbers can occur, cf. for example, 1315. Many
attempts have been made to model turbulence modication
phenomena numerically, for example, 16 21. However, there is
still room for improvement since many of the models are
simplied and apply only to certain ow cases, for example,
small particles, no additional production, and noslip velocity.
To be able to develop better turbulence models that capture
correct turbulence behavior in particulate ows, it is necessary
to understand the two-way coupling between the particles and
the turbulence.
In this work, numerical calculations of particles transport in
airow in a horizontal, two-dimensional channel were made
using an extended Eulerian two-uid model that takes into
account turbulent dispersion and turbulence modication. In
earlier work predictions of particle distribution over the crosssection of a horizontal channel or pipe, using Eulerian models,
have not been presented. The reason for this is believed to be
the difculty to take the turbulent dispersion, that is necessary
to counteract gravity, into account in such a model.
To verify the models, measurements were made of the
particle volume fraction prole of an air/particle suspension in
a horizontal pipe. Comparisons were also made with previous
measurements of gas velocity proles in the same pipe, cf. 22.
Horizontal gas particle ows have previously been studied
experimentally by for instance, 9 who studied a horizontal pipe
ow with air and 0.2 mm particles using a laser doppler
anemometer. They obtained velocity proles for both phases
using thresholding techniques. 23 Studied particle-wall interaction with spherical and rough particles using particle image
velocimetry. As the work was focused on wall interaction,
primarily impact and rebound characteristics of particles close
to the walls were measured. 24 Studied particle-particle and
particle-wall interaction with spherical glass beads in a horizontal channel air ow using a phase doppler anemometer.
They measured air and particle mean velocity proles, as well
as rst-order moments and particle volume fraction proles.

Figure 2. Probe tip geometry.

Experimental setup
Experimental setup is shown in Figure 1. The particle volume fraction prole was measured in a horizontal aluminum
pipe 10 m long and 0.108 m in dia. A classical Venturi tube
with a machined convergent section with a dia. ratio of 0.75,
designed according to the ISO 5167-1 standard, was used to
measure the mean velocity. At low particle concentrations the
extra pressure drop over the Venturi tube due to particles in the
ow is negligible. The suction fan was set to give a mean gas
velocity of 12 m/s. As particles were introduced, the gas
velocity dropped and the fan was adjusted to keep the original
velocity. The particles were injected at the inlet with a loading
ratio of 0.1. The particles used were poly-acrylate particles
having a diameter of 0.4 mm. The density of the particles is
1,000 kg/m3. An optical probe, described in the next section,
was traversed vertically along the pipe diameter, close to the
end of the pipe. Measurements were made both from the top
and the bottom of the test section in order to get a prole for the
entire pipe diameter. The probe introduces incoherent infrared
light into the measuring volume by an optical ber. Another
ber adjacent to the introducing ber, slanted by 45 degrees to
limit the measuring volume, is used to receive the reected
light from the particles. The conguration of the bers in the
probe is shown in Figure 2.

Optical probe

Figure 1. Experimental test rig: 1. Smoothly shaped inlet

section, 2. Aluminium pipe, 3. Removable pipe
section, 4. Venturi tube, 5. Plexiglas section, 6.
Etoile straightener, 7. Bellow, 8. Fan.

December 2005

The particle volume fraction in dilute ows can be measured by

using several different measuring techniques. We chose to work
with an optical probing technique, which is a simple but reliable
method for measuring volume fraction proles in conned ows.
The probe uses infrared light from a diode that is transmitted into
the ow orthogonally to the main ow direction by an optical
ber. Another ber, slanted by 45 degrees, receives light reected
by the particles. This conguration keeps the measuring volume
nite while minimising reection disturbances. The probe has
been manufactured at the department, more details of the technical
specications can be found in 25.
The main problem in using optical probe techniques for
measuring particle volume fraction is the calibration, cf. for
example, 2528. The method described in the last three referVol. 51, No. 12

AIChE Journal

ences is a straightforward procedure to apply, and as the

volume fractions measured in this work are very low, the
theory is clearly applicable. The method is based on the assumption that most of the light reaching the receiving ber is
scattered on a single particle. That is, light paths reecting on
more than one particle are neglected. This is a good approximation, especially at low particle volume fractions. There is,
however, some uncertainty with respect to the inuence of
glare points on the calibration function. For smooth particles,
the light from the probe will form a glare point on the particles,
and the size of this glare point will vary with the distance to the
probe, inuencing the calibration function. If the particles are
rough, there will be no glare point and the calibration function
will be different than if smooth particles were used. In this
case, the situation was even more complicated as the particles
were found to be partially transparent as well as uneven in
shape and roughness, which made determination of the optical
qualities necessary for a successful calibration a difcult task.
In the dilute limit of the ow through the measuring volume,
every particle will contribute to the registered signal. As the
particle volume fraction increases, some particles will be hidden by other particles. This will introduce nonlinearity in the
calibration function. There is however a range in the dilute
region in which this effect is negligible and the calibration
function can be considered linear. With the present probe and
the given particle diameter, this linear range extends to a
volume fraction of about 0.01, depending somewhat on
whether a glare point is present. This justies the use of a linear
calibration function, where the proportionality factor is determined using the numerical calculations.

Two-uid model
The Eulerian two-uid model used in this work is based on
the traditional forms of the continuity and momentum equations, cf. for example, 29 31. The basic form of the Reynolds
decomposed and volume averaged continuity and momentum
equations are

U 0

t k k
x j k k kj

K(U pi U gi) if k g
K(U U ) if k p
M kIi


For a dispersed ow, the generalised drag force per unit volume in a suspension of particles is written
n pF idr
M kIi


where the number of particles per unit of volume, n p , is




The volume of a particle is here denoted V p . The drag force on

a spherical particle can be written

F sp

d p2
1 sp
C d gU rU r


where d p is the particle diameter and U r is the relative velocity

between the phases. The drag function used here for spherical
particles is that given by 32

(1 0.15(Red )0.687 ) if Red 1000
C Red
if Red 1000


Here, Red is the Reynolds number based on the particle diameter and the relative velocity between the phases.
The viscous force in the momentum equation is given by

V ki

U U k
P MkIi
t k k ki
x j k k ki kj
xj ij
k k gi Cki Vki


where the phase index k is g for the gas phase and p for the
particle phase, k is the average volumetric occurrence of
phase k, k is the density, Uki is the velocity, P is the
gas-phase pressure, MkIi
is the generalised drag, gi is the
gravitational acceleration, Vki is the viscous force, and Cki
represents the particle-particle interaction force, which is
zero in the gas phase. These are the laminar forms of the
transport equations. The equations including the additional
terms that take turbulence into account are presented below
in this paragraph. A number of closure equations are also
needed. The gas density is calculated using the ideal gas
law. The particle-particle interaction force is assumed to be
AIChE Journal

zero at the low-particle concentrations used in this study.

The generalised drag is contributed to by stationary drag,
added mass, transverse, history, and other forces. For suspensions with a small gas-to-particle density ratio, the only
force included in the two-uid model is normally the stationary drag, while the remaining forces are neglected. The
generalised drag can then be written

December 2005

x j k kij


The stress tensor is modeled using the Newtonian strain-stress


kij k

1 U kl
U kl
2 k S kij

x l ij
3 x l ij


and the strain-rate tensor is dened by

S kij

1 U ki U kj

2 xj


where k is the dynamic viscosity of phase k. For dilute

suspensions, the bulk viscosity of phase k, k , is often set to
Vol. 51, No. 12


zero in both phases, in accordance with Stokes assumption, cf.

for example, 33. This work, the particulate phase viscosity is set
to zero, as the particle concentration is very low and the
particulate phase viscosity decreases with decreasing concentration.
If turbulence is taken into account, Eqs. 1 and 2 will have
additional terms that can be obtained by Reynolds decomposing and time averaging of the equations, cf. for example, 18,20,34.
These additional terms will, for example, be correlations between uctuating velocities and uctuating particle concentration.
When these terms are included, the continuity equation for
the particulate phase, for steady-state mean ow, can be written

) gug i pup i
Pij K (U

) pup i gug i
Pij K (U

U pupi 0
x i p pi


where p and up are the uctuating particle concentration and

the uctuating particle velocity, respectively. Another approach is to apply mass-weighted averaging or Favre averaging. A Favre-averaged variable is given by

k k k

k k

gg ug i ug j g g gi
xj g gij


p pU
x j p p pi pj

kuki tk


This means that there is no additional term in the continuity

equation due to the turbulence. There will, however, be additional terms in the momentum equations. The Favre-averaged
equations are written as

k k

x i k k gi



ij k


u u
2 gi gi


The turbulent viscosity of the gas phase can be calculated from

the turbulent kinetic energy k, and the rate of dissipation,
using the following relation

tg C



where C 0.09. The Boussinesq assumption extended to the

particulate phase, according to 18, can be written

up i up j tp


ij kp tp


where k p is the turbulent kinetic energy in the particulate phase



u u
2 pi pi


The last term in Eq. 22 is required to give Eq. 22 the correct

trace, that is to ensure that Eq. 23 is valid.
The additional term in the drag force in Eqs. 16 and 17 that


g gU
x j g g gi gj


where k is the turbulent kinetic energy in the gas phase

Using Eqs. 11 and 13, the Favre-averaged continuity equation


x i p pi


Here, tk is a diffusion coefcient for phase k, which can also

be regarded as a turbulent viscosity.
The Reynolds stresses in the gas phase, ugi ugj , are modeled
using the Boussinesq assumption

ug i ug j tg


Here, higher-order correlations are neglected. Correlation

k uki can be modeled using a gradient diffusion model




pp up i up j p p gi
xj p pij


Neglecting density uctuations, a relationship between time the

averaged, and the Favre averaged can be obtained

December 2005

Vol. 51, No. 12

AIChE Journal

occurs from uctuating velocities and particle concentration

can be interpreted as a drag force on the particles by the
uctuating gas velocity.
A number of models for the turbulent viscosity of the
particulate phase have been suggested in the literature, cf. for
example, 19 20. Their models are all rather similar, with a
dependence on the turbulent viscosity of the gas phase tg , the
particle response time m , and the Lagrangian time integral
scale T L . The model of 18, however, can give negative tp
values for pipe ow. The present calculations use the model of

p p

In this work, the turbulence model for the gas-phase is a

modied k -model based on the model by 18. In their
model, the additional dissipation caused by the particles is
taken into account but the additional production due to the
particles is neglected. In the present work a term taking into
account the additional production caused by the larger particles
is also included. This wake production P kp , becomes important
when the size of the particles is large in comparison to the
turbulent scale. The transport equations for the turbulent kinetic energy k, and the dissipation rate , in the gas phase can
be written

g gP k P kp p
tg ( g g)
g g
g g
U gj

g g

(C P C 2) e p
k 1 k

AIChE Journal


ugiupi 2k



kp k


m TL


Constants k , , C 1 , and C 2 are set to the values used for a

standard k -model in single-phase ow, that is, 1.0, 1.3,
1.44, and 1.92 respectively, cf. for example, 36.
The model by 18 does not take into account the additional
production that larger particles can cause. Several models for
wake production can be found in the literature, most of which
can be written in the form
P kp fU


where f may be a function of particle response time, particle

size, wake size, particle and uid density, and drag coefcient.
37 Derived the turbulence energy equation for the gas phase in
the presence of particles based on the volume averaging of the
mechanical energy equation of the gas phase. The derivation
gives a k-equation with additional terms for both the dissipation caused by particles, and the production caused by particles.
The production term, which was used in the present study, can
be written
P kp

December 2005

m TL

Using similar reasoning, the turbulent kinetic energy of the

particulate phase k p , is modeled as


where P k is the production of turbulence by the velocity eld

P k ug i ug j

k T L

m T L ( m T L) 2 x j x j


Turbulence model


By considering the limiting behavior of correlation ugi upi in

the limits of small and large values of m /T L , that is, the limits
of perfectly responsive particles and nonresponsive particles,
respectively, the correlation, according to 35, can be written


The Lagrangian integral time scale is given by

tg ( g gk)
g gk
g gk
U gj

xj k


Term e p is a correlation between the uctuating gas strain rate

and the strain associated with the uctuating slip velocity. A
derivation by 35 shows that e p can be modeled

T L 0.41

ugi upi

xj xj



Comparing the drag term in the model by 18 with the drag term
in Eq. 2, the particle response time can be written

ugiugi ugiupi
u (U
p p gi gi

tp tg

Term p is the additional dissipation owing to the particles, that

is, the turbulent kinetic energy transfer to the particulate phase,
and e p is an additional term that takes into account the inuence of particles on gas-phase dissipation. By deriving the
transport equations for k and for dilute suspensions, 35 obtains
the following expressions for p and e p , respectively

Vol. 51, No. 12

K 2
U gi U pi


This means that the production of turbulent energy will be

dependent of the drag between particles and gas and the slip
velocity, which will be larger for larger particles.

Numerical method
The densities of both phases are assumed to be constant and
a solver for incompressible two-phase ow is used. A thorough
description of the method is given by 38. The transport equations for the turbulence are solved explicitly for each time step
from the latest calculated velocity led. From the k and elds,
new turbulent viscosities for the two phases, and the turbulent
kinetic energy for the particles phase are obtained. These are
then used in the momentum equations for the next time step.
The obtained solution will not completely stationary. The particles will buildup in the channel, and then be swept away from
time to time. This will give a solution that will vary to a small
extent in time, but the variations will be small. This is a true
physical effect if the velocity of the gas is not high enough to
keep all particles suspended. By integrating the velocity and
concentration prole time-averaged proles comparable to the
measured proles are obtained.

Figure 4. Calculated turbulent kinetic energy.

is applied to the particle phase. Here the slip parameter suggested by 39 is used

The computational domain is a two-dimensional channel
with a width of 0.1 m, and a length of 9.0 m. A coordinate
system is chosen that has the x 1 axis along the length of the
channel and the x 2 axis across the channel and directed upwards. The origin of the coordinate system is located at the
lower wall at the inlet. The grid has 32 cells across the channel
and 900 along the channel length. The grid is rened at the
walls, and the wall cell is 0.77 mm in the x 2 direction, which
is at y between 30 and 40 for the present ow cases.
The horizontal walls are impermeable and a noslip condition
is used for the gas phase. A partial slip condition given by
U p1 L

U p1


d p

6 2 p


where is the sphericity of the particles. The pressure and

particle concentration derivatives are set to zero at the walls.
The derivatives of the velocities, the particle concentration and
the turbulent properties are set to zero at the outlet since the
ow eld is fully developed.
The wall boundary conditions for the k and equations are
set by using standard wall functions based on the friction
velocity u * , cf. 36. The inuence of the particles on the law of
the wall is not taken into account and the standard logarithmic
prole is used. At the inlet, the turbulence intensity is set to a
constant value of 1%.
The inlet mean velocity for the gas, and the particles is 12
m/s, and the loading ratio, that is, the ratio between the mass
ow of the particles and the mass ow of the gas, is 0.1.
The spherical particles are poly-acrylate particles with a mean
particle diameter of 450 m and a density of 1,000 kg/m3.

Results and Discussion

Figure 3. Calculated gas and particle velocity over pipe


December 2005

Measurements by, for example, 22 show that the velocity

prole, and, therefore, also the particle concentration, is asymmetric in pipe ow at the particle loading used.
Figure 3 shows the calculated gas and particle velocities.
The maximum velocity for the gas and particle phase is 14.3
m/s and 12.7 m/s, respectively. Both are clearly skewed towards the upper part of the pipe, and the slip velocity is
relatively constant over the cross-section of the channel. This
skewness is also apparent in the turbulent kinetic energy plot
(Figure 4), where the local turbulent intensities are higher in
the lower part of the pipe due to the lower gas velocity.
Figure 5 shows comparisons between measured and calculated particle volume fractions. The agreement is good, but
there is a slight tendency toward stronger segregation in the
measured prole. That is, the simulated prole is slightly more
Vol. 51, No. 12

AIChE Journal

evenly distributed. The main error sources are considered to be

in the disturbance of the ow by the probe, as the noise levels
in the signals are low and ca be neglected.
A comparison between the shapes of the measured and
calculated gas velocities is shown in Figure 6. The measured
gas velocity is obtained from an earlier experimental study of
the same ow case by 22. The mean velocity of the measured
and calculated proles are the same although the maximum
velocity will differ because of the somewhat different shapes of
the proles. It is evident that the simulated prole is more
skewed toward the upper part of the channel than the measured
prole. This can be interpreted such that the particles absorb
more of the gas phase momentum in the simulations than is the
case in reality. This would force the simulated gas to move
faster in the dilute upper regions, thereby skewing the prole.
The explanation for the discrepancy may lie in one or more of
the following four possible explanations:
1. The drag model over predicts the momentum transfer
between the phases. This is unlikely, especially for such dilute
suspensions. The interphase drag is one of the more thoroughly
researched closures and, given the other uncertainties, the
reason for this discrepancy should probably not be sought here.
2. The partial-slip boundary condition for the particle phase
over predicts the momentum transfer from the particle phase to
the wall. This will lead to a larger slip velocity and, thus, a
larger momentum transfer between the phases.
3. The turbulent particle viscosity over predicts the momentum transfer from the particles to the wall. A lower viscosity
would transport momentum less efciently and, thus, reduce
the slip velocity between the phases. This would also lead to a
less skewed velocity prole.
4. The discrepancy is a physical effect of the difference
between the simulated 2-D channel ow, and the 3-D pipe
ow. In the pipe ow, the midsection contains relatively more
momentum than the peripheral upper and lower regions, which
could cause the midregion to act as a momentum buffer. As the
particle phase drains momentum from the gas phase in the
lower parts of the pipe ow, the core region can provide
momentum without decelerating as much. This would imply
that the 2-D channel velocity prole should be more skewed
than the 3-D pipe ow.

Figure 6. Measured and calculated gas velocity.

The correct explanation (or explanations) should be the subject

of further investigation.

The agreement between model predictions and measurements is good. Further investigations should include a 3-D
simulation of a horizontal ow.

The authors would like to thank the Swedish Foundation for Strategic
Research (SSF) and SCA Research AB for nancial support of this work.

C sp
C fib
C ki
F dr
F sp
M kIi

P kp

Figure 5. Particle distribution over the pipe cross-section.

AIChE Journal

December 2005

U gi
U ki
U pi

constant in k -model
constant in k -model
drag function for spherical particle
drag function for straight ber perpendicular to the velocity
drag function for ber
particle-particle force, N/m3
constant in k -model
dia. of test pipe, m
mean dia. of the bers, m
mean dia. of the spherical particles, m
additional term in the equation due to the particle presence, m2/s4
drag force on a single particle, N
drag force on ber perpendicular to the velocity, N
drag force on spherical particle, N
function in wake production model, 1/s
drag function, kg/m3 s
turbulent kinetic energy in the gas phase, m2/s2
turbulent kinetic energy in the particle phase, m2/s2
slip parameter according to 39, m
generalized drag, N/m3
loading ratio
pressure, N/m2
production of gas-phase turbulent energy by the mean velocity
eld, m2/s3
additional production of gas-phase turbulent energy due to the
particles, m2/s3
Lagrangian integral time scale, s
time, s
mean gas velocity in the i direction, m/s
inlet mean velocity, m/s
mean velocity of phase k in the i direction, m/s
mean gas velocity in the i direction, m/s

Vol. 51, No. 12


V ki



relative velocity between the phases, m/s

uctuating gas velocity, m/s
uctuating particle velocity, m/s
viscous force, N/m3
volume of particle, m3
coordinate, m
dimensionless wall distance
mean particle volume concentration
uctuating particle volume concentration
constant in the turbulent diffusion model, kg/ms2
constant in the turbulent diffusion model, kg/m3
Kronecker delta
rate of dissipation of the gas-phase turbulent energy, m2/s3
rate of dissipation of the gas-phase turbulent energy due to particles, m2/s3
dynamic viscosity of phase k, kg/ms
kinematic viscosity of the gas phase, m2/s
turbulent viscosity of the gas phase, m2/s
turbulent viscosity of the particle phase, m2/s
bulk viscosity of phase k, kg/ms
gas density, kg/m3
particle density, kg/m3
constant in k -model
constant in the particle diffusion model
constant in the particle diffusion model
constant in k -model
stress tensor, N/m2
response time for the particles, s
sphericity of particles



k Favre averaged variable for phase k

k time-averaged variable for phase k


Literature Cited
1. Thomas AD. Predicting the deposit velocity for horizontal turbulent
pipe ow of slurries. Int J Multiphase Flow. 1979;5:113129.
2. Oroskar AR, Turian RM. The critical velocity in pipeline ow of
slurries. AIChE J. 1980;26:550 558.
3. Davies JT. Calculation of critical velocities to maintain solids in
suspension in horizontal pipes. Chem Eng Sci. 1987;42:16671670.
4. Cabrejos FJ, Klinzing GE. Pickup and saltation mechanisms of solid
particles in horizontal pneumatic transport. Powder Technol. 1994;79:
5. Tsuji Y, Oshima T, Morikawa Y. Numerical simulation of pneumatic
conveying in a horizontal pipe. KONA. 1985;3:38 51.
6. Huber N, Sommerfeld M. Modelling and numerical calculation of
dilute-phase pneumatic conveying in pipe systems. Powder Technol.
1998;99:90 101.
. Gas-solid ow with
7. Hussainov M, Kartushinsky A, Mulgi A, Rudi U
a slip velocity of particles in a horizontal channel. J Aerosol Sci.
8. Levy A, Mooney T, Marjanovic P, Mason DJ. A comparison of
analytical and numerical models with experimental data for gas-solid
ow through a straight pipe at different inclinations. Powder Technol.
9. Tsuji Y, Morikawa Y. LDV measurements of an aid-solid two-phase
ow in a horizontal pipe. J Fluid Mech. 1982;120:385 409.
10. Rashidi M, Hetsroni G, Banerjee S. Particle-turbulence interaction in
a boundary layer. Int J Multiphase Flow. 1990;16:935949.
11. Kulick JD, Fessler JR, Eaton JK. Particle response and turbulence
modication in fully developed channel ow. J Fluid Mech. 1994;227:
109 134.
12. Kaftori D, Hetsroni G, Banerjee S. The effect of particles on wall
turbulence. Int J Multiphase Flow. 1998;24:359 386.
13. Tsuji Y, Morikawa Y, Shiomi H. LDV measurements of an air-solid
two-phase ow in a vertical pipe. J Fluid Mech. 1984;139:417 434.
14. Sato Y, Hanzawa K, Maeda M. Interactions between particle wake and


December 2005


turbulence, in a water channel ow. in Advances in Multiphase Flow.

Serizawa A, Fukano T, Bataille J, eds. Amsterdam: Elsevier Science;
Boivin M, Simonin O, Squires KD. Direct numerical simulation of
turbulence modication by particles in isotropic turbulence. J Fluid
Mech. 1998;375:235263.
Dannon H, Wolfshtein M, Hetsroni G. Numerical calculations of
two-phase turbulent round jet. Int J Multiphase Flow. 1977;3:223234.
Elgobashi SE, Abou-Arab TW. A two-equation turbulence model for
two-phase ows. Phys Fluids. 1983;26:931938.
Pourahmadi F, Humphrey JAC. Modeling solid-uid turbulent ows
with application to predicting erosive wear. Physico Chemical Hydrodynamics. 1983;4:191219.
Choi YD, Chung MK. Analysis of turbulent gas-solid suspension ow
in a pipe. J Fluids Eng. 1983;105:329 334.
Chen CP, Wood PE. Turbulence closure modeling of two-phase ows.
Chem Eng Commun. 1984;29:291310.
Yokomine T, Shimizu A. Prediction of turbulence modulation by using
k model for gas-solid ows. in Advances in Multiphase Flow.
Serizawa A, Fukano T, Bataille J, eds. Amsterdam: Elsevier Science,
Ljus C, Johansson B, Almstedt AE. Turbulence modication by particles in a horizontal pipe ow. Int J Multiphase Flow. 2002;28:1075
Sommerfeld M, Huber N. Experimental analysis and modelling of
particle-wall collisions. Int J Multiphase Flow. 1999;25:14571489.
Kussin J, Sommerfeld M. Experimental studies on particle behaviour
and turbulence modication in horizontal channel ow with different
wall roughness. Exp in Fluids. 2002;33:143153.
Rundqvist R, Magnusson A, van Wachem BGM, Almstedt AE. Dual
optical bre measurements of the particle concentration in gas-solid
ows. Exp in Fluids. 2003;35:572579.
Bergougnoux L, Misguich-Ripault J, Firpo JL, Andre J. Monte Carlo
calculation of backscattered light intensity by suspension: comparison
with experimental data. Applied Optics. 1996;35:17351741.
Magnusson A. Fluid dynamics of a Wurster-type uidized bed. Doctoral Thesis Dept. of Thermo and Fluid Dynamics. Chulmers University of Technology. Sweden. ISSN: 0346-718X. 2004.
Magnusson A, Rundqvist R, Almstedt AE, Johansson F. Dual bre
probe measurements of solid volume fraction in a circulating uidized
bed. In press: Powder Technol. 2005;151:13. 19.
Ishii M. Thermo-Fluid Dynamic Theory of Two-Phase Flow. Eyrolles;
Boure JM, Delhaye JM. in Handbook of Multiphase Systems. Hetsroni
G, ed. New York: McGraw-Hill; 1982.
Enwald H, Peirano E, Almstedt AE. Eulerian two-phase theory applied
to uidization. Int J Multiphase Flow. 1997;22:21 66.
Schiller L, Naumann ZZ. Ver Deutsch Ing. 1935;77:318.
Panton RL. Incompressible Flow. La Canada, California: DCW Industries, Inc; 1984.
Chung MK, Sung HJ, Lee KB. Computational study of turbulent
gas-particle ow in a venturi. ASME trans. 1986;108:248 252.
Pourahmadi F. Turbulence modeling of single and two phase curved
channel ow. University of California, Berkeley, Department of Mechanical Engineering; 1982. PhD thesis.
Wilcox DC. Turbulence Modeling for CFD. La Canada, California:
DCW Industries, Inc; 1993.
Crowe CT, Gillandt I. Turbulence modulation of uid-particle
owsa basic approach in Third International Conference on Multiphase Flows; France: Lyon; 1998.
Weller HG, Tabor G, Jasak H, Fureby C. A Tensorial Approach to
Computational Continuum Mechanics Using Object Orientated Techniques. Comp in Phys 1998;12:620 630.
Ding J, Lyczkowski RW, Burge SW, Gidaspow D. Three-dimensional
models of hydrodynamics and erosion in uidized-bed combustors.
AIChE Symp Ser. 1992;88:8598.

Manuscript received Apr. 28, 2004, and revision received Apr. 19, 2005.

Vol. 51, No. 12

AIChE Journal


The Effect of Nanobers on LiquidLiquid

Coalescence Filter Performance
C. Shin and G. G. Chase
Microscale Physiochemical Engineering Center, University of Akron, Akron OH 44325

D. H. Reneker
Maurice Morton Institute of Polymer Science, University of Akron, Akron OH 44325
DOI 10.1002/aic.10564
Published online August 10, 2005 in Wiley InterScience (www.interscience.wiley.com).

A number of factors inuence the efciency and the economics of the separation of
dispersed liquid drops in an immiscible liquidliquid mixture. One important factor that
controls the performance of a lter medium in the separation is the ber size. Electrospun
polymer nanobers have diameters in the nanometer range and are arbitrarily long. The
experimental results in this work show that adding nanobers to conventional micronsized brous lter media improves the separation efciency of the lter media but also
increases the pressure drop. An optimum in the performance occurs (signicant increase
in efciency with minimal increase in pressure drop) with the addition of about 1.6% by
mass of 250 nm diameter nylon 6 nanobers to 5 micron diameter glass bers in the lters.
2005 American Institute of Chemical Engineers AIChE J, 51: 3109 3113, 2005

Keywords: nanobers, coalescence, ltration, lter media

Glass-ber lter media augmented with small amounts of
nanobers show improvement of separation efciency. Filter
media with equal amounts of nanobers made of MPD-I [metaaramid, poly(meta-phenyleneisophthalamide)], nylon 6, or
polyacrylonitrile (PAN) were tested for liquidliquid coalescence ltration in a previous article.1 The aim of the study
reported in this article is to present the effects of varying the
amount of nylon 6 nanober added to the glass-ber media on
the pressure drop and the separation efciency in liquidliquid
coalescence lter performance. The nanobers are produced by
electrospinning and have diameters typically in the range of
10 500 nm.2-5
The electrospinning process is driven by the electrical forces
on free charges on the surface or inside of a polymeric liquid.
When the free charges, generally ions, in the polymer solution
move in response to the electric eld, they quickly transfer a
Correspondence concerning this article should be addressed to G. G. Chase at

2005 American Institute of Chemical Engineers

AIChE Journal

December 2005

force to the polymer solution. When the electric eld reaches a

critical value at which the repulsive electric force between
charges on the surface overcomes the surface tension force, a
charged jet of the solution is ejected from the tip of a cone
protruding from a liquid drop of the polymer. As the jet
stretches and elongates in the air, the solvent evaporates, leaving behind a charged solid polymer ber that lays itself randomly on a collecting metal screen. Thus, continuous bers are
produced to form a nonwoven fabric.2,6 Nonwoven mats of
electrospun bers have a large surface area per unit mass and
small pore sizes compared to those of commercial textiles,
making them excellent materials for use in ltration applications.7
Water-in-oil emulsion separation is important to the petroleum and chemical industries to remove the dispersed liquid for
safety, ecologic, and economic reasons. The coalescence lter
is effective for separation of secondary dispersions, although
dispersions of drops with diameters 100 m can be difcult
to separate. Fibrous lter media can provide high ltration
efciency with low pressure drop, depending on the ow rate,
bed depth, ber surface properties, ber size, drop size, and
ber orientation.8-21
Vol. 51, No. 12


of the lter (the inlet stream) and downstream of the settling

tank as shown in Figure 1 (the outlet stream). The size distributions of water droplets in the emulsion samples are measured
with a particle size analyzer (Hyac Royco BR8 particle
counter, eight channels with sizes from 1 to 150 microns).
Some of the coalesced drops exiting the lter media are too
large for the BR8 to measure. Because of this the measurement
point for the outlet stream was positioned downstream of the
settling tank. The measured performance is a combination of
the lter and the settling tank.
Separation performance can be characterized on a total mass
basis, but the mass basis is biased to larger drop sizes. To
determine how the separation efciency varies with droplet
size, the separation efciency of droplets of size x is dened as
Figure 1. Experimental apparatus for lter testing.
Oil from the oil tank is pumped through a mixing pipe,
through a lter sample holder, into a settling tank, and into a
reservoir. Water droplets are dispersed into the oil by injection
from a syringe into the mixing pipe. Emulsion samples are
removed from the ow stream for drop size measurements at
the sample points upstream (U) and downstream (D) of the
lter. The latter is located between the settling tank and the

Number of drops of size x removed

Number of drops of size x in the inlet

This equation is rewritten in terms of the number of drops of

size x measured in the outlet stream as
e x 1

Experimental Setup and Procedures

The experimental apparatus is shown in Figure 1. A mixture
of water-in-oil was used in the experiments, in which deionized
water droplets were dispersed in an oil phase having properties
similar to those of diesel fuel (Viscor 1487, specic gravity 0.83; Rock Valley Oil and Chemical Company). The
Viscor 1487 is pumped from the oil tank by a peristaltic
pump, at a constant ow rate, through a mixing pipe (where the
water drops are mixed with the oil), through the lter sample
and into a settling tank and a reservoir. The ow rate is
controlled by selection of the tube diameter for the peristaltic
pump. A ow rate of 100 mL/min was used in all experiments
in this work.
The water-in-oil emulsion is produced in the mixing pipe.
The mixing pipe is a Plexiglas tube with a 1 mm inside
diameter. The mixing pipe is inserted into the exible tube
between the pump and the lter sample. Water is pumped by a
syringe pump (Model Sp101i, with ow ranges from 0.001 to
1.175 mL/min; World Precision Instruments, Sarasota, FL)
through a hypodermic needle into the middle of the mixing
pipe through which the oil ows. A very ne water-in-oil
emulsion is produced at the outlet of the hypodermic needle,
arising from the shear force created by the owing oil phase, in
which 99.8% of water droplets are 30 microns,. The concentration of water and size of the droplets are controlled by
adjusting the water ow rate using the syringe pump and the
velocity of the oil past the tip of the needle.
The water droplets produced in the mixing pipe do not settle
out of the oil for a reasonable length of time. The water
droplets in the oil come into contact with each other in the lter
medium and form larger coalesced droplets of water that are
carried downstream to the settling tank. The larger drops are
easily separated from the oil in the settling tank. Smaller drops
that do not settle out by gravity are carried into the reservoir
Samples of the water-in-oil emulsion are taken at intervals of
10 to 25 min from the ow system at sample points upstream

December 2005


n O x
n I x


where nO(x) is the number count of drops of average size x for

a size range x in the outlet stream and nI(x) is the number
count of drops of the same average size and size range in the
inlet stream.
The lter media samples used in the experiments are formed
from glass bers, nanobers, and a binder (Carboset 560, BF
Goodrich) by vacuum molding an aqueous slurry of bers and
binder. Sulfuric acid is added to the slurry to adjust the pH in
the range of 2.0 to 2.5 to disperse the glass bers. The nanobers are electrospun directly into the surface of the aqueous
slurry of glass bers and binder solution and the mixture is
agitated using injected air.25
Figure 2 shows the apparatus used for electrospinning the
nanobers. A nylon-6 solution is prepared by dissolving 16 wt
% nylon-6 by mass in 84 wt % formic acid.1 The polymer

Figure 2. Experimental apparatus of electrospinning.

The polymer is pumped through the hypodermic needle. The
needle is charged to 20,000 volts. The polymer jets from the
needle, elongates into nanobers, and collects on the surface
of the slurry of glass bers.

Vol. 51, No. 12

AIChE Journal

Table 1. Masses of Fibers in the Filter Media and Measured

Separation Efciencies


Mass (g)


Mass (g)

Total Filter
Mass (g)






*Calculated by subtracting the ber amounts from the total lter mass.

solution is loaded into the syringe and pumped through a

needle suspended about 20 cm above the aqueous slurry surface. The syringe pump is set at a rate of 4.1 microliters per
minute and the needle is charged to 20,000 to 25,000 volts to
spin the polymer nanobers.

Results and Discussion

The aim of this research is to investigate the effects of
adding varying amounts of polymer nanobers to glass-ber
media on lter media performance. Five lters labeled A
through E, as listed in Table 1, were made by electrospinning
varying amounts of nylon-6 nanobers in each lter sample.
Table 1 lists the masses of each ber type, the mass of binder
added to the lter to glue the bers together, and the total mass
of each lter sample. Each sample was made in triplicate and
the average values of the three samples are listed in the table.
The nanobers spun from the apparatus described in Figure 2
had diameters varying from 200 to 300 nm, with an average
diameter of about 250 nm (Figure 3). SEM images such as
Figure 3 show that nanobers appear as a single ber intermingled with the larger glass bers. The nanobers crisscross
over the void space between the larger glass bers and are
supported by the glass bers. The glass bers have diameters
varying from about 2 to about 7 microns, with an average
diameter of about 5 microns and lengths varying from about
100 microns to 1 mm.
The steady-state number counts of particles for different bin
sizes from the ltration experiments are shown in Figure 4. The

Figure 3. SEM image of a 5 micron diameter commercial

glass ber and 250 nanometer diameter electrospun nylon 6 polymer nanobers.
AIChE Journal

December 2005

Figure 4. Steady-state drop size distributions measured

in the inlet stream before the lter holder and
in the outlet stream between the settling tank
and the reservoir.
The curve marked as No Filter is the measured outlet distribution for when no lter is present in the lter holder. The
curves marked Filter A through Filter E are the distributions
for when the lters are tested in the holder.

drop size on the horizontal axis is the average drop size of the
particular bin (the bin sizes in the BR8 are specied by the
largest particle size in that particular bin). The size distribution
marked as Inlet is the size distribution measured from the
sample point upstream of the lter. The distribution marked as
No-Filter was measured at the outlet stream sample point
between the settling tank and the reservoir when no lter was
in the holder. The difference between the Inlet and NoFilter curves shows some of the particles are separated out by
the settling tank even when there is no lter in the line. The
distribution curves marked as Filter A through Filter E
were similarly measured at the outlet stream sample point and
are the steady-state downstream distributions when the lters
are placed in the line. With lters in the line the downstream
numbers of particle concentrations are signicantly reduced.
The design of the settling tank may inuence these results;
thus, the same settling tank was used for all experiments and
these results are comparable between the experiments for relative performance.
Figure 5 shows the separation efciency distributions. The
system with no lter approaches 100% efciency for drop
sizes 125 microns. All of the lters approach 100% efciency for sizes 100 microns. Filter A shows a separation
efciency of 67% for drop size of 125 microns, but this low
efciency is attributed to one drop detected in the outlet stream
from Filter A, whereas the inlet stream contained three such
In the 0 to 100 micron drop size range signicant differences in performance are noted. Filter A performs better
than No Filter; Filter B performs better than Filter A; and
Filters C, D, and E all perform about the same, all better than
Filter B.
The steady-state pressure drops across the lters are shown
in Figure 6. The lters supplemented with nanobers have a
higher pressure drop than that of lters without nanobers. The
change in pressure drop with the addition of nanobers is
almost linear. Because Filters C, D, and E have similar high
capture efciencies then an optimum design is one that has the
Vol. 51, No. 12


least pressure drop among the lters with the highest capture
efciencies. This means that, of the lters tested here, Filter C
with 0.014 g of nanober (about 1.6% of the lter mass) is the
optimum design.
The pore sizes in the microber lters (no nanobers) are
typically the same order of magnitude as that of the ber
diameter (about 5 microns). Large droplets are easily captured
by interception with multiple bers and the efciency for
microber lters is high for the larger drops. A greater fraction
of smaller drops near the pore size may pass through the pores
without coalescing into larger drops. Thus the lter efciency
decreases for smaller droplet sizes.
Results of this study conrm that the addition of nanobers
improves the lter performance. The nanobers crisscross over
pore spaces, making the pore openings much smaller. Thus it is
more difcult for smaller drops to pass through the pores
without coalescing with other drops.
For the drop sizes considered herein the dominant mechanism of capture is by direct interception and inertial impaction
with multiple bers. Most of the droplets are larger than the
pore openings, and thus the mechanism of single-ber capture
does not apply. The small diameter of the nanober allows us
to reduce the pore sizes with the addition of only small amounts
of nanobers and yet maintain a high porosity.
Other mechanisms, such as surface wetting properties, may
inuence the coalescence performance. We consider the interception of the drops by the nanobers to be more important in
this work.

In this work, the effect of different amounts of polymer
nanobers added to the lter media is experimentally evaluated. The addition of nanobers to glass-ber lter media
improves the capture efciency, but also causes an increase in
pressure drop. The experimental results show that an optimal
amount of nanober enhances the capture efciency but does
not cause excessive pressure drop. For the materials tested
herein 1.6% nanober by mass is optimal.

Figure 5. Separation efciency by drop size.

Separation efciencies (especially for the smaller drop sizes)
increase with the nanober amount up to about 0.014 g of
nanober. Filters of equal or greater than 0.014 grams of
nanober performed at about the same efciency.


December 2005

Figure 6. Pressure drop vs. the amount of nanober

added to the glass-ber lters.

This work was supported by Ahlstrom Paper Group, Donaldson Company, Fleetguard Inc., Hollingsworth and Vose, Parker Hannin, and
Produced Water Seminar. This work was also partially supported by
National Science Foundation Grants CTS-0310429 and DMI-0403835.

Literature Cited
1. Shin C, Chase GG. Water-in-oil coalescence using glass ber lters
augmented with polymer nanobers. AIChE J. 2004;50:343-350.
2. Reneker DH, Chun I. Nanometre diameter bres of polymer, produced
by electrospinning. Nanotechnology. 1996;7:216-223.
3. Reneker DH, Yarin AL, Fong H, Koombhongse S. Bending instability
of electrically charged liquid jets of polymer solutions in electrospinning. J Appl Phys. 2000;87:4531-4547.
4. Shin YM, Hohman MM, Brenner MP, Rutledge GC. Electrospinning:
A whipping uid jet generates submicron polymer bers. Appl Phys
Lett. 2001;78:1149-1151.
5. Gibson P, Schreuder-Gibson H, Rivin D. Transport properties of
porous membranes based on electrospun nanobers. Colloids Surf A.
6. Doshi J, Reneker DH. Electrospinning process and applications of
electrospun bers. J Electrostat. 1995;35:151-160.
7. Deitzel JM, Kleinmeyer J, Harris D, Beck Tan NC. The effect of
processing variables on the morphology of electrospun nanobers and
textiles. Polymer. 2001;42:261-272.
8. Sareen SS, Rose PM, Gudesen RC, Kintner RC. Coalescence in brous
beds. AIChE J. 1966;12:1045-1050.
9. Fahim MA, Akbar AM. Removal of ne oily hazes from wastewater
using deep brous bed coalescer. J Environ Sci Health. 1984;A19:
10. Othman FM, Fahim MA, Jeffreys GV, Mumford GJ. Prediction of
predominant mechanisms in the separation of secondary dispersion in
a brous bed. J Dispers Sci Technol. 1988;9:91-113.
11. Secerov Sokolovic RM, Sokolovic SM, -Okovic BD. Effect of
working conditions on bed coalescence of an oil-in-water emulsion
using a polyurethane foam bed. Ind Eng Chem Res. 1997;36:49494953.
12. Moses SF, Ng KM. A visual study of the breakdown of emulsions in
porous coalescers. Chem Eng Sci. 1985;40:2339-2350.
13. Hajra MG, Mehta K, Chase GG. Humidity, temperature, and polymer
nanobers on drop coalescence in glass ber media. Sep Purif Technol.
14. Hazlett RN. Fibrous bed coalescence of water: Role of a sulfonate
surfactant in the coalescence process. Ind Eng Chem Fundam. 1969;
15. Hazlett RN. Fibrous bed coalescence of water: Steps in the coalescence process. Ind Eng Chem Fundam. 1969;8:625-632.
16. Akbar AM, Othman FM. Prediction of oil saturation in a brous bed
coalescer from pressure drop data. J Dispers Sci Technol. 1989;10:
17. Chase GG, Beniwal V, Venkataraman C. Measurement of uni-axial
ber angle in non-woven brous media. Chem Eng Sci. 2000;55:21512160.

Vol. 51, No. 12

AIChE Journal

18. Sherony DF, Kintner RC. Coalescence of an emulsion in brous bed:

Part I. Theory. Can J Chem Eng. 1971;49:314-320.
19. Sherony DF, Kintner RC. Coalescence of an emulsion in brous bed:
Part II. Experimental. Can J Chem Eng. 1971;49:321-325.
20. Spielman LA, Goren LG. Theory of coalescence by ow through
porous media. Ind Eng Chem Fundam. 1972;11:66-72.
21. Chokdeepanich S. Coalescence Filtration and Drainage Design. PhD
Dissertation. Akron, OH: The University of Akron; 2002.
22. Allen T. Particle Size Measurement. Vol. 1, 5th Edition. London, UK:
Chapman & Hall; 1997.

AIChE Journal

December 2005

23. Svarovsky L. SolidLiquid Separation. 3rd Edition. London, UK:

Butterworths; 1990.
24. Brown RC. Air Filtration: An Integrated Approach to the Theory
and Applications of Fibrous Filters. New York, NY: Pergamon;
25. Chase GG, Reneker D, Rangarajan S, Mehta K. Method and Apparatus
of Mixing Fibers. WO Patent No. 01/68228A1; 2001.

Manuscript received Feb. 20, 2004, and revision received Apr. 5, 2005.

Vol. 51, No. 12


Application of CAMD in Separating Hydrocarbons

by Extractive Distillation
Biaohua Chen, Zhigang Lei, Qunsheng Li, and Chengyue Li
The Key Laboratory of Science and Technology of Controllable Chemical Reactions, Ministry of Education, Beijing
University of Chemical Technology, Box 35, Beijing, 100029, China
DOI 10.1002/aic.10562
Published online August 22, 2005 in Wiley InterScience (www.interscience.wiley.com).

The solvent is the core of extractive distillation, and a suitable solvent plays an
important role in the economical design of extractive distillation. Computer-aided molecular design (CAMD) has been applied to rapidly screen the solvents for separating
hydrocarbons by extractive distillation. The systems of propane/propylene, n-butane/1butene, and n-heptane/benzene, respectively, as the representatives of C3, C4, and C6
hydrocarbons were investigated, and the potential solvents were selected by means of
CAMD. The designed results were further proven by experiments and process simulation.
The mechanism for separating hydrocarbons by extractive distillation is based on the
different uidities of the electron cloud of COC (no double bond), CAC (one double
bond), and ACH (aromatic carbon ring) bonds and thus different interactions between
solvent and hydrocarbon molecules. To improve the separation ability of the main solvent,
one strategy is to add some additive that can form hydrogen bonding with the main solvent
to make into a mixture. 2005 American Institute of Chemical Engineers AIChE J, 51:
3114 3121, 2005

Keywords: computer-aided molecular design (CAMD), extractive distillation, relative

volatility, hydrocarbons, application

Extractive distillation is commonly applied for separating
hydrocarbons with close boiling points, such as C4, C5, and C6
mixtures and so on.1-3 In extractive distillation, an additional
solvent (entrainer or separating agent) is used to alter the
relative volatility of the components to be separated. In this
way, it is possible to obtain one pure component at the top of
one column and the other, together with the solvent at the
bottom, which may then be easily separated in a secondary
distillation column because of the high boiling point of the
solvent. The solvent does not need to be vaporized in the
extractive distillation process.
It is known that the solvent is the core of extractive distillation, and a suitable solvent plays an important role in the
Correspondence concerning this article should be addressed to B. Chen at
chenbh@mail.buct.edu.cn or Z. Lei at leizhg@mail.buct.edu.cn.

2005 American Institute of Chemical Engineers


December 2005

economical design of extractive distillation. However, it is

tiresome to choose the best solvent from thousands of different
substances for a given system through experiments. The computer-aided molecular design (CAMD) developed in the 1980s
may break new ground in this aspect by largely reducing the
amount of experimental work.4-15 The application of CAMD in
chemical engineering is mostly based on the UNIFAC group
contribution. Incalculable molecules would be formulated by
simply joining UNIFAC groups, if they were without any
constraint. In accordance with certain combination rules, however, the size of the combinatorial problem can be greatly
reduced and only then are the chemically feasible molecules
generated. Furthermore, in terms of given target properties, the
desired molecules are screened from chemically feasible molecules. The groups of UNIFAC provide building blocks for
assembling molecules. CAMD is essentially the inverse of
property prediction by group contribution. Given a set of
desirable properties, it is proposed to nd a combination of
structural groups satisfying the property specications. In most
Vol. 51, No. 12

AIChE Journal

cases more than one solution is produced. Thus, a screening is

needed because only one of the alternatives may be chosen for
the specied problem. Moreover, such factors as corrosion,
prices, source, azeotrope, and so on should be taken into
consideration. Of course, this procedure is performed after
Many studies have been done in such unit operations as gas
absorption, liquidliquid, and so on. However, the application
of CAMD in exploring the separation mechanism of hydrocarbons by extractive distillation has rarely been systematically
reported. In this work, CAMD is applied for screening the
solvents of extractive distillation, and some common and important systems in industry are investigated.

CAMD Program
CAMD is conducted in the following four steps.

Step 1: group sorting and preselection

Groups are the basis of CAMD and molecular design makes
full use of the group concept raised by Franklin,16 which is
built on the UNIFAC groups. The groups must be systematically ordered to facilitate their use. We adopt the approach
proposed by Gani et al.17-19 to sort the groups. That is, a certain
group is characterized by the number of attachments present in
a given group (or the valence number of the group) and the
degree of difculty that the group combines other groups (or
the type of attachment). If a group belongs to several different
classes, it means that this group takes on different types of
characterization in different molecules.
The attachment types are indicated as ordered pairs (i, j),
where i is the type of attachment and j is the number of i
attachments. Five types of attachments are put forwarded for
nonaromatic groups:
N single molecules as a group having no attachment with other groups,
such as H2O
K severely restricted attachment, such as OH
L partially restricted attachment, such as CH2Cl
M unrestricted carbon attachment in linear dual-valence or single-valence groups, such as CH3
J unrestricted carbon attachment in radial dual-valence groups, such as

For aromatic molecules, two new attachments are introduced:

I aromatic carbon ring attachment, such as ACH
H substituted aromatic carbon ring attachment, such as ACCl

Types M and J attachments are extended to aromatic groups:

M unrestricted attachment in a carbon linked to an aromatic carbon,
such as ACCH2
J unrestricted attachment in a radial carbon linked to an aromatic
carbon, such as ACCH

The chemically feasible molecules will be generated from

the characterized UNIFAC groups in terms of combination
rules. Not all of the UNIFAC groups need to be used in CAMD
for extractive distillation. The groups may be prescreened
according to the criteria: availability of binary parameters for
the synthesis group and elimination of unsteady, corrosive,
toxic compounds. For instance, some groups, that is, Cl and
F, may cause corrosion to the equipment and must be avoided
in the extractive distillation process.
AIChE Journal

December 2005

Step 2: combination of groups

The greatest difculty in CAMD is assembling groups into
one molecule. To generate chemically stable molecules, the
assembly must t to the following rules:
The chemical valence of a molecule must be zero.
The neighborhood effect of groups must be avoided. Fortunately, many researchers17-20 have discussed this rule from
different perspectives and given the corresponding restriction
conditions. Their work is helpful for us to program CAMD.
In general, a molecule is composed of not more than eight
groups and the polar groups cannot number more than three.
The groups in a molecule must accord with the following
attachment criterion: K M J/2 2 for aliphatic compounds; I H 6, H 2, or H 3 for aromatic compounds.
For aliphaticaromatic compounds these restrictions must be
satised simultaneously; otherwise, this molecule is unsteady
under normal conditions.
The group parameters and group interaction parameters
must be known in advance. We have collected a set of parameters for 109 subgroups and 51 main groups based on the
original UNIFAC model. Some parameters directly come from
literature values21-25 based on the original UNIFAC model,
whereas others are obtained by transformation from the modied UNIFAC model.26-30 In total about 1200 binary interaction parameters of UNIFAC groups are present. The properties
concerned can be calculated by using these parameters.
The combination procedure is carried out in such a manner
that the only combinations considered are those resulting in the
generation of chemically feasible structures. The combinations
from a preselected number of groups and testing for their
chemical feasibility take place simultaneously.

Step 3: prediction of target properties

For CAMD for extractive distillation, such properties as
relative volatility at innite dilution ( ij ), selectivity at innite
dilution (S ij ), solubility capacity (SP), molecular weight (MW),
and boiling point (Tb) are important. Specication of the problem type identies the corresponding target properties. Because
not all of the target properties are computable, it is convenient
to classify them as explicit target properties and implicit target
properties. Prediction methods for explicit target properties are
available and can be implemented automatically by computer.
Prediction methods for implicit properties are not presently
available and thus a combination of experience, information
from the open references, and experiments are needed to determine them. The prediction methods for explicit properties
are given in Table 1, where i is the activity coefcient of
component i at nite dilution, P 0i is saturated pressure, and MW
is molecular weight.
On the other hand, a databank of pure components constituting the physical properties of about 500 compounds, such as
molecular weight, normal boiling point, critical properties,
vapor pressure constants in Antoine equation, ideal gas heat
capacity, and the likeis also set up. If target properties can be
sought in the databank, the values from the databank (not from
calculation) are regarded as the ultimate results.

Step 4: sort order and selection of potential solvents

As we know, relative volatility (which is consistent with
selectivity) is the most important among all the explicit propVol. 51, No. 12


Table 1. Explicit Properties Estimation

for Screening Solvents
Relative volatility at
Selectivity at innite
Solubility capacity20
Molecular weight
Normal boiling point
Critical properties
Vapor pressure
Azeotropic judgment



i Pi0
j Pj0

1 MWi
i MWs
Pure component data bank or by adding
group parameters32
Pure component data bank or by adding
group parameters3234
Pure component data bank or Joback
Antoine equation or Riedel equation32,33
By drawing the curves of x (mole fraction in
the liquid phase)y (mole fraction in the
vapor phase) to judge

erties. The solvent with the highest relative volatility is always

considered to be the most promising solvent for a given separation task. For this reason, solvents are ranked in order of
decreasing relative volatility (or selectivity). Moreover, azeotropic judgment is carried out in this step by calling the dew
point subroutine. Its role is to determine whether the solvent
can form azeotropes with the components to be separated and
at which concentration. Because it takes extended CPU time to
draw an xy curve for each solvent and component to be
separated, this function is set in the end applying to the potential solvents.
In extractive distillation it is necessary to evaluate such
implicit properties as toxicity, cost, stability, and material
source. The solvents that do not satisfy the requirement of
implicit properties are crossed out from the order. The remaining solvents ranked in the front are possible to be the potential
solvents we seek.
CAMD has been programmed with Visual Basic for Windows. This program can provide much important information
for extractive distillation. CAMD is an easy-to-use software
with a user-friendly interface. One with only slight knowledge
of extractive distillation can become familiar with it within a
short time. A ow diagram of CAMD program is illustrated in
Figure 1. For the solvent mixture, one is the main solvent, the
other is the additive. In CAMD, the rst step is to nd the main
solvent; when the basic solvent is determined, the next step is
to nd the additive in the same way as single solvent.

n-butane, n-heptane) is brought out as a light component and

propylene (or 1-butene, benzene) as a heavy component (although the boiling point of n-heptane is higher than that of
benzene). That is, propane (or n-butane, n-heptane) would be
obtained as the overhead product in extractive distillation column, which is propylene (or 1-butene, benzene) and the solvent, the bottom product.

The system of n-heptane and benzene

It is assumed that n-heptane and benzene are taken on as the
representatives of nonaromatics and aromatics, respectively,
because aromatics and nonaromatics are mainly composed of
them and their separation mechanism is consistent. Extractive
distillation is used here for the separation of n-heptane and
benzene, and the solvents are screened by means of CAMD.
The restrictions for solvents in CAMD are listed as follows:
Preselected group types: CH3, CH2, CH3COO, CH3CO,
ACOH, H2O, CH3OH, (CH2OH)2, acetonitrile (ACN), N,Ndimethylformamide (DMF), N-methyl-2-pyrrolidone (NMP),
morpholine (Morph), furfural, dimethyl sulfoxide (DMSO)
Expected group number: 1 6
Maximum molecular weight: 150
Minimum boiling point: 323.15 K
Maximum boiling point: 503.15 K
Temperature: 303.15 K
Minimum relative volatility at innite dilution: 5.0
Minimum solubility capacity: 0.10

Case Studies
The systems of propane/propylene, n-butane/1-butene, and
n-heptane/benzene are investigated because they are commonly
encountered in industry and may be separated by extractive
The separation mechanism for separating hydrocarbons by
extractive distillation is based on the different uidities of the
electron cloud of COC (no double bond), CAC (one double
bond), and ACH (aromatic carbon ring) bonds. The greater the
uidity, the easier can the group be polarized by the polar
solvent. The uidity of the electron cloud of CAC (or ACH)
bond is greater than that of the COC bond so that propane (or

December 2005

Figure 1. CAMD program.

Vol. 51, No. 12

AIChE Journal

Table 2. Results of CAMD for the Separation of n-Heptane (1) and Benzene (2) at T 303.15 K
Azeotropic Judgment

Molecular Structure


T b (K)




1 and Solvent

2 and Solvent










By means of CAMD, the design results are obtained and

listed in order of decreasing relative volatility at innite dilution as shown in Table 2. Note that molecules 3, 5, 7, 9, and 10
are bought in chemical markets only with difculty (or bought
only at very expensive prices). From this perspective, they are
excluded as the potential solvents. Furthermore, after considering such implicit properties as toxic (excluding molecules 2
and 4), boiling point (excluding molecule 8), and chemical
stability (excluding molecule 1), only molecule 6, NMP, remains. NMP has many outstanding advantages as the solvent,
such as nontoxicity, noninjurious, facility of ecological treatment, and high separation ability.
The reliability of NMP as a solvent has already been veried
by our experiments.2 However, CAMD goes a step further to
provide a theoretical foundation for the selection of NMP. This
also indicates that the designed results from CAMD are acceptable in this regard.
Process simulation of the separation of n-heptane and benzene by extractive distillation was made with PRO/II software.
The process consisted of extractive distillation column with 30
theoretical plates and a solvent recovery column with 15 theoretical plates. The feeding mixture was composed of 45.0 wt
% n-heptane and 55.0 wt % benzene with a ow rate 100 kg/h,
and the feed/solvent volume ratio was 1:3. The extractive
distillation column was operated at pressure 101 kPa and reux
ratio 0.5, whereas the solvent recovery column was operated at
pressure 10 kPa and reux ratio 1.0. When the concentration of
benzene reached 99.5 wt % at the top of solvent recovery
column, the yield ratio of benzene was 94.2%.

The system of propane (1)/propylene (2)

Because most of the propylene comes from pyrolysis gases,
the separation of propylene in ethylene projects has great
commercial signicance. In ethylene projects, propylene is
puried from a mixture mainly composed of propylene and
propane. It is known that the pure-component boiling points are
very close over a large range of pressure. However, in industry
they are currently separated by ordinary distillation at pressures
of 1910 2000 kPa with the number of theoretical trays at about
120 and reux ratio up to 1520. Extractive distillation is here
used for the separation of propane and propylene, and the
solvents are screened by means of CAMD. The restrictions for
solvents in CAMD are listed as follows:
Preselected group types: CH3, CH2, CH3COO, CH3CO,
AIChE Journal

December 2005

acetonitrile (ACN), N,N-dimethylformamide (DMF), N-methyl-2-pyrrolidone (NMP), morpholine (Morph), furfural

Expected group number: 1 6
Maximum molecular weight: 150
Minimum boiling point: 323.15 K
Maximum boiling point: 503.15 K
Temperature: 303.15 K
Minimum relative volatility at innite dilution: 1.45
Minimum solubility capacity: 0.10
For the additive added to a main solvent, its concentration
is 10% in mass fraction
By means of CAMD, the design results are obtained and
listed in order of decreasing relative volatility at innite dilution as shown in Table 3. It seems that for single solvent, the
solvent with molecular structure of CH2CNOCH2CN (no. 1)
has the largest relative volatility of propane to propylene,
although it has strong toxicity. Therefore, ACN (no. 2) is
regarded as the potential solvent. For the solvent mixture, the
separation ability of the mixture of ACN and water (10 wt %)
is a little lower than that of the mixture of NMP and water (10
wt %). However, the boiling point of NMP (Tb 477.2 K) is
much higher than that of ACN (Tb 354.8 K). The extractive
distillation process must be operated at high pressures because
of the low boiling points of propane and propylene, which
easily leads to decomposition of NMP. On the other hand, the
separation ability of the mixture of ACN and water is higher
than that of single ACN. Therefore, it may be advisable to
select the mixture of ACN and water as the solvent for separating propane and propylene. One must be extremely cautious
in the process design because ACN tends to hydrolyze, which
contributes to the loss of ACN and corrosion of equipment
during production.
The experimental results obtained by the methods of inert
gas stripping and gas chromatography are also given in parentheses in Table 3. Details of the apparatus and experimental
procedure used in this work have been described in a previous
publication.35 It can be seen that the mixture of ACN and water
can apparently improve the relative volatility of propane to
propylene, and 12 is far away from unity. It is generally
thought that because 12 1.20, extractive distillation could
be adopted for separating the components with close boiling
points or forming azeotropes. From an economic perspective,
the use of the solvent with the highest relative volatility will
always give the lowest total annual cost of extractive distillation process.
Vol. 51, No. 12


Table 3. Results of CAMD for the Separation of Propane (1) and Propylene (2) at T 303.15 K
Azeotropic Judgment

Molecular Structure


Single solvent
Additive added to ACN to make into a mixture (10 wt %)
Additive added to NMP to make into a mixture (10 wt %)





1 and

2 and


1.704 (1.620)






1.846 (1.750)











Process simulations of the separation of propane and propylene by ordinary distillation and by extractive distillation
with the mixture of ACN and water as the solvent were respectively made with PRO/II software. The ordinary distillation column was operated at pressure 1910 kPa and reux ratio
15.0 with 120 theoretical plates. The extractive distillation
process consisted of four sections: extractive distillation column with 50 theoretical plates; stripping column with 10 theoretical plates; water scrubber column with 10 theoretical
plates; and solvent recovery column with 20 theoretical plates.
In the case of the feeding composition of propylene 82.0 wt %,
feed rate 32,000 kg/h, production rate 26,500 kg/h, and molar
composition of product propylene 99.3%, the extractive
distillation process saved 13.2% reboiler load, 79.1% condenser load, and 25% theoretical plates when compared to
ordinary distillation process.

The system of n-butane (1)/1-butene (2)

In the production of ethylene and/or propylene by pyrolysis
of liqueed petroleum gas or naphtha, C4 mixtures are produced. In general, C4 mixtures mainly contain n-butane, isobutane, isobutene, 1-butene, trans-2-butene, cis-2-butene, 1,3butadiene, and vinylacetylene (VAC). Earlier, we paid more
attention to extracting 1,3-butadiene, which is an important raw
material, from C4 mixtures. However, among the residua,
n-butene (that is, the mixture of 1-butene, trans-2-butene, and
cis-2-butene) is able to be used as one monomer of synthesizing polymers and one reactant of producing acetone by reacting
with water. Isobutene can be selectively combined with methanol to make methyl tert-butyl ether (MTBE). Unfortunately,
the separation of butane and butene has rarely been studied.
Similarly, extractive distillation is here used for the separation
of butane and butene, and the solvents are screened by means
of CAMD. n-Butane and 1-butene are regarded as the key
components. The restrictions for solvents in CAMD are the
same as in the system of propane (1)/propylene (2).
By means of CAMD, the design results are obtained and
listed in order of decreasing relative volatility at innite dilu3118

December 2005

tion as shown in Table 4. For the solvent mixture, after such

factors as source, toxic, and miscibility are taken into consideration, the rst potential solvent is the mixture of ACN and
water that are totally miscible, whereas the second potential
solvent is the mixture of ACN and ethylenediamine (no. 18)
that are also totally miscible.
Process simulation of the separation of n-butane and
1-butene by extractive distillation with the mixture of ACN and
water as the solvent was performed with PRO/II software. The
process consisted of extractive distillation column with 70
theoretical plates and solvent recovery column with 30 theoretical plates. The feeding mixture was composed of 6.1 wt %
n-butane and 93.9 wt % 1-butene with a ow rate 2450 kg/h,
and the feed/solvent mass ratio was 1:10. The extractive distillation column was operated at pressure 500 kPa and reux
ratio 5.0, whereas the solvent recovery column was operated at
pressure 450 kPa and reux ratio 4.0. When the concentration
of 1-butene reached 99.0 wt % at the top of solvent recovery
column, the yield ratio of 1-butene was above 95.0%.
Table 5 gives the number of molecules assembled with two
to six groups from eight kinds of different groups (excluding
molecular groups) for single solvents. It shows that if without
any constraint, incalculable molecules would be formulated.
After the constraints of steps 2 and 3 in CAMD, however, the
number of chemically feasible molecules is greatly reduced,
and thus the potential solvents can be rapidly screened.

Thermodynamic Analysis
By comparison of Tables 3 and 4, it is found that from
numbers 1 to 13 the solvents screened by CAMD and their
order in both systems is consistent. The reason may be explained by the similar molecular structure of C3 and C4 hydrocarbons (a difference of only one carbon chain) and the
same separation mechanism.
The experimental results obtained by inert gas stripping and
gas chromatography are also given in parentheses in Table 4.
However, it is found that the mixture of ACN and ethylenediVol. 51, No. 12

AIChE Journal

Table 4. Results of CAMD for the Separation of Butane (1) and Butene (2) at T 303.15 K
Azeotropic Judgment

Molecular Structure





1 and

2 and


1.684 (1.696)






1.824 (1.880)
1.640 (1.725)
1.631 (1.707)
1.626 (1.787)
1.579 (1.713)












Single solvent
Additive added to ACN to make into a mixture (10 wt %)
Additive added to NMP to make into a mixture (10 wt %)

amine also can give rise to higher relative volatility than single
ACN or single ethylenediamine.
Why do the mixtures of ACN and water and that of ACN and
ethylenediamine show good separation ability? This may be
explained from Prausnitz and Andersons solution thermodynamics.36,37
For the hydrocarbons to be separated by extractive distillation, selectivity (which is consistent with relative volatility) is
related to the various energy terms leading to the desired
nonideality of solution, which is the basis of extractive distillation, and can be expressed as

is, the polar effect, the dispersion effect, and the inductive
effect of the solvent. It is convenient to rewrite as
RT ln S23 P D I


where P 21p (V2 V3), D V2(1n 2)2 V3(1n 3)2,

and I 2V313 2V212.
In general, the polar term P is considerably larger than the
sum of D and I. Thus, Eq. 2 becomes
V2 V3
RT ln S23 1p


RT ln S23 V2 V3 V2 1n 2 V3 1n 3

2V3 13 2V2 12 (1)

where subscripts 1 to 3 represent solvent, the light component,
and the heavy component to be separated by extractive distillation, respectively; V is the molar volume, is Hansens
solubility parameter; and is induction energy per unit volume.
The three bracketed terms in Eq. 1 show, respectively, the
separate contributions of physical force to the selectivity, that

Table 5. Number of Molecules Assembled with Two to Six

Groups from Eight Kinds of Different Groups (Excluding
Molecular Groups) for Single Solvents
Without any
After step 2
After step 3

AIChE Journal

Case 1

Case 2

Case 3




December 2005

Equation 3 not only shows the effect of molecular size but

also predicts that when one separates hydrocarbons of different
molar volumes, the selectivity is sensitive to the polar solubility parameter. This indicates that the effectiveness of a solvent
depends on its polarity, which should be large, and on its molar
volume, which should be small.
In Tables 2 4, most of the solvents screened by CAMD are
of two polar functional groups, showing high polarity. In
addition, the inuence of molar volume is investigated for the
separation of n-butane (1) and 1-butene (2) with the solvents
whose molecular structures are ROC'N. It is assumed that
the polar functional group C'N is kept constant, but the
carbon number in the R group changes. In this case the change
of carbon number with relative volatility of n-butane to
1-butene at innite dilution is shown in Figure 2. It can be seen
that as the carbon number increases, the relative volatility
decreases. This conrms that the solvent with high polarity and
small molar volume will be desirable to produce a much higher
relative volatility.
Vol. 51, No. 12


Figure 2. Inuence of carbon number on relative volatility of n-butane to 1-butene at innite dilution
for the solvents whose molecular structures
are ROC'N: R CnH2[in]n1 (n carbon number 1, 2, 3, 4, 5, 6).

Besides, for a given separation task, the term (V2 V3) in

Eq. 3 cannot be changed. For the components with similar
molecular structures this value is even small. Therefore, to
improve the relative volatility, the only way is to increase 1p,
which can be fullled by introducing an additive into the main
solvent. The schematic representation of the molecular interaction among main solvent (ACN), additive (water or ethylenediamine), and hydrocarbon is illustrated in Figure 3.
It is conceivable that hydrogen bonding is formed between
molecules of ACN and ethylenediamine (or water) because
ACN is a hydrogen bond acceptor and ethylenediamine (or
water) is a hydrogen bond donor. This would result in the
improvement of 1p by a transfer of electron cloud and thus
intensify the interaction between the main solvent and the
hydrocarbon. Therefore, relative volatility will increase in
terms of Eq. 3. To verify this analysis, a set of inert gas
stripping and gas chromatography equipment has been built to
measure the relative volatility of n-butane (1) to 1-butene (2).
The experimental results at different additive concentrations in
the mixture of ACN and ethylenediamine are shown in Figure
4. It can be seen that the mixture of ACN and ethylenediamine
performs more satisfactorily than single ACN.

Figure 3. Molecular interaction among main solvent, additive, and hydrocarbon.

(a) ACN (main solvent), water (additive), and 1-butene system; (b) ACN (main solvent), ethylenediamine (additive), and
1-butene system.

For separating hydrocarbons by extractive distillation, the

separation mechanism is consistent based on the different uidities of electron cloud of hydrocarbon molecules. The solvent
with high polarity and small molar volume will be desirable.
To improve the separation ability of the main solvent, one

CAMD has been programmed to screen the solvents for
separating hydrocarbons by extractive distillation. The systems
of propane/propylene, n-butane/1-butene, and n-heptane/benzene are investigated because they are commonly met in industry and may be separated by extractive distillation. By
means of CAMD, it is found that for the separation of propane
and propylene, the nal potential solvent is ACN (no hydrolysis) or the mixture of ACN and water (hydrolysis); for the
separation of n-butane and 1-butene, the nal potential solvent
is ACN (no hydrolysis) or the mixture of ACN and water
(hydrolysis), although the mixture of ACN and ethylenediamine (no hydrolysis) also shows high separation ability; for the
separation of n-heptane and benzene, the nal potential solvent
is NMP.

December 2005

Figure 4. Inuence of concentration of ethylenediamine

in the mixture on relative volatility of n-butane
to 1-butene at innite dilution at 293.15 and
303.15 K.
Vol. 51, No. 12

AIChE Journal

strategy is to add some additive that can form hydrogen bonding with the main solvent to make into a mixture.
In this work the solvents of extractive distillation are restricted only to liquid solvents. In fact, so far they have been
developed to include inorganic salts, polymers, and ionic liquids. Unfortunately, the group parameters of polymer and ionic
liquid are vacant in the parameter table except that there are a
few group parameters for inorganic salts. Even for liquid
solvents, there are still about 54% of the UNIFAC group
interaction parameters missing because experimental data are
needed to ll them. This is where CAMD is limited. For the
development of calculation techniques, it is thought that, besides the efforts to blend the computation methods of target
properties with the screening of potential solvents and replenish the group parameters, were CAMD combined with other
software (such as Excel, PRO/II, ASPEN PLUS), its functions
would become stronger and the designed results would be more


This work is nancially supported by the National Nature Science
Foundation of China under Grant 20406001.


Literature Cited


1. Berg L. Separation of benzene and toluene from close boiling nonaromatics by extractive distillation. AIChE J. 1983;29:961-966.
2. Chen B, Lei Z, Li J. Separation on aromatics and non-aromatics by
extractive distillation with NMP. J Chem Eng Jpn. 2003;36:20-24.
3. Lei Z, Li C, Chen B. Extractive distillation: A review. Sep Purif Rev.
4. Joback KG, Stephanopoulos G. Searching spaces of discrete solutions:
The design of molecules possessing desired physical properties. Adv
Chem Eng. 1995;21:257-311.
5. Venkatasubramanian V, Chan K, Caruthers JM. Genetic algorithmic
approach for computer-aided molecular design. ACS Symp Ser. 1995;
6. Churi N, Achenie LEK. Novel mathematical programming model for
computer aided molecular design. Ind Eng Chem Res. 1996;53:37883794.
7. Marcoulaki EC, Kokossis AC. Molecular design synthesis using stochastic optimisation as a tool for scoping and screening. Comput Chem
Eng. 1998;22:S11-S18.
8. Mavrovouniotis ML. Design of chemical compounds. Comput Chem
Eng. 1998;22:713-715.
9. Meniai AH, Newsham DMT, Khalfaoui B. Solvent design for liquid
extraction using calculated molecular interaction parameters. Chem
Eng Res Des. 1998;76(A11):942-950.
10. Ourique JE, Telles AS. Computer-aided molecular design with simulated annealing and molecular graphs. Comput Chem Eng. 1998;22:
11. Pistikopoulos EN, Stefanis SK. Optimal solvent design for environmental impact minimization. Comput Chem Eng. 1998;22:717-733.
12. Raman VS, Maranas CD. Optimization in product design with properties correlated with topological indices. Comput Chem Eng. 1998;
13. Harper PM, Gani R, Ishikawa T, Kolar P. Computer aided molecular
design with combined molecular modeling and group contribution.
Fluid Phase Equilib. 1999;158-160:337-347.
14. Hostrup M, Harper PM, Gani R. Design of environmentally benign

AIChE Journal

December 2005


processes: Integration of solvent design and process synthesis. Comput

Chem Eng. 1999;23:1395-1414.
Sinha M, Achenie LEK, Ostrovsky GM. Environmentally benign
solvent design by global optimization. Comput Chem Eng. 1999;23:
Franklin JL. Prediction of heat and free energies of organic compounds. Ind Eng Chem Res. 1949;41:1070-1076.
Gani R, Brignole EA. Molecular design of solvents for liquid extraction based on UNIFAC. Fluid Phase Equilib. 1983;13:331-340.
Gani R, Nielsen B, Fredenslund A. A group contribution approach to
computer-aided molecular design. AIChE J. 1991;37:1318-1332.
Gani R, Fredenslund A. Computer aided molecular and mixture design
with specied property constraints. Fluid Phase Equilib. 1993;82:3946.
Pretel EJ, Lopez PA, Bottini SB, Brignole EA. Computer-aided molecular design of solvents for separation processes. AIChE J. 1994;40:
Jorgensen SS, Kolbe B, Gmehling J, Rasmussen P. Vaporliquid
equilibria by UNIFAC group contribution. Ind Eng Chem Process Des
Dev. 1979;18:714-722.
Gmehling J, Rasmussen P, Fredenslund A. Vaporliquid equilibria by
UNIFAC group contribution. 2. Revision and extension. Ind Eng
Chem Process Des Dev. 1982;21:118-127.
Macedo EA, Weidlich U, Gmehling J, Rasmussen P. Vaporliquid
equilibria by UNIFAC group contribution. 3. Revision and extension.
Ind Eng Chem Process Des Dev. 1983;22:676-678.
Tiegs D, Gmehling J, Rasmussen P, Fredenslund A. Vaporliquid
equilibria by UNIFAC group contribution. 4. Revision and extension.
Ind Eng Chem Res. 1987;26:159-161.
Hansen HK, Rasmussen P, Fredenslund A, Schiller M, Gmehling J.
Vaporliquid equilibria by UNIFAC group contribution. 5. Revision
and extension. Ind Eng Chem Res. 1991;30:2352-2355.
Gmehling J, Li J, Schilleer M. A modied UNIFAC model. 2. Present
parameter matrix and results for different thermodynamic properties.
Ind Eng Chem Res. 1993;32:178-193.
Gmehling J, Lohmann J, Jakob A, Li J, Joh R. A modied UNIFAC
(Dortmund) model. 3. Revision and extension. Ind Eng Chem Res.
Lohmann J, Gmehling J. Modied UNIFAC (Dortmund): Reliable
model for the development of thermal separation processes. J Chem
Eng Jpn. 2001;34:43-54.
Gmehling J, Wittig R, Lohmann J, Joh R. A modied UNIFAC
(Dortmund) model. 4. Revision and extension. Ind Eng Chem Res.
Wittig R, Lohmann J, Gmehling J. Vaporliquid equilibria by UNIFAC group contribution. 6. Revision and extension. Ind Eng Chem
Res. 2003;42:183-188.
Momoh SO. Assessing the accuracy of selectivity as a basis for solvent
screening in extractive distillation processes. Sep Sci Technol. 1991;
Tong JS. The Fluid Thermodynamics Properties. Beijing: Petroleum
Technology Press; 1996.
Poling BE, Prausnitz JM, OConnell JP. The Properties of Gases and
Liquids. 5th Edition. New York, NY: McGraw-Hill; 2001.
Joback KG, Reid RC. Estimation of pure-component properties from
group-contributions. Chem Eng Commun. 1987;57:233-243.
Lei Z, Wang H, Zhou R, Duan Z. Inuence of salt added to solvent on
extractive distillation. Chem Eng J. 2002;87:149-156.
Prausnitz JM, Anderson R. Thermodynamics of solvent selectivity in
extractive distillation of hydrocarbons. AIChE J. 1961;7:96-101.
Lei Z, Chen B, Ding Z. Special Distillation Processes. Amsterdam:
Elsevier; 2005.

Manuscript received Dec. 3, 2004, and revision received Mar. 27, 2005.

Vol. 51, No. 12


Rate Constants of Mass Transfer Kinetics in

Reversed Phase Liquid Chromatography
Lan Hong, Fabrice Gritti, and Georges Gulochon
Dept. of Chemistry, The University of Tennessee, Knoxville, TN, 37996 and Division of Chemical Sciences,
Oak Ridge National Laboratory, Oak Ridge, TN, 37831

Krzysztof Kaczmarski
Faculty of Chemistry, Rzeszow University of Technology, 35-959, Rzeszow, Poland
DOI 10.1002/aic.10565
Published online September 2, 2005 in Wiley InterScience (www.interscience.wiley.com).

The parameters of the kinetics of mass transfer of several n-alkylbenzenes were

measured by the method of moments on a series of columns prepared with different
samples of the same RPLC packing material having widely different average particle
diameters, from 3 to 50 m. These data were analyzed using the available models, and
correlations. The best agreement between experimental and theoretical data was obtained
under the assumption that the rate constant for the external mass transfer increases with
increasing average particle size, an unexpected conclusion. It was also shown that the
interpretation of the relative importance of the roles of pore diffusivity and surface
diffusivity in the internal mass transfer kinetics is somewhat ambiguous and that the
conclusion to be drawn from experimental results depends on the assumptions made
regarding the tortuosity model and the relationship between kext and the average particle
size. 2005 American Institute of Chemical Engineers AIChE J, 51: 31223133, 2005
Keywords: liquid chromatography, mass transfer kinetics, method of moments, pore
diffusion, surface diffusion

Liquid chromatography has become an important process for
the extraction and/or the purication of drug intermediates in
the pharmaceutical industry. It is of special interest for fragile
molecules, for those that are obtained as components of very
complex mixtures of similar compounds, and for difcult separations. Typical examples are the preparation of peptides,
proteins, and enantiomers. In all these instances, the interest for
the modeling of specic HPLC separations has much increased
during the last few years because computer-assisted modeling
of processes permits a considerable reduction in the time and
manpower needed to start a process or to start up a unit after
Correspondence concerning this article should be addressed to G. A. Gulochon at

2005 American Institute of Chemical Engineers


December 2005

maintenance and brings signicant economies in the design and

operation costs. Process modeling is further encouraged by the
strategic initiative of the Food and Drug Administration that
will impact on the current Good Manufacturing Practices
(cGMP) and will much affect chemical engineering applications in the pharmaceutical industry by requiring a better understanding of the processes involved.1
The modeling of the separation processes encountered in the
various implementations of preparative chromatography, including simulated moving bed processes, requires an accurate
knowledge of the specic equilibrium thermodynamics and
equilibration kinetics involved.2 The nature of the data required, the procedures to follow to acquire them, and the
measurements to perform in order to obtain an accurate, reliable model of the competitive equilibrium isotherms of the feed
components of importance are well known.2,3 Different procedures allow the acquisition of data of increasing accuracy at an
Vol. 51, No. 12

AIChE Journal

increasing cost in time spent and chemicals consumed.4 Current research aims at reducing these costs but will not change
the need for these data. If the relationship between equilibrium
isotherms, the calculation of elution band proles, and the
modeling of process performance is clearly established, the
inuence of the mass transfer kinetics on band proles and
process performance remains insufciently understood.
For its better understanding, the kinetics of equilibration
between the uid and the solid phases in chromatography is
divided into several steps that can be grouped into two broad
classes, axial dispersion and the mass-transfer resistances. The
former class involves axial diffusion, the consequence through
Fick laws of the concentration gradient that migrates along the
column, and eddy diffusion, which arises from the ow pattern.
The chromatographic columns are operated under a rather slow
stream of the uid phase. In liquid chromatography, the Reynolds numbers of these ows are typically between 0.01 and
0.001, ensuring creeping ow conditions.5 There are no significant eddies but the distribution of the average residence times
along the different stream-lines has a signicant width, causing
eddy diffusion.2,5 The contributions of axial dispersion to band
broadening tend to decrease with increasing ow velocity.
By contrast, the effects of the mass-transfer coefcients
increase with increasing mobile phase velocity. They include
mass transfer between the uid phase stream percolating
through the bed and the uid stagnant inside the particles (or in
the porons, in the case of monolithic columns), diffusion
through this stagnant uid across the particles, and the kinetics
of adsorption desorption onto the solid surface. The rst two
of these contributions are called the external and the internal
mass transfer resistances, respectively. The latter contribution
has been abundantly studied and our recent results have demonstrated the importance of surface diffusion in reversed-phase
liquid chromatography.6 The external mass transfer resistance
has been much less investigated. It is usually estimated in
kinetic studies from the Wilson and Geankoplis correlation.7
This remarkable equation was derived, however, from measurements carried out on a column packed with 1/4 solid beads
of benzoic acid percolated with a stream of water. The validity
of this equation for columns packed with silica particles that
are a thousand times smaller and exchange solutes with the
external uid through a multitude of tiny pores has never been
The goal of this work was to investigate the external mass
transfer kinetics by determining, under linear conditions, its
rate constant for a series of compounds on columns packed
with the same packing material, but using lots having average
particle sizes spanning a wide range, between 3 and 50 m,
hence, under experimental conditions spanning a 100-fold
range of the Fick number. The dependence of the external rate
constant on the particle size should provide new clues on the
mass transfer-kinetics in chromatography.

General rate model (GR)
Any serious investigation of the inuence of the masstransfer kinetics on the band proles in chromatography must
use the general rate model of chromatography. Although there
are no algebraic, only numerical solutions for this model, a
solution is available for the Laplace transform in the linear
AIChE Journal

December 2005

case. This solution cannot be inverted into the time domain.

However, its moments can be derived using simple algebraic
manipulations. This approach constitutes the one that was
followed in this work.
The following modied general rate (GR) model was used.8
It consists of two equations.
A mass transport equation for the mobile phase percolating
through the bed

eD L

1 3 k

C C p,i r R i



A mass transport equation for the liquid stagnant in the pores

of the solid phase

q i
C p,i
C p,i
1 p
D eff 2
r r


In these equations, C and C p are the concentrations of the

solute in the bulk liquid stream outside the particle and in the
stagnant liquid phase in the pores, respectively, q is the concentration of the adsorbed component, t is the time, r is the
distance from the particle center, R is the particle radius, e is
the external porosity, p is the internal or particle porosity, and
D eff is the effective diffusivity coefcient. Subscript i denotes
the i-th fraction of the particle-diameter distribution (see later),
and the external porosity ei is a hypothetical external porosity
when only the i-th fraction exists, calculated by use of the
e,i 1 1 e



where V i is the volume of the i-th fraction and V a is the total

volume of adsorbent.
The GR model (Eqs. 1 and 2) was coupled with typical
boundary and initial conditions, assuming Danckwerts boundary conditions for Eq. 1, continuity of the mass transfer on the
outer surface of the particle, and a symmetrical concentration
distribution inside the particles. It was also assumed that the
initial concentration of each experiment (elution) was equal to

Theory of second moment analysis

The chromatographic band proles were analyzed using the
conventional method of moments.2,6,9 13 Information on thermodynamics of equilibrium between the mobile and the stationary phase and on the mass-transfer kinetics between the
two phases of the system is obtained from the rst absolute
moment and the second central moment, respectively. The rst
absolute moment (1), and the second central moment (2 ) of
the chromatographic peaks are expressed as follows

Vol. 51, No. 12



Ctt 1 2dt


where C(t) is the chromatographic band prole, L is the length

of the column, and u 0 the supercial mobile phase velocity.
The second central moment is related to the different kinetic
characteristics of the column, as follows


2u 01 e P 1 P K 2 R

3k ext 15D eff


1 e P 1 P K 2

eD L 02
u 02



f d

3k ext 15D eff

f 1 eR/ 2k ext p 1 p K 2


d 1 eR 2/15D eff P 1 p K 2


0 e 1 e P 1 P K




L L 2


2u 0 e
L f d




The HETP used in conventional chromatography is reported in

length units. However, the parameter dened in this work H is
reported in time units.
It should be noted that the equilibrium constant, which is
included in the above equations, can be easily derived from the
experimental parameters and is given by

o e

1 e1 p 1 p


Theoretical calculation of the molecular diffusivity, the

dispersion coefcient and the external mass transfer
In this work all the mass transfer kinetic parameters and the
dispersion coefcient parameter were determined experimentally. The obtained results were compared with theoretical
calculation of these parameters based on probably most frequently used correlations.
The External Mass-Transfer Coefcient. The external
mass-transfer coefcient was calculated from the generally
accepted correlation, the Wilson-Geankoplis equation7 valid
according to these authors for 0.0015 Re 55. This includes
the range under which liquid chromatography is carried out.

This equation relates the Sherwood number Sh, a function of

the external mass transfer coefcient, to the Schmidt (Sc) and
the Reynolds (Re) numbers. These numbers are dened by


By applying Eq. 7 to 9 to Eq. 13 we obtain the form

December 2005

eD L
2 L
2 H0
12 2u 0

Sh 1.09/ eSc 1/3Re1/3

By applying Eqs. 4 to 6 to Eq. 12, the HETP is expressed as



Both A and H o are independent of the mobile phase ow

velocity, insofar as the dependency of k ext on the mobile phase
velocity can be neglected in the increment investigated. Now,
Eq. 12 can be rewritten as follows

where ( 2 ) inj and ( 2 ) sys denote the contributions to the

second central moment arising from the injection of the probe
compound into the stream of mobile phase and from dispersion
in the extra-column void volumes of the instrument, respectively. L , d , and f are the contributions to 2 of axial
dispersion, the uid-to-particle mass-transfer resistance, and
intraparticle diffusion, respectively. The other parameters in
these equations are: e , the external (bed) porosity p , the
internal (particle) porosity, R, the particle radius, k ext , the
external mass transfer coefcient, D eff , the coefcient of effective diffusivity, and K q/C, the equilibrium constant. The
concentrations in the uid phase, C, and at the adsorbent
surface, q, are referenced to the uid volume and to the
adsorbent matrix volume, respectively.
Equation 6 shows that the contributions to the second moment of the band of each source of mass transfer resistance are
separate and additive. The rst and the second moments are
simply related to the number of theoretical plates N, of the
column, and to its high equivalent to a theoretical plates
(HETP), H, namely

For the sake of simplication, we dene the following auxiliary

parameters, A and H 0

f d 2 inj 2 sys
u0 L


2 eD L

Vol. 51, No. 12

Re ud/


Sc / D m

AIChE Journal

Sh k extd/D m

From the correlation Eq. 19, it follows that the external mass
transfer rate constant is inversely proportional to the particle
diameter to the power (2/3).
k ext

1.09u 0 1/3 D m



In the earlier equation, is the mobile phase viscosity and its

density. This relationship permits the calculation of k ext knowing the molecular diffusivity, D m , of the solute in the mobile
phase. Because the exact value of D m is seldom available,
however, it must be estimated. There are several different
approaches available. The one most frequently used in the
HPLC literature is the Wilke-Chang Eqs. 2, 12, 14, which is
claimed to be accurate within 10% for small to medium-size
molecules. Its formulation is
D m 7.4 10 8T

AM s 0.5
V 0.6


where T is the absolute temperature, V A is the molar volume of

the solute at its normal boiling point (calculated from LeBas
correlation), M s is the molecular weight of the uid phase, is
the uid viscosity (in cP), and A is the association factor for
the uid which accounts for the solute-solvent interactions.
This factor is 1.9 for methanol and 2.6 for water.
Dispersion Coefcient. The dispersion coefcient was calculated from the Gunn15 equation

eD L
B1 2 2
du 0
Re Sc



Re Sc2
Re Sc
1 p2
4 1 e
1614 1 e 2

p1 p3 exp

4 1 e
p1 pRe Sc

p 0.17 0.33 exp24/Re


with Pe being the particle Peclet number (Pe ud/D L e ), and

Re given by Eq. 20.

Particle-size distribution
A relationship can be derived easily between the rst two
moments of the elution bands, the column HETP under linear
conditions, and the parameters of the general rate model if the
column bed is assumed to consist in spherical particles having
all the same diameters. However, all actual packing materials
have a particle-size distribution (PSD) of nite width. Typical
examples of the PSD of contemporary packing materials and
adsorbent particles are discussed by Neue.17 For practical calculations, the PSD is replaced by the mean-particle diameter.
There exist several methods to calculate the mean diameter of
a particulate material, but for the description of the mass
transfer processes, such as those that take place in chromatography, the volume moment mean diameter, noted d 43 , is recommended17

d 43

i n id i4
i n id i3


In this equation, n denotes the number of particles of size d,

and the subscript i denotes the i-th fraction of the distribution.
Recently it was proved8 that the band proles calculated from
the PSD and those derived from the volume moment mean,
d 43 , for different isotherm models and for the particle dia. of 5
[m], the most frequently used in HPLC, are practically identical.
Since in this work packing materials with average particle
diameters between 3 and 50 [m] were used, we decided to
check whether the inuence of the PSD on the rst and second
moments was signicant (see later).


where 1 is the rst root of the zero-order Bessel function, is

the bed package tortuosity factor (assumed to be equal to 1.4),
and v2 is the dimensionless variance of the distribution of the
ratio between the local uid linear velocity and the average
velocity over the column cross-section. In this work, this variance was assumed to be equal to zero. Finally, p is a parameter
dened by

The other popular correlation between the ow velocity of the

stream percolating through the bed and the axial dispersion
coefcient is the following equation elaborated by Chung and
AIChE Journal

ePe 0.2 0.011 Re0.48


December 2005

An HP 1090 (Hewlett-Packard, Palo Alto, CA) liquid chromatograph was employed for all the experimental determinations. The instrument is equipped with a multisolvent delivery
system, an automatic sample injector with a 100-L loop, a
diode array UV-detector and a computer data station that
controls its operation. The signal of the UV detector was
acquired at a wavelength of 254 nm.

Mobile phase and chemicals

The mobile phase used was a mixture solution of methanol
and water (70:30 v/v). Methanol and water were HPLC grade
solvents from Fisher Scientic (Fair Lawn, NJ). Thiourea was
purchased from Sigma Chemicals (St. Louis, MO). The solutes
studied, toluene, ethylbenzene, propylbenzene, butylbenzene,
and amylbenzene were purchased from Aldrich Chemical (Milwaukee, WI).
Vol. 51, No. 12


Table 1. Characteristics of the Columns Used


3 [m]

5 [m]

10 [m]

15 [m]

50 [m]

Mean particle size. Average from several particle

size analyses [m]
Error of mean particle size evaluation
Particle distribution, 90/10%a
Pore diameter,
Surface area [m2/g]
Total carbon, %
Surface coverage [mol/m2]
Total porosityb
External porosityb
Internal porosityb








Ratio of the diameters below which 90% and 10% of the particles, respectively, are found.
Values measured in our laboratory.

Chromatographic conditions
Sets of identical experiments were carried out under different ow rates on a series of stainless steel columns (all 150
4.6 [mm]). Five columns were packed by the manufacturer
with Luna C18 (Phenomenex, Torrance, CA) in different average particle sizes, 3, 5, 10, 15, and 50 [m]. This packing
material is made of spherical particles of porous silica, chemically bonded with octadecylsilane. Their chemical properties
are broadly similar as far as alkylaromatics are concerned. The
details of the column parameters, as supplied by the manufacturers, are shown in Table 1. These columns are similar to one
that was investigated by Kele18 but they belong to more recent
lots and most of them have different particle sizes. The column
temperature was kept constant at 25.0C.
To achieve accurate results, the column was equilibrated for
at least three hours prior to any measurement when the mobile
phase composition was changed. The run-to-run reproducibility of the instrument was characterized by a RSD of the peak
area which was better than 0.50% (six replicates) for an injection volume of 5 L. The rst and second moments of the
elution bands of the solutes were measured for the following
mobile phase ow rates: 0.8, 1.0, 1.2, 1.5, and 1.8 [mL/min] for
average particle sizes of 5, 10, and 15 [m]; 0.6, 1.0, 1.2, and
1.4 [mL/min] for an average particle sizes of 3 [m]; and 1.2,
2.0, 2.4, and 2.8 [mL/min] for an average particle size of 50
The total porosity t , of each column was derived from the
retention volume of thiourea, that was considered to be unretained on all the columns, an assumption consistent with the
results previously obtained by this group under the experimental conditions used here. The external porosity e , was obtained
from measurements carried out under ISEC mode (see later,
section 3.6). The internal porosity P , was calculated according
to the conventional equation t e (1 e ) p . The
porosities are also listed in Table 1.

Selection of solutes and columns

Toluene, ethylbenzene, propylbenzene, butylbenzene, and
amylbenzene were tested with different columns. It was found
that ethylbenzene, propylbenzene, and butylbenzene, eluted on
columns packed with particles having dia. of 3, 5, 10, 15, and
50 [m], could provide peaks of suitable width for an accurate
measurement of the second moment under the experimental
chromatographic conditions required. The peak of toluene was

December 2005

too narrow and the peak of amylbenzene had too high a

retention time, more than 60 min. The use of either solute
would have resulted in excessively large errors in the measurement of their second moments.

Procedures for moment analysis

In this work, all the measurements reported were carried out
under linear chromatographic conditions. The lowest sample
size that the sampling system of the instrument could deliver (5
[L]) was used. Dilute solutions (ca. 130 300 ppm, v/v) were
used for the measurements. Nevertheless, the signal to noise
ratio was sufcient to afford precise values of the peak moments. The corresponding elution peaks were recorded, and the
rst and second moments were obtained from the reports
generated by the program of the HP ChemStation for LC-3D.
The start and end points of the peaks were automatically
selected by the parameter of slope sensitivity. The accuracy
and precision of the moments measured by the Chemstation has
been previously assessed, using the same instrument.19
The values of the extra-column volumes between the injection valve, and the column and between the column and the
detector were derived from the retention volumes of the solute
measured on the instrument with the column being replaced
with a zero-volume connector. The contributions of ( 2 ) inj and
( 2 ) sys to the second moment 2 , of the solute were calculated
by measuring the second moment of the solute on the system
with a zero-volume connector, and without column at each
ow rate and subtracting it from the second moment of the

Determination of the external porosity of the column via

The external porosity of a column is conveniently evaluated
using the method of inverse size exclusion chromatography
(ISEC).20,21 As commonly understood, conventional SEC permits the separation of molecules according to their size. It uses
as the stationary phase a porous material of known pore-size
distribution, under experimental conditions (that is, with a
mobile phase) such that the sample components do not adsorb
on the pore surface of the packing material. There is a general
relationship between the size of a molecule, and the size of the
pores from which it is excluded. In ISEC, the converse principle is exploited. The pore-size distribution of an adsorbent is
Vol. 51, No. 12

AIChE Journal

Results and Discussion

Inuence of the particle-size distribution on the second

Figure 1. Logarithm of the molecular mass of the samples vs. their retention volumes for a series of
polystyrene standards.

derived from the retention volumes of standard samples of

known molecular mass. More simply, in the present case, the
external and internal porosities are derived from the experimental relationship between the molecular mass of the polymeric compounds used and the average size of the pores from
which they are excluded. Accordingly, estimate of the poresize distribution of the packing material studied, and of its
external porosity can be derived from this plot. The successful
application of ISEC requires that a few simple conditions be
fullled:21 there should be no adsorption of the polymeric
samples on the stationary phase; the standard molecules should
not aggregate; all their chains should remain in the same
conformation; there should be instantaneous equilibrium between the two phases of the system during the whole experiment; the stationary phase should remain rigid (that is, should
not swell). Experimental investigations showed that tetrahydrofuran or methylene chloride are ideal for performing such
measurements on RPLC packing materials, using polystyrene
Figure 1 shows a typical result obtained, the plot of the
logarithm of the molecular mass of the samples vs the retention
volumes of a series of polystyrene standards (with molecular
masses between 1,780 and 3,840,000). Samples of 10 L of a
1 g/L solution were injected for each polystyrene standard. The
chromatograms were recorded at 265 nm. Two linear regression lines (solid lines) identify two specic regions in the
pore-size distribution corresponding to the interparticle, and
the intraparticle pore volumes, respectively. The intersection of
these two linear regression lines gives the value of the external
pore volume, which allows the calculation of the external
The values obtained for the external column porosities are
summarized in Table 1, together with the values calculated for
the internal porosities. The column-to-column uctuations of
the results reect a combination of the precision of the measurements, and the reproducibility of the packing procedure.
AIChE Journal

December 2005

Inuence of the Mean Particle Size. The actual mean particle size can differ by as much as 10% from the nominal value,
as seen in Table 1. To gain an accurate knowledge on how the
value of the particle-mean diameter can inuence the second
moment, numerical solutions of the GR model were calculated
for particle diameters equal to 5 and 5.3 [m], on the one hand,
to 50 and 53.5 [m], on the other hand. The mobile phase
velocity was assumed to be equal 1 [ml/min]. The other GR
model parameters (the porosities, the dispersion coefcient,
and the mass-transfer coefcients) were taken from those derived from our experimental data and which are reported later
in this work or can be found in Table 1.
According to the calculated peak proles, the relative difference between the second moment calculated for the two
pairs of actual particle diameters was derived according to the

2,d2 2,d1


The indices d1 and d2 in Eq. 30 denote the smaller and the

higher particle diameters, respectively. The relative difference
for dia. of 5 and 5.3 [m] (a relative difference of 6%) was
18%; for dia. 50 and 53.5 [m] (a relative difference of 7%),
it was about 13%.
The Particle-Size Distribution. The equations giving the
peak moments (Eqs. 4 and 5) are correct only when it is
assumed that all the particles in the bed have the same diameter. However, the particles of all particulate materials have a
certain diameter distribution. The integral distribution of the
volume fraction f, as a function of the particle diameter is
presented in Figure 2 for the 15 m particles. This result is
typical of those obtained for all the materials investigated in
this work. The PSD were determined with a Mastersizer (Mavern Instrument) by the column supplier.
To investigate the inuence of the PSD on the second
moment, numerical solutions of the GR model were calculated.
Figure 3 compares the peak proles obtained for a packing
material made of uniform particles having a diameter equal to
the mean moment volume diameter (d 43 46.49 [m]), and
for the correct PSD. The peak obtained with the actual PSD is
slightly shorter and wider than the one corresponding to identical beads. For particles having a smaller size, these discrepancies are smaller. The relative differences between the values
calculated for the rst moment of the peaks by the two methods
just described were less than 1 106. The relative difference
between the second moments

2,PSD 2


is reported in Table 2. In this equation, 2,PSD

and 2 denote
the second moments calculated with the actual PSD, and with
the mean-particle diameter, respectively.
As seen in Table 2, the relative differences between the
correct value of the second moment and the one calculated with
Vol. 51, No. 12


Table 2. Relative Differences Between the Moments

Calculated from the PSD and from the Mean
Particle Diameter
d 43 [m]
u o [ml/min]

Figure 2. Integral volume fraction f, or ratio of the volume of the particles with a diameter less than
d to the total volume of the particles vs. the
particle diameter.
The volume mean diameter of this material is d 43 12.77

the average particle size is negligible for particle diameters less

than about 10 [m]. For a nominal particle diameter of 50
[m], the relative difference can be as large as 7%. It decreases
slowly with increasing mobile phase ow velocity.

Determination and calculation of the axial dispersion

According to Eq. 17, a series of elution chromatograms were
run for each probe compound, on each column, at different
mobile phase ow rates. The dependence of H on the mobile
phase velocity results from the effect of axial dispersion. A plot
of ( 2 L)/(2 21 u 0 ) vs. 1/u 20 should be a straight line, having for

Figure 3. Comparison between the peak proles calculated by GR for the average particle dia. d43
46.49 [m] (higher peak), and for the true PSD
(smaller peak).

December 2005






slope the coefcient of axial dispersion D L , and for intercept,

the value of H 0 . The plots of H vs. 1/u 20 for the ve columns
studied are shown in Figure 4. The experimental data points
span nearly perfect straight lines for all ve-particle diameters
investigated. All the R 2 values were larger than 0.99. From this
result, it follows that the value of H o is practicably independent
of the uid velocity in the range studied, which shows that, in
this range, the coefcient k ext is practically constant. The
values of the axial dispersion coefcient calculated from the
slopes of the straight lines in Figure 4 are reported in Table 3,
and compared with the theoretical value derived from Gunn
correlation, Eq. 25. As expected, the experimental values of the
dispersion coefcient increase with increasing particle size.
However, there is no systematic relationship between the axial
dispersion coefcient, and the molecular diffusivity of ethylbenzene, propylbenzene and butylbenzene, which decrease in
this order.
This result suggests that the inuence of molecular diffusivity on the apparent dispersion coefcient is rather minor, meaning that the contribution of eddy diffusion is much larger than
that of molecular axial diffusion.
The two values of D L derived from Eq. 25, and reported in
Table 3 were calculated for the smallest, and the largest mobile
phase velocities for which moment measurements were carried
out, for each particle diameter. Except for the largest particle
diameter for which this ratio is about two, the second value (at
the largest ow rate) is between three and four times larger than
the rst one. The agreement between the calculated and the
experimental values is relatively good for small particle diameters (3 and 5 [m]), the experimental values being intermediate between the two values derived from the calculations. For
larger particle diameters (10 to 50 m), the calculated dispersion coefcient is about two to three times larger than the
experimental one. The worst agreement was obtained for the
medium particle size. Finally, values of the dispersion coefcient calculated from Eq. 28 are given in the last column of
Table 3. These values are generally much larger than the
experimental ones, especially in the size range most frequently
used in chromatography. From these results it seems that Gunn
correlation approximates better our experimental data.
Relatively large discrepancies between the experimental and
the theoretical values can result from our neglecting so far the
inuence of the PSD, and using the average particle size in the
calculations of the moments and of the axial dispersion coefcients. We need to evaluate the possible effect due to PSD on
the determination of an estimate of the apparent dispersion
coefcient. The largest difference between the second moments derived from the mean particle diameter and from the
PSD was found earlier to take place for the largest particle
diameter (see Table 2). Because the nite width of the particlesize distribution causes an increase in the second moment, it
will result in the experimental value of H being larger than
theoretical value that would be calculated for a bed of identical
Vol. 51, No. 12

AIChE Journal

Figure 4. Typical plot of H vs. 1/u02 according to Eq. 17.

H is derived from the second moment of the propylbenzene peaks at different mobile phase ow velocities. (a) Particles between 3 and 15
m, and (b) Particles of 50 m.

particles. Moreover, the value of the second moment depends

on the mobile phase velocity and decreases with increasing
uid velocity. All this means that the fact that the PSD has a
nite width affects the slope of the straight lines in Figure 4.
To account for the inuence of the PSD and evaluate the
maximum possible error made on the dispersion coefcient, the
H values calculated for the largest particle size were multiplied by (1 ), with given in Table 2. The sign minus
accounts for the fact that the second moment of a hypothetical
bed of identical particles is smaller than that of a bed of
particles with a PSD having a nite width. The difference
between the two values of the dispersion coefcients calculated
after and before this correction was less than 5%. So, the error
made from neglecting the nite width of the PSD cannot
explain the differences observed between the experimental
values of D L , and those derived from classical correlations,
Eqs. 25, 28. Neither can the uncertainty of the actual particle
size can explain these discrepancies. The difference between

the values of D L calculated from Eq. 25 for particle dia. of 10

and 9 [m] is less than about 15%.

Determination of the external mass-transfer coefcient

and the effective diffusivity
According to Eq. 17, the value of H o can be derived from the
intercept of the straight lines with the ordinate axis. This value
of H o is affected with the systematic error caused by neglecting
the inuence of the PSD. However, we did not try to correct for
this error because of the uncertainty of the actual mean-particle
diameter, as just discussed, and because the error is relatively
small, anyway.
The value of H o is directly related to the rate constant of the
external mass transfer kinetics and to the effective diffusivity.
Equation 16 shows that the effective diffusivity can be derived
from a plot of H o vs. R 2 , as the slope of the straight line
obtained while the intercept of this plot with the ordinate axis

Table 3. Calculated and Measured Values of DL

D L [cm2/s]
d [m]


Eq. 25


4.28 10


6.30 105


8.13 105


1.34 104


1.17 103



1.45 10
5.86 105
2.16 105
8.61 105
1.32 104
5.37 104
2.48 104
9.08 104
1.52 103
3.53 103

Eq. 25

4.23 10

5.96 105
8.88 105
2.02 104
1.27 103



1.51 10
6.06 105
2.28 105
9.23 105
1.42 104
5.67 104
2.56 104
9.50 104
1.58 103
3.61 103

Eq. 25b

3.53 10

7.15 105
9.29 105
1.46 104
1.28 103

Eq. (28)b

1.55 10
6.38 105
2.37 105
9.72 105
1.49 104
5.89 104
2.68 104
9.81 104
1.62 103
3.66 103

1.1 104
2.6 104
2.1 104
4.7 104
4.3 104
9.5 104
6.0 104
1.3 103
1.9 103
3.3 103

Values derived from the slopes of the lines in Figure 4, using Eq. 17.
Values calculated from Eq. 25 or 28.

AIChE Journal

December 2005

Vol. 51, No. 12


gives the external mass-transfer coefcient. However, such a

simple method cannot be used in this study because the external mass-transfer coefcient is a function of the particle size.
This problem is usually circumvented by calculating the external mass transfer coefcient from one of the generally accepted
correlations and then using Eq. 16 to calculate the effective
One of the most widely accepted correlations for the rate
constant of the external mass transfer kinetics is the WilsonGeankoplis equation.7 It indicates that the external mass-transfer coefcient should be inversely proportional to (d) 2/3 . The
correlation is valid for 0.0015 Re 55. In our experiments,
the Reynolds number was between: 0.001 and 0.04, so it was
around the low end of the range of validity of the WilsonGeankoplis equation. In this work, we will not assume any a
priori correlation for the external mass-transfer coefcient, but
will estimate directly k ext from the dependency of H o on the
particle diameter, using the following equation

3k ext


1 pRD sK R

m p


Table 5. Experimental Values of the External Mass

Transfer-Coefcient and the Surface Diffusivity


Tortuosity factor from Eq. 34

0.66 (E)
0.66 (P)
0.66 (B)
Tortuosity factor from Eq. 35
0.66 (E)
0.66 (P)
0.66 (B)
1 (E)
1 (P)
1 (B)
1 (E)
1 (P)
1 (B)

D m pR
1 pRD sK R


the superscript R in the last two equations was added to

emphasize that parameters A, p , and K depend all on the
particle size. This means that the parameters listed earlier, have
different values for different columns lled with particles of the
same adsorbent but having different diameters. The tortuosity
parameter was calculated from one of the following two correlations2,12

2 p 2


p 1.51 p


As indicated earlier, the Wilson-Geankoplis correlation7 states

that k ext is proportional to 1/(R) 2/3 . On the other hand, for a
Table 4. Comparison of Experimental and Calculated Rate
Constants of External Mass Transfer for Different Average
Particle Diameter
k ext a [cm/s]

d [m]






Values for n-propylbenzene.

Values calculated using Eq. 23. Tortuosity factor from Eq. 34. 2/3.


December 2005

D s [cm2/s]


6.34 107
4.926 107
5.04 107


3.45 107
3.44 107
4.23 107
5.57 107
1.44 106
4.59 107
2.45 107
2.28 107
2.19 107

1 104

The symbols E, P, and B denote ethyl, propyl and butylbenzene, respectively.

The values of the external mass-transfer coefcient were calculated for a
particle diameter equal to 5.12 m.

stagnant uid phase, the external mass-transfer coefcient is

inversely proportional to the particle dia.,12 and k ext D m /R p .
Taking all that into account, it was assumed that

This equation is derived from Eq. 16 assuming that the

contributions of the pore diffusivity and the surface diffusivity
D s , to the effective diffusivity are parallel. Thus,
D eff

k ext

k ext

R p


combination of Eqs. 32, 33, 36 and either Eq. 34 or 35 gives an

equation with three unknown parameters, , , and D s . However, there are only ve experimental data points, for the ve
different particle diameters used in this work. This is too small
a data set to perform a robust estimation of all three parameters.
To avoid this problem, a value of was assumed arbitrarily in
Eq. 36, and the other two parameters were then estimated,
using the LevenbergMarquardt method modied by
Fletcher.22 The external mass transfer coefcient was calculated on the basis of the assumed value of and of the
estimated value of . Typical examples of the results obtained
for this estimation are reported in Tables 4 and 5.
Table 4 lists the values of the external mass-transfer coefcient k ext , calculated for propylbenzene on the ve columns
packed with particles of ve different diameters, assuming that
the parameter in Eq. 36 is equal 2/37 and that the tortuosity
factor is given by Eq. 34. The values calculated for this coefcient, listed in the column experiment, are compared with
those calculated from Eq. 23, listed in the column theory.
The smaller and the higher values in this column were calculated from the experimental data corresponding to the minimum and maximum mobile phase velocities. The calculated
values are about twice smaller than the experimental ones. For
the other compounds studied, the results (not shown) are very
The estimated values of the external mass-transfer coefcient, and of the surface diffusivity obtained for different
values of the parameter in Eq. 36, and for all the compounds
studied are given in Table 5. In the last column of this table, the
value of the Fisher test, F, is also given. The higher F, the
better the estimate. The value 2/3 is that of the WilsonGeankoplis correlation. The value 1 corresponds to a
stagnant uid. A value 1 corresponds to the assumption
Vol. 51, No. 12

AIChE Journal

uid, so the layer can be regarded as a stagnant lm of liquid.

From this theory, the following relationship can be derived
k ext

Figure 5. Comparison of exemplary experimental and

theoretical value of Ho calculated for propylbenzene and model parameter estimated for



where is the thickness of the thin lm. The main disadvantage

of this theory is that it is practically impossible to calculate .
However, there is a general agreement that this layer thickness
decreases with increasing Reynolds number. In the case of our
experiments, the Reynolds number increases 13-fold with increasing particle radius. This would explain the important
increase of the mass-transfer kinetics, and the increase of k ext
with increasing particle size that we have observed.
We understand that the assumption that we have just formulated needs to be veried carefully, under more sophisticated
experimental conditions. Such an investigation is now in
progress. Nevertheless, in the following, the interpretation of
the experimental data is performed assuming the validity of the
parameters estimated with the assumption that 1, parameters that are given in Table 5.

Contributions of axial dispersion, external and internal

mass-transfer resistances to HETP
that the rate of the external mass transfer is proportional to the
particle diameter. The smallest value of F was obtained for a
stagnant uid phase, F is slightly higher for 2/3 and,
surprisingly, the highest of F corresponds to the assumption
that 1. A slightly better agreement with the experimental
data was achieved when the tortuosity was calculated from Eq.
35. A comparison of the experimental and calculated values of
H o for propylbenzene and for 1 is shown in Figure 5.
Before analyzing further the data in Table 5, the following
remark must be made. The differences between the values
derived for k ext for different solutes on a given column (hence
particle size) may arise only from differences between molecular diffusivities, so the quotient of the external mass-transfer
coefcients calculated for the different compounds studied
should be proportional to the ratio of their molecular diffusivities. These ratios, calculated from the Wilke-Chang equation,
Eq. 24, and the ratios of the experimental external masstransfer coefcients, are reported in Table 6. It is obvious that
the proportionality was observed only for 1 and, moreover, that there is equality between the ratios of the external
mass-transfer rate coefcients and of the molecular diffusivities. This suggests that k ext depends linearly on the molecular
Another interesting remark follows from the analysis of the
surface diffusivity data in Table 5. Surface diffusivity should
decrease with increasing molecular weight of the compound
studied in this homologous series. It is obvious that only the
data obtained under the assumption that 1 fulll this
requirement. Summarizing: the best and most satisfactory
agreement between the current theory of mass transfer in
porous media and our experimental results is obtained under
the assumption that k ext R.
This result can be explained by considering the simple
Lewis-Whitman stagnant lm model.23 When a uid ows over
a surface, a boundary layer always forms against this surface.
The ow velocity in this layer is much lower than in the bulk
AIChE Journal

December 2005

The contributions of H L , H k , and H D to the total HETP (see

Eq. 14) are plotted in Figure 6, as a function of the particle size,
for a mobile phase velocity equal to 1 [ml/min]. The calculations were made for propylbenzene. The largest contribution to
the HETP is that of H D that characterizes the internal masstransfer resistances. The smallest one is the one due to the
external mass-transfer resistances, characterized by H k . The
axial dispersion, characterized by H L , plays an intermediate
role. All these contributions depend on the mobile phase velocity. The dimensionless contributions of these three contributions to the HETP, for mobile velocities of 0.8 and 1.2
[ml/min], are shown in Figures 7 and 8, respectively.
For the smallest particle diameter, at the smallest mobile
phase velocity, the contribution of axial dispersion to the HETP
prevails over the other two contributions. The relative importance of the internal mass-transfer resistance increases rapidly
with increasing particle size and its contribution dominates for
particles larger than about 10 [m]. However, axial dispersion
can never be neglected. The contribution of the external masstransfer resistance is marginal for particle sizes larger than
about 15 [m]. Similar conclusions can be derived from the
plots shown in both Figures 7 and 8. It should be noticed,
Table 6. Ratios of the Molecular Diffusivities and the
Experimental External Mass-Transfer Coefcients
k i /k i c














D mi /D mj


Symbols E, P and B denote ethyl, propyl and butylbenzene.

Molecular diffusivities from Eq. 24.
Tortuosity parameter from Eq. 35.


Vol. 51, No. 12


Figure 6. Contributions HL, Hk, and HD to the total HETP

as a function of the particle radius.

Figure 8. Same as in Figure 7, except mobile-phase velocity, u 1.2 [ml/min].

Plot made for propylbenzene.

however, that the importance of the contribution of the internal

mass resistance to band broadening increases with increasing
mobile phase velocity.

Contributions of pore diffusion and surface diffusion to

effective intraparticle diffusion
Equation 33 describes the relationship between the effective
diffusivity, D eff , the pore diffusion, D p , the surface diffusion
D s , and the effective surface diffusion D e,s , coefcients. The
ratio of these diffusion coefcients is reported in Table 7. The
ratio of the surface to the molecular diffusivity is also given.
The ratio D s /D m is similar to that reported previously.6 However, the ratio D e,s /D p is much smaller.6 This discrepancy
comes from the method used to interpret the experimental data.
In the previous work, the external mass transfer was calculated
from Wilson-Geankoplis correlation Eq. 23 in which the masstransfer coefcient is inversely proportional to the particle

diameter to the power (2/3). A value of D e,s /D p similar to the

one in Ref. 6, equal to ca Eq. 14, could be obtained from the
data in this work, if we assume 0.66, and estimate surface
diffusivity with the tortuosity factor calculated from Eq. 34.
To avoid possible ambiguities in the interpretation of experimental data, more sophisticated experiments that would allow
an independent evaluation of the kinetic rate constants must be

Although the apparent dispersion coefcient and the external
mass-transfer coefcient may depend on the mobile phase
velocity, our experimental results show that this inuence is
smaller than the experimental errors in the range of velocity
investigated. The particle diameter inuences the external mass
transfer coefcient. The experimental results show that the
dependence of the rate constant of the external mass transfer on
the particle diameter is not consistent with the Wilson-Geankoplis correlation. This correlation, often used to account for
HPLC results, is generally accepted as valid in chemical engineering but it was developed for particle sizes nearly 1,000
times larger than commonly used in contemporary chromatography, and for Reynolds numbers that were, mostly, much
larger than those used in liquid chromatography. The range of
validity of the correlation is 0.0015 Re 55 while columns
are run usually at 0.001 Re 0.05, which places them at the
very low end of the recommended range of validity of the
Our experimental results suggest, by contrast, that the exterTable 7. Ratios of Effective Surface to Pore Diffusivities and
of Surface to Molecular Diffusivities

Figure 7. Dimensionless contributions, HL, Hk and HD, to

the HETP for propylbenzene.
Indexes 1, 2, 3, 4, 5 on the abscissa axis refers to average
particle sizes of 3.74, 5.21, 10.6, 14.92 and 47.1 [m] respectively. Mobile phase velocity, u 0.8 [ml/min].


December 2005


D e,s /D p

D s /D m



4.04 102
3.97 102
4.09 102

Symbols E, P, and B denote ethyl, propyl and butylbenzene.

Tortuosity parameter was calculated from Eq. 35.

Vol. 51, No. 12

AIChE Journal

nal mass-transfer coefcient is proportional to the particle

diameter. This result is surprising. The external mass-transfer
coefcient does increase linearly with increasing particle size
in the investigating range of Reynolds numbers. These results
should be considered as preliminary and must be conrmed by
carrying out more sophisticated experiments.
Finally, it was also shown that the interpretation of the
relative role of pore and surface diffusivity in the internal mass
transfer is ambiguous, at least if our results are conrmed. The
conclusion on this point will depend on the assumption made
on the best model of tortuosity and on the exact relationship
between k ext , and the particle size.


absolute moment
second central moment
uid density

a denote adsorbent
i i-th fraction of adsorbent particle
s solid phase

R means that parameter depends on used particle radius


Literature Cited

This work was supported in part by grant CHE-02-44693 of the National

Science Foundation, and by the cooperative agreement between the University of Tennessee and the Oak Ridge National Laboratory. We thank
Tivadar Farkas (Phenomenex, Torrance, CA) for the generous gift of the
columns used in this work and for fruitful and creative discussions.

1. Burlington DB, Haggard D, Kaminski E, Sakers F, Grazal J, Whitaker

J, King E. A Risk-Based Approach to cGMPs. FDA Report; 2003.
2. Guiochon G, Golshan-Shirazi S, Katti AM. Fundamentals of Preparative and Nonlinear Chromatography. Boston: Academic Press; 1994.
3. Ruthven DM. Principles of Adsorption & Adsorption Processes. New
York; NY: Wiley; (1984).
4. Felinger A, Cavazzini A, Guiochon G. Numerical Determination of the
Competitive Isotherm of Enantiomers. J Chromatogr A 2003: 986: 207.
5. Bird RB, Stewart WE, Lightfoot EN. Transport Phenomena, New York;
NY, Wiley, (2002).
6. Miyabe K, Guiochon G. Measurement of the Parameters of the Mass
Transfer Kinetics in High Performance Liquid Chromatography. J
Separat Sci 2003: 26: 155.
7. Wilson EJ, Geankoplis CJ. Liquid Mass Transfer at Very Low Reynolds
Numbers in Packed Beds. Ind Eng Chem Fundam 1966: 5: 9.
8. Kaczmarski K, Bellot JC. Effect of particle-size distribution and particle
porosity changes on mass-transfer kinetics. Acta Chromatographica
2003: 13: 22.
9. Kucera E. Contribution to the theory of chromatography. Linear nonequilibrium theory of chromatography. J Chromatogr 1965: 19: 237.
10. Kubin M. Collection Czechoslov. Chem Commun 1965: 30: 1104.
11. Schneider P, Smith MJ. Adsorption rate constants from chromatography. AIChE J 1968: 14: 762.
12. Suzuki M. Adsorption Engineering. The Netherlands: Amsterdam;
Elsevier, 1990.
13. Miyabe K, Guiochon G. Fundamental interpretation of the peak proles in linear reversed-phase liquid chromatography. Adv Chromatogr
2000: 40: 1.
14. Wilke CR, Chang P. Correlation of diffusion coefcients in dilute
solutions. AIChE J 1955: 1: 264.
15. Gunn D. Axial and radial dispersion in xed beds. Chem Eng Sci 1987:
42: 363.
16. Chung SF, Wen CY. Longitudinal dispersion of liquids owing
through xed and uidized beds. AIChE J 1968: 14: 857.
17. Neue UD. HPLC Columns. Theory, Technology, and Practice. Wiley
VCH; 1997.
18. Kele M, Guiochon G. Repeatability and reproducibility of retention
data and band proles on reversed-phase liquid chromatography columnsIV. results obtained with luna C18 columns. J Chromatogr A
2000: 869: 181.
19. Felinger A, Guiochon G. Validation of a chromatography data analysis
software. J Chromatogr A 2001: 913: 221.
20. Halasz I, Martin K. Pore Sizes of Solids. Angew Chem. (Int. Ed. Engl.)
1978: 17: 901.
21. Guan H, Guiochon G. Study of the physico-chemical properties of
some packing materials. I. measurement of the external porosity of
packed columns by inverse size exclusion chromatography. J Chromatogr A 1996: 731: 27.
22. Fletcher R. A modied Marquardt subroutine for nonlinear least
squares. AERE-R6799 Harwell.
23. Skelland AHP. Diffusional Mass Transfer. New York: Wiley; 1974.

A parameter dened by Eq. 15
C concentration in the mobile phase
C p concentration in the stagnant uid phase contained inside pores
d equivalent particle diameter
d 43 volume moment mean diameter
D m diffusion coefcient
D eff effective (or inside-pore) diffusion coefcient
D p pore diffusivity
D L dispersion coefcient
D s surface diffusivity
D s,e effective surface diffusivity
H high equivalent to theoretical plates
H L , H k , H D parameters dened by Eq. 14
H o , H parameters dened by Eqs. 16 and 17
k ext external mass-transfer coefcient
K equilibrium constant
L column length
N number of theoretical plates
M s molecular mass
Pe Peclet number, uL/D L e
q concentration of the solute in the stationary phase
r radial distance from the center of the particle
Re Reynolds number, ud/
R particle radius
Sc Schmidt number, / D m
Sh Sherwood number, k ext d/D m
t time
T absolute temperature
u o average supercial velocity of the mobile phase
V volume
V A molar volume of solute
x longitudinal distance along the column

Greek symbols
A association factor
exponent of the dependence of the rate constant of
external mass transfer on the particle diameter (Eq. 36)
thickness of the thin lm
L , f , d , o parameters dened by Eqs. 710
tortuosity parameter
e external porosity
p internal porosity
t total porosity

AIChE Journal

December 2005

Manuscript received Dec. 20, 2004, and revision received Apr. 7, 2005.

Vol. 51, No. 12



Offset-free Receding Horizon Control of

Constrained Linear Systems
Gabriele Pannocchia
Dept. of Chemical Engineering, Industrial Chemistry and Science of Materials, University of Pisa, Via Diotisalvi, 256126
Pisa, Italy

Eric C. Kerrigan
Dept. of Engineering, University of Cambridge, Cambridge CB2 1PZ, U.K.
DOI 10.1002/aic.10626
Published online August 19, 2005 in Wiley InterScience (www.interscience.wiley.com).

The design of a dynamic state feedback receding horizon controller is addressed, which
guarantees robust constraint satisfaction, robust stability and offset-free control of
constrained linear systems in the presence of time-varying setpoints and unmeasured
disturbances. This objective is obtained by rst designing a dynamic linear offset-free
controller and computing an appropriate domain of attraction for this controller. The
linear (unconstrained) controller is then modied by adding a perturbation term, which
is computed by a (constrained) robust receding horizon controller. The receding horizon
controller has the property that its domain of attraction contains that of the linear
controller. In order to ensure robust constraint satisfaction, in addition to offset-free
control, the transient, as well as the limiting behavior of the disturbance and setpoint need
to be taken into account in the design of the receding horizon controller. The fundamental
difference between the results and the existing literature on receding horizon control is
that the transient effect of the disturbance and set point sequences on the so-called target
calculator is explicitly incorporated in the formulation of the receding horizon controller. An example of the control of a continuous stirred-tank reactor is presented. 2005
American Institute of Chemical Engineers AIChE J, 51: 3134 3146, 2005

Keywords: model predictive control, constraints, offset-free control, tracking, robust


The control of systems in the presence of constraints is an
important task in many application elds because constraints
always arise from physical limitations and quality or safety
reasons. Moreover, in practical applications, disturbances are
usually present and often they are not measurable or predictable. For example, in the chemical industries disturbances arise
from interactions between different plant units, from changes
Correspondence concerning this article should be addressed to G. Pannocchia at

2005 American Institute of Chemical Engineers


December 2005

in the raw materials and in the operating conditions (such as

ambient temperature, humidity, and so on).
It is well-known that with an unmeasured persistent disturbance offset-free control is, in general, not possible. However,
if the disturbance has the additional property that it is integrating or periodic, then offset-free control may be an achievable
goal. In many practical applications, especially in the process
industries, disturbances are often integrating and reach, after
some transient, a constant value. Hence, one basic objective of
an effective control algorithm is that it guarantees offset-free
control whenever this is possible. Moreover, an effective control algorithm is also applicable to cases in which the set points
of the controlled variables are allowed to vary over time.
Vol. 51, No. 12

AIChE Journal

The design of control algorithms able to stabilize constrained

linear plants with input and state constraints subject to unknown, but bounded disturbances, has been the subject of
several works over the last few decades; a number of excellent
surveys are available,15 which discuss how the important goal
of guaranteeing closed-loop stability and constraint satisfaction
can be obtained. Existing control algorithms, which address the
problem of robust control of constrained systems, are usually
based on ideas from set invariance,1,6,7 reference governors,8 12
or receding horizon control.1322 It is interesting to note that,
despite the practical importance of guaranteeing offset-free
control in the presence of integrating disturbances, none of the
existing receding horizon control algorithms with robust stability and robust constraint satisfaction guarantees are able to
guarantee offset-free control.
Compared to linear (unconstrained) control, the rigorous
study of designing controllers that guarantee offset-free control
has received very little attention in the constrained control
community, until relatively recently.2330 However, though the
receding horizon control algorithms presented in 2330 guarantee offset-free control around a neighborhood of the steadystate, they do not guarantee robust constraint satisfaction for all
initial states over which the controller is dened (in other
words, they do not guarantee feasibility of the optimization
problem for all time and for all allowable disturbance and set
point sequences). Furthermore, with the exception of 8 10,14,16
(which do not guarantee offset-free control for disturbances
that decay to nonzero values), none of the existing receding
horizon control algorithms that guarantee robust stability and
robust constraint satisfaction, address the problem of tracking
arbitrary set points.
In this article, we present a novel receding horizon control
algorithm for controlling constrained linear systems subject to
unmeasured, bounded disturbances. The proposed algorithm is
guaranteed to remove steady-state offset in the controlled variables whenever the disturbances reach an (unknown) steadystate value. Moreover, the set points of the controlled variables
are allowed to vary arbitrarily with time within a prespecied
set, provided they also converge to some limit. Importantly, the
algorithm also guarantees to satisfy input and state constraints
for all allowable disturbance sequences. None of the existing
receding horizon control algorithms are able to provide similar
The difculty with guaranteeing robust constraint satisfaction, in addition to offset-free control, is that the transient, as
well as the limiting behavior of the disturbance and set point
need to be taken into account during the computation of the
control input. The fundamental difference between the algorithm proposed in this article, and those available in the literature, is that we consider both the transient and limiting effect
of all allowable future disturbance and set point sequences
when computing the receding horizon control action. This is
the key idea that allows us to guarantee offset-free control and
robust constraint satisfaction. Existing approaches either neglect the transient response or they neglect the ultimate behavior of the disturbance and set point, hence, why they are unable
to offer the same kind of guarantee.
This article is organized as follows. First, the problem definition is given, followed by the design of a linear offset-free
controller. Following this, we show how a receding horizon
controller can be used to improve on the linear controller by
AIChE Journal

December 2005

enlarging the region of attraction. The main characteristics of

the proposed receding horizon controller are illustrated on an
example of a continuous stirred-tank reactor. Finally, the main
contributions of this work are summarized and some possible
extensions are discussed. To simplify the presentation and
reading of this article, the proofs of results that can be derived
using well-known methods have been omitted. The only proofs
that have been included are for those cases where we were
unable to nd equivalent and sufciently detailed proofs,
where we have used unconventional proof techniques, or where
we feel that a particular detail of the proof is important.
Notation: If a and b are vectors, then the column vector (a,
b) : (a T b T ) T . Given an arbitrary set , N denotes the
Cartesian product . . . .

Problem Description and Preliminary Results

In this article, we consider a discrete-time linear time-invariant plant
xk 1 Axk Buk Edk


zk C zxk


in which k is the sample instant, x n is the plant state, u

m is the control input (manipulated variable), d r is an
unmeasured disturbance and z p is the controlled variable,
that is, the variable to be controlled to a given (time-varying)
set point s. Afne inequality constraints are given on the state
and input,1 that is
x ,


where is a polyhedron (that is, a closed and convex set that

can be described by a nite number of afne inequality constraints) and is a polytope (that is, a bounded polyhedron);
the interior of contains the origin. We also make the
following standard assumption:
Assumption 1 (General). A measurement of the plant state
is available at each sample instant, ( A, B) is stabilizable, ( A,
C z ) is detectable and (see for example,28)

I C A B0 n p


Remark 1. The rank condition (Eq. 3) is used to guarantee

the existence of an offset-free steady-state for any constant set
point and disturbance pair. In general, the steady-state is not
necessarily unique for a given set point and disturbance. Note
also that this condition implies that the number of controlled
variables cannot exceed the number of control inputs or the
number of states, that is, p min{m, n}. It is easy to nd
examples for which offset-free control is not possible if Eq. 3
is not satised.
We also consider the following assumptions on the set point
and disturbance sequences:
Assumption 2 (Set point). At each time instant, the current
The results in this article can easily be extended to the case with mixed constraints
on the state and input.

Vol. 51, No. 12


set point is known but future set point values are unknown. The
set point sequence s takes on values in a polytope p
containing the origin and asymptotically reaches an unknown
steady-state value, that is, s(k) for all k , and there
exists an s such that limk3 s(k) s .
Assumption 3 (Disturbance). At each time instant, current
and future disturbances are unknown. The disturbance sequence d takes on values in a polytope r containing
the origin and asymptotically reaches an unknown steady-state
value d , that is, d(k) for all k and there exists a d
such that limk3 d(k) d .
Remark 2. Note that, unlike many existing results, we do
not assume that the disturbance or set point is constant. Furthermore, unlike,25 we do not assume that the set point and/or
disturbance are generated by a known nite-dimensional exogenous system. The lack of these assumptions in this article
complicates the design of the controller.
Under the earlier assumptions we present a novel method for
designing a dynamic, state feedback receding horizon controller that, for any allowable disturbance and set point sequences
(that is, any innite disturbance and set point sequences that
satisfy Assumptions 2 and 3), accomplishes the goal of asymptotically driving the controlled variable to any given allowable
asymptotic set point, while respecting the state and input constraints, that is
lim zk s



xk ,

uk ,

k .


Before proceeding, we present here the following wellknown result:6,15,18,20,31

Proposition 1. Let the polyhedron
: t F g Hw for all w
where F qt and H qs are matrices, g q is a
vector and is a compact (that is closed and bounded) subset
of s , then
tF g min Hw

where the minimization is performed row-wise, that is if H i

denotes the i th row of H, then
min Hw :

min H1 w min Hq wT

Since our system, in closed-loop with the receding horizon

controller, is nonlinear, and we are interested in robust stability
results, we review the following denitions and results, adapted
from 32 for a generic nonlinear perturbed discrete-time system

k 1 F k wk


Denition 1. The origin is a robustly asymptotically stable

xed point of Eq. 5 if for all 0, there exist a 0 and a
0 such that for all initial conditions (0) and
perturbation sequences w satisfying w(k) for all k
, the following two conditions are satised:
1. (Robust stability) The solution of Eq. 5 satises (k)
for all k ;
2. (Robust convergence) The solution of Eq. 5 satises
limk3 (k) 0 if limk3 w(k) 0.
Denition 2. If w : limk3 w(k) is the limit point of the
perturbation sequence w, then a vector satisfying
F( ) w is a robustly asymptotically stable xed point of Eq.
5 if the origin is a robustly asymptotically stable xed point of
the system (k 1) G( (k)) (k), in which :
, : w w and G( ) : F( ) F( ).
Theorem 1. [32, Thm. 3] Let F : 3 be a Lipschitz
continuous function in a neighborhood of the origin with
F(0) 0. If the origin is an exponentially stable xed point of
the unperturbed system (k 1) F( (k)), then it is a
robustly asymptotically stable xed point of the perturbed
system (k 1) F( (k)) w(k).

Linear Controller Design

In order to guarantee robust constraint satisfaction and stability in receding horizon control, it is by now standard practice
to compute a suitable terminal constraint and terminal cost,
based on a stabilizing linear controller.25 This section shows
how one can compute such a linear controller and an appropriate terminal constraint. The difference between this article
and standard results is that the controller in this article is
dynamic, rather than static, hence, the terminal constraint is
computed in the joint plant-controller state-space, rather than
just the plant state-space.
In order to guarantee offset-free control when disturbances
are asymptotically constant and nonzero, it is standard practice
to augment the plant model with a disturbance model and use
this combined model to estimate the size of the disturbance.
However, as pointed out in,27,28 this is not necessarily as
straightforward as is commonly thought. Care has to be taken
in constructing the controller, since offset-free control is guaranteed only if the combined plant-disturbance system is detectable and the closed-loop system (feedback gain observer
target calculator plant) satisfy a couple of additional, technical assumptions (see 27,28 for details). This section will,
therefore, show in detail how one can design a dynamic controller (observer and static state feedback gain) for Eq. 1 that
guarantees offset-free control. The general structure of the
proposed controller is depicted in Figure 1, and each block is
described next.

Augmented system
We will make use of the following auxiliary system in order
to dene the controller dynamics
x k 1 Axk Buk d k xk x k

in which F : 3 and F(0) 0. Let r : {

r} if r 0.

December 2005

d k 1 d k xk x k
Vol. 51, No. 12


AIChE Journal

subject to

I C A B0 ux 0I

Figure 1. Proposed dynamic offset-free feedback controller.

Note that the system (Eq. 6) corresponds to using a dead-beat

observer for the following system
xk 1 Axk Buk d k
d k 1 d k
yk xk
in which it is clear that d n is an integrated (step)
disturbance acting on the state x n . The role of d is
essential in removing steady-state offset in the presence of an
unknown persistent disturbance.27,28 As will be seen later, the
dimensions of d and d need not be the same in order to
guarantee offset-free control.
By combining the plant dynamics (Eq. 1), and the auxiliary
system (Eq. 6), we obtain the following augmented system

k 1 k uk dk


zk k


in which : ( x, x , d ) and , , and are suitablydened.

Target calculation
When a nonzero disturbance affects a system (and/or the
current set point s is different from zero), the steady-state
target value of the state and input may need to be shifted in
order to cancel the effect of such a disturbance on the controlled variable.33,34 To this aim, at each sample instant we use
the estimate of the future disturbance and compute the steadystate target ( x , u ) such that one can drive the controlled
variable to the current set point, by solving the following
least-squares problem 34 in which R mm is a positive
denite matrix
x * , s, u * , s : argmin
x ,u

AIChE Journal

1 T
u R u


December 2005

0 0 I s


where it can be seen that the current estimate of the limiting

value of the disturbance is given by [I I I] (k)
x(k) x (k) d (k) d (k 1).
For a given augmented state , and a given set point s, one
can think of ( x *( , s), u *( , s)) as the new origin around
which the system should be regulated. Solving for ( x *( , s),
u *( , s)) is trivial:
Lemma 1 (Target calculation). If Assumption 1 holds,
then the target calculation (Eq. 8) has a unique solution that is
linear with respect to the augmented state and the set point
s, and has the form

ux **,, ss



x, x,
u, u, u,s s


for suitably-dened constant matrices x, nn , u,

mn , x,s np and u,s mp .
Proof. The statement follows immediately from the Lagrangian/KKT conditions for Eq. 8,35 Sect. 16.1. It is possible
to verify that the matrix, which is to be inverted when obtaining
the expression for the stationary point of the Lagrangian, is
nonsingular because of the rank condition in Eq. 3, and the fact
that R is positive denite. Hence, the target calculation (Eq. 8)

has a unique minimizer.36

Remark 3. Recall that Eq. 3 on its own is not sufcient to
guarantee that the minimizer of Eq. 8 is unique. However, if R
is positive denite, then we do not need to include a penalty on
x in the cost (Eq. 8a) in order to guarantee that the minimizer
is unique.
Remark 4. Note that, unlike,27,28 the constraints on the state
and input are not included in Eq. 8. In order for the receding
horizon controller to take account of the transient of the disturbance and set point in its predictions, it needs to calculate
how ( x *( , s), u *( , s)) will vary over the control horizon. As
will be shown later, this is easy if the mapping ( , s) ( x *( ,
s), u *( , s)) is linear, as earlier. If inequality constraints had
been included in Eq. 8, then the mapping ( , s) ( x *( , s),
u *( , s)) would have been nonlinear, hence, the computation
of ( x *( , s), u *( , s)) at each time instant in the prediction
horizon, and, hence, the computation of the receding horizon
control input, would be considerably more complicated.

Unconstrained offset-free controller design

We now consider what would happen if one were to choose
a gain matrix K such that A BK is strictly stable and, as is
common practice, let the control input in the augmented system
(Eq. 7) be given by
u u * , s K x x * , s


Clearly, as can be seen in Figure 1, there is a feedback path

around the target calculator and therefore it is important to
verify that the closed-loop system (observer target calculaVol. 51, No. 12


tor state feedback gain plant) is stable. We, therefore,

present the following intermediate result:
Lemma 2 (Stability). Suppose that Assumption 1 holds and
K mn is such that A BK is strictly stable. If mn
is any constant matrix and
: K




is strictly stable.
Proof. The statement follows from straightforward block
matrix manipulations,36 where one can show that 2n of the
eigenvalues of are at zero and the rest are the eigenvalues
of A BK.

By dening
: u, Kx, ,

: u,s Kx,s


and substituting Eq. 9 into Eq. 10 it easily follows that

u s


where is dened as in Eq. 11.

After substituting Eq. 14 into Eq. 7, one can write an
expression for the augmented system (Eq. 7) under the linear
control u s as

k 1 k dk sk,


zk k,


where is suitably-dened.
It immediately follows from Lemma 2 that the closed-system
(Eq. 15) is stable if the disturbance and set point are zero (or
asymptotically constant). As a consequence, we introduce the
following standing assumption:
Assumption 4 (Stabilizing gain). The matrix K mn is
chosen such that A BK is strictly stable, is given by Eq.
11, and given by Eq. 13 and : .
The following result states that if the control is given by u
s, then the value of the controlled variable for Eq. 15
is guaranteed to converge to the asymptotic set point s , given
any allowable innite set point and disturbance sequence:
Lemma 3 (Offset-free control). If Assumptions 1 4 hold,
then the solution of the closed-loop system (Eq. 15) satises
limk3 z(k) s for all (0) 3n .
Proof. See the Appendix.

Remark 5. In this article, we have assumed that there is

no mismatch between the plant model ( A, B, E), and the
actual plant dynamics ( Ap, Bp, Ep). However, it is important
to point out that, as in,24,28 it is possible to verify that the
offset-free property in Lemma 3 holds even if there is a
mismatch between the plant model ( A, B, E), and the actual
plant dynamics ( Ap, Bp, Ep), provided the augmented closedloop system is stable. This fact is not surprising. Recall that

December 2005

the internal model principle37 (roughly) states that a controller that guarantees offset-free control even when the
system parameters are perturbed, must incorporate a model
of the dynamic structure of the disturbance, and the reference signals; in our case, the observer contains a model of
the disturbance and the target calculator contains a model of
the disturbance and set point. This point will be further
illustrated in the example section.

Maximal constraint-admissible robust positively

invariant set
As mentioned earlier, since the constraints on the state and
input (Eq. 2) are not included in the target calculation (Eq. 8),
there is no guarantee that the steady-state target will satisfy constraints. Furthermore, there is also no guarantee that the linear
controller dened earlier will guarantee the satisfaction of the
constraints during the transient. We, therefore, proceed along
similar lines as in conventional receding horizon control 25 and
compute a constraint-admissible, robust invariant set that could be
used as a terminal constraint in our receding horizon controller (to
be dened in the next section). However, since our linear controller is dynamic and not static, we depart from convention and
compute an invariant set in the space of the augmented (plant
controller) state : (x, x, d) of the closed-loop system (Eq. 15),
rather than in the plant state-space (as is usually done 25).
Let the constraint-admissible set be dened as all augmented states for which the constraints on the plant state and
plant input are satised, for any choice of set point s , if
the control is given by u s
: 3n x and s for all s


The maximal constraint-admissible robust positively invariant set for the closed-loop system (Eq. 15) is dened as all
initial states in for which the evolution of the system remains
in for all allowable innite set point and disturbance sequences, that is
: 0 k 1 k dk
sk , sk , dk , k


Assumption 5 (Maximal invariant set). The set is

nonempty, contains the origin in its interior and is nitely
determined (that is, can be described by a nite number of
afne inequality constraints).
Note that since and are polyhedra given by afne
inequalities, is easily computed by applying Proposition 1 to
the earlier denitions. Since Eq. 15 is linear and time-invariant
and is given by a nite number of afne inequality constraints, (or an inner approximation to it) is easily computed
by solving a nite number of LPs.31
We are now in a position to collect all of the results in this
section into a single statement. The following result states that,
provided the state of the augmented system (controller plant)
is in at time k 0, then the evolution of the augmented
system under the linear control u s is such that
offset-free control is guaranteed, the state and input constraints
are satised for all allowable set point and disturbance sequences, and robust stability is guaranteed:
Vol. 51, No. 12

AIChE Journal

Theorem 2 (Linear controller). Suppose that Assumptions

15 hold. The solution of the closed-loop system (Eq. 15)
satises Eq. 4 if and only if the initial augmented (controller
plant) state (0) . Furthermore, : (I ) 1 (d
s ) is the robustly asymptotically stable xed point of Eq.
Proof. See the Appendix.

Before proceeding, we also dene X 0 , the set of plant states

for which there exists a controller state such that the augmented
state is in , as
X 0 : x n x, d 2n such that


Receding Horizon Controller Design

The set X 0 , dened in Eq. 18, is the set of initial plant states
for which one can initialize the controller state such that the
controlled variable will ultimately be driven by the linear
controller to the asymptotic set point s . Clearly one would like
to enlarge this set, if possible. It is well-known that receding
horizon control allows one to achieve this aim.25 This section,
therefore, presents an approach for computing a (dynamic)
receding horizon controller, which enlarges the set of initial
plant states for which the controlled variable can ultimately be
driven to the asymptotic set point.
As mentioned earlier, in order to guarantee robust constraint
satisfaction and stability in receding horizon control, it is by
now standard practice to compute a suitable terminal constraint
and terminal cost based on a stabilizing linear controller.25
However, all existing results on receding horizon control do
not include the effect of all allowable disturbance and set point
sequences on the target calculator in their predictions. This is
an important difference that, coupled with the fact that the
controller in this paper is dynamic and not static (as in 25),
makes it necessary to show in detail how the results from the
previous section can be used to dene an appropriate receding
horizon controller.

Denition of the receding horizon controller

For the sake of simplicity of exposition and implementation,
we follow the approach of 14,16 18,20 by prestabilizing the
plant and letting the linear control in Eq. 14 be modied with
a perturbation term as follows
u s


where m is the perturbation term. The solution to the

nite horizon optimal control problem (dened later) is then a
nite sequence of input perturbations that guarantees robust

N , s : v mN

AIChE Journal

constraint satisfaction over the horizon and optimizes some

cost function.
Under the control in Eq. 19, the augmented state dynamics in
Eq. 7 become

k 1 k k dk sk

Remark 6. At this point, it is useful to recall that , ,

and in Eqs. 19 20 implicitly contain the solution of the
target calculation (Eqs. 8 and 9). Since these equations are
linear, it is easy to compute the set of admissible input perturbations, as detailed later.
Remark 7. Obviously, there exist many alternative robust receding horizon control formulations that can also be
generalized to incorporate the results in this article. For
example, one could adopt a less conservative framework,
such as the ones proposed in.15,19,22,38 However, this will
come at a greater cost in off-line and/or on-line implementation. We believe that the prestabilizing framework adopted
here is a practical approach that can be used to illustrate the
main ideas of this article without introducing unnecessary
detail. The results in this article are easily extended using
the ideas in.15,19,22,38
Before proceeding, we need to dene some notation. Let the
horizon length N be a positive integer, and the block vectors v
mN , s p(N1) , d rN be dened as v : ( 0 ,
1 , . . . , N1 ), s : (s 1 , s 2 , . . . , s N1 ), and d : (d 0 ,
d 1 , . . . , d N1 ). Note that s and all related terms are present
only if N 1. Let i denote the predicted solution to Eq. 20
i time steps into the future (time k i), given the augmented
state : (k) at the current time k, a nite sequence of
control perturbations v, the current set point s 0 : s : s(k),
a nite sequence of future set points s and a nite sequence of
future disturbances d. The corresponding predicted plant state
x i : [I n 0] i and input u i : i s i i are similarly
Given the above, we can now dene the set of admissible
input perturbations N(, s) as the set of input perturbations
of length N, such that for all allowable future set point
sequences of length N 1 and disturbance sequences of
length N, the input constraints are satised over the
horizon i 0, . . . , N 1, the state constraints are
satised over the horizon i 0, . . . , N 1 and the augmented state at the end of the horizon is in (hence, the
predicted plant state at the end of the horizon is also in ),
that is

0 , s0 s,
i1 i i di si , i 0, . . . , N 1,
xi In 0 i , i 0, . . . , N 1, N ,
ui i si i , i 0, . . . , N 1
for all s N1 and all d N

December 2005


Vol. 51, No. 12



Remark 8. Note that the predictions in Eq. 21 take into

account the solution of the target calculation in Eqs. 8 and 9 on
the control and state trajectories over the horizon i 0, . . . ,
N. This is the fundamental difference between this article and
existing results in receding horizon control and, therefore,
allows one to provide a guarantee of robust constraint satisfaction during transients.
In order to compute the receding horizon controller, we need
to dene a cost and set up an appropriate nite horizon optimal
control problem (FHOCP). We choose to dene the FHOCP to
be solved for the current augmented state and set point s, as
V *N , s : min

v N , s

VN , s, v


where the cost to be minimized is

x x*, s Qx x*, s u


V N , s, v :


u *, sT Rui u *, s xN x *, sT PxN x *, s
with the matrices Q nn , R mm and P nn
positive denite. The vectors x i n and u i m are dened
x 0 x I n 0
x i1 Ax i Bu i x x d ,

uk k sk *0 k, sk

Properties of the receding horizon controller

It is well-known that if there are disturbances present and the
receding horizon control action is computed by optimizing over
open-loop input sequences, rather than feedback policies, then
the optimization problem may become infeasible for large
control horizons.3,22 Since we are not optimizing over feedback
policies, but over a sequence of perturbations to a stabilizing
control law (which is equivalent to optimizing over open-loop
input sequences if the system is open-loop stable and the
chosen control law is zero), it is important to be able to
characterize the properties of the set of plant states for which
one can guarantee that the FHOCP (Eq. 22) has a solution.
Fortunately, optimizing over perturbations to a stabilizing
control law does not suffer the same drawbacks as optimizing
over open-loop input sequences.14,16 18,20 However, since we
are including the solution of the target calculator in our control
input and predictions, it is still important to verify that feasibility of the optimization problem is not lost in the formulation
presented in this article. We will, therefore, proceed to show
that if the FHOCP (Eq. 22) is feasible at time k 0, then the
FHOCP (Eq. 22) is feasible at all future time instants and
offset-free control is guaranteed in the limit.
The set of plant states X Nv for which one can initialize the
controller state ( x , d ) such that the set of admissible input
perturbations N ( , s) is nonempty for all s (and, hence,
the FHOCP (Eq. 22) has a solution), is given by

i 0, . . . , N 1

X Nv : x x, d 2n such that
N , s A for all s

u i u * , s K x i x * , s i,
i 0, . . . , N 1


The target state x *( , s) and target input u *( , s) are given by

Eqs. 8 and 9.
Remark 9. Note that the choice of cost (Eq. 23) corresponds to assuming the disturbance sequence is constant over
the control horizon and equal to the current estimate, and
similarly with the set point. However, note that in the denition
of the constraints on the input perturbations (Eq. 21), it is
assumed that the disturbance and set point may change over the
prediction horizon. As will be shown later, these two facts can
be combined to guarantee, respectively, offset-free control in
the limit and robust constraint satisfaction during transients.
The minimizer of the FHOCP (Eq. 22) is dened as
v*, s : *0 , s, . . . , *N1 , s : argmin VN , s, v
v N , s

As is standard in receding horizon control, for the current
augmented state and set point s, we only keep the rst
element *0 ( , s) of the solution to the FHOCP. Using this
receding horizon principle, we dene our receding horizon
control input as


December 2005


As will be shown later, X Nv is the set of plant states in for

which the controlled variable will ultimately be driven to the
set point s by the receding horizon controller.
We can now give our rst main result:
Theorem 3 (Domain of RHC law). Suppose that Assumptions 15 hold. If X 0 is dened as in Eq. 18 and each X vj , j
{1, . . . , N}, is dened as in Eq. 27 with N j, then all the sets
in {X 0 , X v1 , . . . , X Nv } contain the origin in their interior and
X 0 X 1v XN1


Proof. See the Appendix.

Theorem 3 is very important because it shows that, under the

above assumptions, the set of states for which the FHOCP (Eq.
22) has a solution is nonempty, and an increase in the horizon
length does not decrease the size of the set of initial plant states
for which the controlled variable can be driven to the set point.
Furthermore, it also implies that the domain of attraction of the
receding horizon controller contains the domain of attraction of
the linear (unconstrained) controller dened in the previous
We now proceed to give conditions under which one can
guarantee offset-free control in addition to robust constraint
satisfaction. For this purpose, we introduce the following assumption:
Vol. 51, No. 12

AIChE Journal

Assumption 6. The matrices Q and R are chosen to be

positive denite, the matrix P is the positive denite solution of
the discrete algebraic Riccati equation P Q A T PA
A T PB(R B T PB) 1 B T PA, and the matrix K is the corresponding gain K (R B T PB) 1 B T PA. is given by Eq.
11 with given by Eq. 13, : and is given by
Eq. 13.
Before giving our second main result, we need the following
two lemmas, which are generalizations of results in:17
Lemma 4 (FHOCP equivalence). Suppose that Assumptions 1 and 6 hold. If the matrix W mm is given by W :
N1 T
R B T PB, then v*( , s) argminv{ i0
i W i v N ( ,
Proof. A similar result, for robust receding horizon controllers that do not provide offset-free control, is well-known,17
Rem. 3. A detailed proof for the offset-free receding horizon

controller proposed in this article is reported in.36

Lemma 5 (Robust feasibility and perturbation sequence).
Suppose that Assumptions 13 and 5 6 hold. If the set
N ( (0), s(0)) is nonempty, then the set N ( (k), s(k)) is
nonempty for all k and
lim *0 k, sk 0

By applying Proposition 1 one can compute an equivalent

expression for N ( , s) in terms of a nite number of afne
inequality constraints
N , s v mN Fv c H Hs s


c : b min Gd d min Gs s



d N

s N1

Since and (and, hence, N and N1 ) are polytopes

and can, therefore, be described by a nite number of afne
inequality constraints, c can be computed efciently by solving
q LPs. However, if and are given only by upper and lower
bounds on the components of d and s, respectively, then it is
not necessary to solve LPs in order to compute c; computing
the absolute values of the components of G d and G s is sufcient. For example, let the disturbance and the set point take on
values in the hypercubes : {d r d } and :
{s p s }. Recall now that the -norm is the dual
norm of the 1-norm,39 that is, a 1 max{a T x x 1} for
any vector a. Hence, it is easy to show that


Proof. The proof follows fairly standard arguments in re

ceding horizon control and is reported in.36
Theorem 4 (Offset-free control, robust constraint satisfaction and stability of RHC). Let Assumptions 13 and 5 6
hold. One can choose the initial controller state ( x (0), d (0))
such that the FHOCP (Eq. 22) has a solution, and the evolution
of the augmented system (Eq. 7) in closed-loop with the receding horizon control (Eq. 26) satises Eq. 4 if and only if the
initial plant state x(0) X Nv . Furthermore, : (I
) 1 (d s ) is the robustly asymptotically stable xed
point of Eq. 20 if it is in the interior of .
Proof. See the Appendix.

c b absGd 1 absGs 1

where abs(M) is the matrix of the absolute values of the

corresponding components of the matrix M and 1 is a column
vector of ones of appropriate length.
It is also important to note that the number of constraints q
in Eq. 30 is not dependent on the description for and , but
only dependent on N, and the number of constraints that
describe , and . As a matter of fact, it is easy to show
that the number of constraints q O(N).
Given all of the above, it is now clear that the solution to the
FHOCP (Eq. 22) exists, if and only if N ( , s) A. The
solution of the FHOCP (Eq. 22) is the solution to the following
tractable, strictly convex QP

Implementation of the receding horizon controller

Recall from Eq. 21 that the constraints on the input perturbations have to hold for all allowable set point and disturbance
sequences. In this section, we point out the fact that the
(innite) set of constraints in Eq. 21 is easily rewritten in terms
of a nite and tractable set of linear inequality constraints. This
then allows one to compute the receding horizon control action
by setting up and solving a tractable, strictly convex quadratic
program (QP) at each sample instant.
Since , and are polyhedral sets with nonempty
interiors, they are given by a nite number of afne inequality
constraints. As a consequence, one can obtain an expression for
the set of admissible input perturbations N ( , s) as
N , s v mN Fv b Gd d Gs s H
Hs s, s N1 , d N (29)
where the matrices F qmN , G d qrN , G s
qp(N1) , H q3n , H s qp , and the vector b q
depend on the augmented system dynamics (Eq. 20) (see 36 for
the exact expressions).
AIChE Journal

December 2005



v*, s argmin

iT Wi Fv c H Hs s



where W : R B T PB as in Lemma 4.

Illustrative Example
Process and constraints
As an example, we consider a jacketed continuous stirred
tank reactor (CSTR) studied by Henson and Seborg40 in which
an irreversible liquid-phase reaction occurs. A detailed nonlinear model has two states (reactant concentration and reactor
temperature), one input (cooling liquid temperature) and two
disturbances (feed temperature and feed reactant concentration). This CSTR shows three steady states, two of which are
open-loop unstable, and for quality and safety reasons the
middle conversion open-loop unstable steady-state is chosen as
a nominal operating set point. Using a sampling time of t s
0.1 min and introducing deviation variables (from the corresponding steady state) a linearized model is as follows
Vol. 51, No. 12


state ( x (0), d (0)) (as well as the initial perturbation term) is

computed from the following strictly convex QP

x 0, d 0, v*0, s0 : argmin ((x x(0))T (x

x ,d

W Fv c H (0) H s(0)


x(0)) dT d)



Figure 2. Domain of attraction (XN
) for different xed

0.7776 0.0045 x k
xx kk 11 26.6185
1.8555 x k
0.0002 0.0893 d k
0.2907 uk 0.1390 1.2267 d k

in which 1,000.
For the receding horizon controller based on Q I 2 and
R 0.2, the plant state sequence, x, is also shown in Figure
2. Notice that the state sequence x initially starts at the
boundary of the domain of attraction X v10 and enters the domain
of attraction of the linear controller X 0 in nite time. As
expected from Theorem 4 the proposed controllers asymptotically drive the controlled variable to the asymptotic set point
despite the presence of persistent unmeasured disturbances.
Also, when the set point is changed the controllers drive the
controlled variable to the new set point. Moreover, it is interesting to notice that the choice of penalty matrices has a direct
impact on the closed-loop performance. As expected, when a
lower input penalty R is chosen, the disturbance is rejected

x 1k
zk 0 1 x k

in which x 1 and x 2 represent the reactant concentration and the

reactor temperature, respectively; u represents the coolant temperature; d 1 and d 2 represent the feed temperature and the feed
reactant concentration, respectively. Notice from the structure
of C z that the controlled variable is the reactor temperature, for
which offset-free control to the set point s is required. The
following constraints on the plant states and input and on the
admissible disturbances and set point are considered
15 u 15,
5 x 5 ,
dd 0.12 ,

1 s 1

Results and comments

We present in Figure 2 the domain of attraction (that is, X Nv )
of four receding horizon controllers using different xed horizons (specied in the gure) and the same penalty matrices:
Q I 2 and R 0.2. Notice that X 0 is the domain of attraction
of the linear controller. As expected from Theorem 3 we have
that an increase in the xed horizon length results in a larger
feasible region, and also that the domain of attraction of the
linear controller is included in that of the receding horizon
We present in Figure 3 the closed-loop simulation results
(controlled variable and input) obtained with four receding
horizon controllers using the same xed horizon, N 10,
different penalty matrices (Q I 2 for all controllers and R
specied in the gure). The initial plant state is x(0)
[0.258 5] T , the disturbances and the set point vary during
the simulation time as reported in Table 1. The initial controller

December 2005

Figure 3. Closed loop comparison of different receding

horizon controllers (linear plant): controlled
variable (top) and input (bottom).
Vol. 51, No. 12

AIChE Journal

Table 1. Disturbances and Set Point

t (min)

[0, 4)

0.1] T

[4, 8)

0.1] T

[8, 12)

[12, 16)

0.1] T

0.1] T


[16, 24)

0.1] T

[20, 24]

0.1] T

(and the set point is reached) more quickly and a larger control
input is used.
In Figure 4 we present the closed-loop simulation results
obtained with the same four controllers as in Figure 3 when the
plant is described by the original nonlinear system.40 From
these results it is clear that the proposed controllers are able to
achieve offset-free control even when there is a mismatch
between the plant model, and the actual plant dynamics. As
mentioned after Lemma 3, as long as the closed-loop system is
stable, it is possible to show that offset-free control holds
independently of the actual plant dynamics.
We nally present in Figure 5 a comparison of the proposed
receding horizon controller with a standard (that is, nonoffset-free) robust receding horizon controller. As an example we
chose the approach in,17 which is similar to the one proposed in
this article, in the sense that a prestabilizing gain matrix is used,
and the plant state prediction at the end of the horizon is
restricted to be in the maximal disturbance invariant set .
Both controllers are based on the same stabilizing gain matrix
K, which is the optimal LQR gain with Q I 2 and R 0.2.

The xed horizon used for both controllers is N 10 and the

perturbation penalty for the standard controller is chosen as
W R B T PB with P the solution to the corresponding
steady-state Riccati equation. The initial plant state is x(0)
[0.258 5] T and the disturbance varies as specied in Table
1. In this comparison the set point is the origin since the
method in 17 does not apply to set points different from the
origin (an extension of 17 to the set point tracking problem has
been proposed in 16; however, the controller proposed in 16 still
does not guarantee offset-free control). As expected, the goal of
offset-free control is achieved by the proposed method whereas
the controller of 17 leaves a signicant and undesired steadystate offset.

Figure 4. Closed loop comparison of different receding

horizon controllers (nonlinear plant): controlled variable (top) and input (bottom).

Figure 5. Closed loop comparison of offset-free and

standard robust receding horizon controllers:
controlled variable (top) and input (bottom).

AIChE Journal

December 2005

This article has shown how one can design a dynamic
receding horizon controller that guarantees robust constraint
satisfaction, robust stability and offset-free control in the pres-

Vol. 51, No. 12


ence of asymptotically constant disturbances and set points.

The design of the controller was split into two parts:
The design of a dynamic linear time-invariant controller.
A deadbeat observer is used to estimate the disturbance, the
new steady state is given as a linear function of the current
plant and observer states and of the current set point, and the
controller aims to regulate the plant state and input to the new
target steady-state.
The design of a dynamic nonlinear time-invariant receding horizon controller. In order to increase the region of
attraction of the linear controller a robust receding horizon
controller, which computes perturbations to the linear control
law, was proposed. The receding horizon controller includes
the state and input constraints explicitly in its computations, as
well as the transient effect of the unknown disturbance and
time-varying set point on the target calculator and closed-loop
response, thereby guaranteeing robust constraint satisfaction. It
was shown that the specic formulation of the proposed receding horizon controller improves on the linear controller in
terms of the domain of attraction. The proposed controller is
computationally tractable since one has to solve, at each sampling time, a QP whose dimension increases linearly with an
increase in the horizon length.
This article also demonstrated the effectiveness of using the
results in this paper in designing a controller for guaranteeing
offset-free control of a continuous stirred tank reactor with
respect to existing non offset-free algorithms. The simulation
results were shown to be in agreement with the theory.
We conclude this article with some recommendations on
how the results in this article may be extended:
The choice of auxiliary system (that is, observer) has an
impact on the region of attraction and closed-loop performance
of the system. A more detailed investigation into this topic
could be undertaken.
The constraints on the state and input were not included in
the target calculation in Eq. 8. If the constraints are included in
the target calculation, then the optimal steady-state target is no
longer a linear function of the augmented state and set point.
Clearly, this complicates the receding horizon controller design. However, the inclusion of constraints in the target calculation will enlarge the domain of attraction and increase the
size of the disturbance and set points that can be handled by the
controller. An extension of this article, which includes constraints in the target calculation, could combine the results in 28
with those in.16
Due to the requirement that all set point sequences need to
be handled, the approach presented in this article is potentially
conservative, since the maximal constraint-admissible robust
positively invariant set may be small or empty for the given
range of set points. Further work could involve removing this
source of conservativeness to allow a larger range of set points
to be tracked without offset. Once again, this may be possible
by combining results in 28 with those in.16
Clearly, the rank condition in Eq. 3 is not always satised.
If this assumption is violated, then one might have to relax the
requirement that offset-free control be achieved on all controlled variables. One possible approach to resolving this problem is to prioritize the controlled variables when performing
the target calculation. The framework proposed in 41 may be
useful in this context.
Rather than optimizing over perturbations to a prestabi3144

December 2005

lizing control law, one could consider optimizing over feedback policies.2,3,15,19,22 This will enlarge the region of attraction
of the receding horizon controller at the expense of an increase
in computational complexity.
The important problem of guaranteeing robust stability,
performance, constraint satisfaction and offset-free control
when output feedback (rather than state feedback) is used,
remains to be addressed.

Literature Cited
1. Blanchini F. Set Invariance in Control. Automatica. 1999;35:1747
1767. Survey paper.
2. Mayne DQ. Control of constrained dynamic systems. Euro J of Control. 2001;7:8799. Survey paper.
3. Mayne DQ, Rawlings JB, Rao CV, Scokaert POM. Constrained model
predictive control: Stability and optimality. Automatica. 2000;36:789
814. Survey paper.
4. Findeisen R, Imsland L, Allgower F, Foss BA. State and output
feedback nonlinear model predictive control: An overview. Euro J of
Control. 2003;9:190 206. Survey paper.
5. Morari M, Lee JH. Model predictive control: Past, present and future.
Computers & Chem Eng. 1999;23:667 682. Survey paper.
6. Blanchini F. Ultimate boundedness control for uncertain discrete-time
systems via set-induced Lyapunov functions. IEEE Trans Automatic
Control. 1994;39:428 433.
7. Mayne DQ, Schroeder WR. Robust-Optimal Control of Constrained
Linear Systems. Automatica. 1997;33:21032118.
8. Bemporad A, Casavola A, Mosca E. Nonlinear Control of Constrained
Linear Systems via Predictive Reference Management. IEEE Trans.
Automatic Control. 1997;42:340 349.
9. Bemporad A. Reference governor for constrained nonlinear systems.
IEEE Trans. Automatic Control. 1998;43:415 419.
10. Bemporad A, Mosca E. Fullling hard constraints in uncertain linear
systems by reference managing. Automatica. 1998;34:451 461.
11. Gilbert EG, Kolmanovsky I. Nonlinear tracking control in the presence
of state and control constraints: A Generalized Reference Governor.
Automatica. 2002;38:20632073.
12. Gilbert, EG, Kolmanovsky I, Tan KT. Discrete-time reference governors and the nonlinear control of systems with state and control
constraints. Int J of Robust and Nonlinear Control. 1995;5:487504.
13. Badgwell TA. Robust model predictive control of stable linear systems. Int J of Control. 1997;68:797 818.
14. Bemporad A. Reducing conservativeness in predictive control of constrained systems with disturbances. In: Proc. 37th IEEE Conference on
Decision and Control; Tampa, FL. 1998.
15. Bemporad A, Borrelli F, Morari M. Min-max control of constrained
uncertain discrete-time linear systems. IEEE Trans Automatic Control.
2003;48:1600 1606.
16. Chisci L, Zappa G. Dual mode predictive tracking of piecewise constant references for constrained linear systems. Int J Control. 2003;
17. Chisci L, Rossiter JA, Zappa G. Systems with persistent disturbances:
predictive control with restricted constraints. Automatica. 2001;37:
1019 1028.
18. Kerrigan EC, Maciejowski JM. On Robust Optimization and the
Optimal Control of Constrained Linear Systems with Bounded State
Disturbances. In: Proc. European Control Conference: Cambridge,
U. K.; 2003.
19. Kerrigan EC, Maciejowski JM. Feedback min-max predictive control
using a single linear program: Robust stability and the explicit solution. Int J of Robust and Nonlinear Control. 2004;14:395 413.
20. Lee YI, Kouvaritakis B. Constrained receding horizon predictive control for systems with disturbances. Int J Control. 1999;72:10271032.
21. Mayne DQ, Langson W. Robustifying Model Predictive Control of
Constrained Linear Systems. Electronics Lett. 2001;37:14221423.
22. Scokaert POM, Mayne DQ. Min-Max Feedback Model Predictive
Control for Constrained Linear Systems. IEEE Trans Automatic Control. 1998;43:1136 1142.
23. Rawlings JB, Meadows ES, Muske KR. Nonlinear Model Predictive
Control: a Tutorial and Survey. In: AD-CHEM Conference: Kyoto,
Japan; 203214 1994.

Vol. 51, No. 12

AIChE Journal

24. Maciwjowski JM. The implicit daisy-chaining property of constrained

predictive control. Appl Math and Comp Sci. 1998;8:695711.
25. Magni L, De Nicolao G, Scattolini R. Output feedback and tracking of
nonlinear systems with model predictive control. Automatica. 2001;
26. Magni L. On robust tracking with non-lineare model predictive control. Int J Control. 2002;75:399 407.
27. Muske KR, Badgwell TA. Disturbance Modeling for Offset-Free Linear Model Predictive Control. J Proc Cont. 2002;12:617 632.
28. Pannocchia G, Rawlings JB. Disturbance Models for Offset-free
Model Predictive Control. AIChE J. 2003;49:426 437.
29. Sakizlis V, Dua V, Perkins JD, Pistikopoulos EN. Robust ModelBased Tracking Control using Parametric Programming. Comp &
Chem Eng. 2004;28:195207.
30. Sakizlis V, Kakalis NMP, Dua V, Perkins JD, Pistikopoulos EN.
Design of Robust Model-Based Controllers via Parametric Programming. Automatica. 2004;40:189 201.
31. Kolmanovsky I, Gilbert EG. Theory and Computation of Disturbance
Invariance Sets for Discrete-Time Linear Systems. Mathematical
Problems in Engineering: Theory, Methods and Applications. 1998;4:
32. Scokaert POM, Rawllings JB, Meadows ES. Discrete-time Stability
with Perturbations: Applications to Model Predictive Control. Automatica. 1997;33:463 470.
33. Kwakernaak H, Sivan R. Linear Optimal Control Systems. Wiley;
34. Muske KR, Rawlings JB. Model Predictive Control with Linear Models. AIChE J. 1993;39:262287.
35. Nocedal J, Wright SJ. Numerical Optimization. Springer; 1999.
36. Pannocchia G, Kerrigan EC. Offset-free receding horizon control of
constrained linear systems subject to time-varying set points and
persistent unmeasured disturbances. Tech. Rep. CUED/F-INFENG/
TR.468 Department of Engineering, University of Cambridge; Cambridge, UK; 2003.
37. Francis BA, Wonham WM. The internal model principle of control
theory. Automatica. 1976;12:457 465.
38. Kerrigan EC, Maciejowski JM. Properties of a new parameterization
for the control of constrained systems with disturbances. In: Proc.
American Control Conference: Boston, MA; 2004.
39. Horn RA, Johnson CR. Matrix Analysis. Cambridge University Press,
UK; 1985.
40. Henson MA, Seborg DE. Nonlinear Process Control. Upper Saddle
River, New Jersey: Prentice Hall PTR, 1997.
41. Kerrigan EC, Maciejowski JM. Designing model predictive controllers
with prioritised constraints and objectives. In: Proc. IEEE Conference
on Computer Aided Control System Design: Glasgow: Scotland; 2002.

Appendix: Selected Proofs

From Eq. A2c we immediately obtain that x x (note that

x x , and, hence, the rest of the proof holds independently
of Eq. A2a that is, independently of the actual plant dynamics
and disturbance), which combined with Eq. A2b leads to
x Ax Bu x x d


Let ( x , u ) denote the solution to the target calculation

problem (Eq. 8) for the augmented state , and the set point s .
From Eq. 8b and Eq. A3 we can write
x x A x x Bu u A BK x x
where the last step comes from Eq. 10. It is important to notice
that Eq. A4 and Assumption 4 imply that
x x


Finally, from Eq. A5 and from Eq. 8b we obtain

s C zx C zx lim k

Proof of Theorem 2
Robust constraint satisfaction follows immediately from the
fact that is robust positively invariant for the closed-loop
system (Eq. 15), and the fact that is constraint-admissible.
Since is strictly stable, (I ) 1 exists and hence
is well-dened and unique. Note also from the proof of Lemma
3 that : limk3 (k). Robust asymptotic stability
follows from Theorem 1 by dening

: ,

w : d d s s

Hence, we can write the closed-loop system dynamics in

terms of the shifted variables as (k 1) (k)
w(k). The proof is completed by noting that limk3 w(k) 0.

Proof of Lemma 3
Since limk3 s(k) s and limk3 d(k) d we have from
Eqs. 14 and 15 and from Lemma 2 that

: lim k d s u d ,

in which u s . Let be partitioned as follows:
( x , x , d ) in which each block is a column vector of
length n. We can rewrite Eq. A1 explicitly as follows
x Ax Bu Ed


x Ax Bu x x d


d x x d


AIChE Journal

December 2005

Proof of Theorem 3
Although a result, similar to the one stated here, appears to
be known 17, Sect. 4.2, we have been unable to nd a detailed
proof in the literature. Classical robust open-loop receding
horizon control,3 Sect. 4.5 is well-known to exhibit infeasibility
problems if the plant is open-loop unstable and no prestabilizing policy is used in the predictions (Eq. 22). However, it is a
remarkable fact that one can remove this problem by optimizing over a sequence of perturbations to a prestabilizing control
law. To show that this is indeed still true for the control
algorithm proposed in this article, we present a detailed proof.
It follows trivially from Assumption 5 that X 0 contains the
origin in its interior. The rest of the proof is by induction.
Let the plant state x X vj , where j {1, . . . , N 1}, the
controller state ( x , d ) be such that j ( , s) is nonempty and vj
: ( 0 , . . . , j1 ) j ( , s) be an admissible perturbation
sequence of length j. Also, let sj1 : (s 1 , . . . , s j1 ) j1
and dj : (d 0 , . . . , d j1 ) j be allowable set point and
disturbance sequences of length j 1 and j, respectively.
Vol. 51, No. 12


From the denition of j ( , s), it follows that j for

all sj1 j1 , and all dj i . Recall that is disturbance
invariant and constraint-admissible for the closed-loop system
(Eq. 15), hence, is disturbance invariant and constraintadmissible for system (Eq. 20) under the innite perturbation

sequence { (k)} k0
: {0, 0, . . .}.
It follows that if j for all sj1 j1 , and all dj
i , then j1 for all si i , and all di1 i1 .
This implies that if vj j ( , s), then (vj , 0) j1 ( ,
s). Hence, if j ( , s) is nonempty, then j1 ( , s) is nonempty. It follows from the denition of X vj that if x X vj , then
x X j1
, hence, X vj X j1
Using similar arguments as above, the result is completed by
noticing that X 0 X v1 .

Proof of Theorem 4
Sufciency. Suppose that x(0) X Nv , then it immediately
follows from Eq. 27 that for any initial set point s(0) one
can choose a controller state ( x (0), d (0)) such that N ( (0),
s(0)) A, and, hence, the FHOCP (Eq. 22) has a solution.
This implies from Lemma 5 we have that N ( (k)) A for all
k , and also that

: lim k : lim *0 k, sk 0


The fact that Eq. 4a holds can now be shown exactly as in

the proof of Lemma 3, since from Eqs. 20 and A6 it follows

lim k d s d s

u d
in which u s s .
The fact that Eq. 4b holds follows trivially from Lemma 5,
and the denition of N .
Necessity. This is obvious because if x(0) X Nv , then we
either have that x(0) , or that there exists an s(0)
such that for all ( x (0), d (0)) 2n , N ( (0), s(0)) A,
and, hence, the control input is undened at time 0.
Finally, robust asymptotic stability follows from Theorem 1
and can be shown in a similar fashion as in the proof of
Theorem 2. This is because it is easy to show that for any
, the optimal perturbation *0 ( , s) 0 for all s .
Hence, we can write the closed-loop system dynamics in a
neighborhood of , in terms of the shifted variables, as (k
1) (k) w(k). Again, the proof is completed by noting
that limk3 w(k) 0.



December 2005

Manuscript received Dec. 8, 2004, and revision received Jun. 3, 2005.

Vol. 51, No. 12

AIChE Journal

Passivity Based Control of Transport

Reaction Systems
Martin Ruszkowski, Vianey Garcia-Osorio, and B. Erik Ydstie
Chemical Engineering Dept., Carnegie Mellon University, 5000 Forbes Avenue, Pittsburgh, PA 15213
DOI 10.1002/aic.10543
Published online September 13, 2005 in Wiley InterScience (www.interscience.wiley.com).

Inventory control is based on the idea manipulating process ows so that the inventories
follow their set points. The operator mapping ows to inventories in a coarse grained
(macroscopic) system is passive and any input strictly passive (ISP) feedback controller can,
therefore, be used to achieve input-output stability. Examples of ISP controllers include the
PID controller, parameter adaptive feedforward control, optimal controllers and many
nonlinear and gain scheduling controllers. Input-output stability and convergence of inventories, such as total mass and energy can be used to show that other internal state variables
are bounded and belong to an invariant set. However, this does not necessarily imply
convergence of all state variables to stationary valves because a stabilizability condition,
known as strict state passivity, must be satised. The 2nd law of thermodynamics is used to
develop sufcient conditions for strict state passivity in the space of intensive variables. The
theory relies on following two assumptions concerning nonequilibrium systems: (1) The
hypothesis of local equilibrium, and (2) That a local entropy is dened using semiclassical
statistical mechanics. These assumptions allow us to dene a local entropy function for the
coarse grained system which is homogeneous degree one, concave, and has positive temperature. Subject to these conditions stability theory is developed for conjugate variables, such as
temperature, pressure and chemical potential in innite dimensional reaction-diffusion-convection systems. In classical irreversible thermodynamics this type of analysis assumes
linearity and symmetry of transport relations. The use of Gibbs tangent plane condition allows
us to dene a Lyapunov like storage function for passivity design, which gives stability criteria
for nonlinear problems. The resulting sufcient condition for stability can be expressed in
terms of dimensionless groups, similar to the (second) Damkohler number. Simple simulation
examples illustrate the application of the theory. 2005 American Institute of Chemical Engineers
AIChE J, 51: 3147-3166, 2005

Keywords: process control, coarse graining, distributed parameter system, irreversible

thermodynamics, feedback, passivity, invariant set, stability, entropy, reciprocity relations, nonlinear, convexity, minimum entropy production.

In the context of control of distributed parameter systems in
chemical engineering, four main categories of control problems
Correspondence concerning this article should be addressed to B. E. Ydstie at

2005 American Institute of Chemical Engineers

AIChE Journal

December 2005

and their applications were reviewed by Christodes (2001).1

They included (1) control of spatial proles with application to
packed-bed reactors, CVD and etching, (2) control of size
distributions with applications in aerosol production, crystallization and emulsion polymerization, (3) control of uid ows
with application in uid mixing, wave suppression and drag
reduction, and (4) control of material microstructures with
applications for thin lm growth and nanostructured coatings.
Vol. 51, No. 12


This work expands on the list since the objective is to control

inventories by adjusting ows. The inventory of a transportreaction process (TRP) is an integral measure that represents
physical properties like total mass, energy or the holdup of a
chemical constituent. The inventory is controlled by adjusting
the ows at the boundary or the reaction rate. Inventory control
is closely related to mass and energy balance control (Buckley,
1964; Shinskey, 1967),2,3 and extensive variable control (Georgakis, 1986).4 The method was developed to its nal form by
Farschman et al. (1998)5 who used Lyapunov theory to show
that inventory control systems with PID feedback are indeed
input-output stable.
In this current article we show that the inventory control
method can be applied to coarse-grained TRPs using strictly
input passive feedback control. We also show that latent (intensive) variables like temperature, pressure and chemical potentials converge. A distinctive feature of the control problem
at hand is that control may only be possible at one or only a few
points at the boundary. Examples where this type of control
may nd application include: plasma etching for microelectronics processing, moving and xed-bed reactors for chemical
and metallurgical processes, bioreactors, membrane and hollow
ber reactors for nanoprocessing.
Many methods have been proposed for control of TRPs
(Ray, 1978; Christodes, 2001).6,1 In one approach the partial
differential equations are spatially discretized to form large
systems of ordinary differential equations using the method of
lines. Controllers can then be synthesized using classical theory
like pole placement, linear quadratic or model predictive control. However, neglecting the innite dimensional nature of the
original system may lead to erroneous conclusions concerning
the stability, controllability and observability properties of the
system since the number of modes required to derive a nite
dimensional ODE systems that yields a desired degree of
approximation can in fact be very large (Aling et al., 1998;
Balas, 1986; Christodes, 2000).7,8,9
Methods have been developed to reduce the dimensionality
of the system before the controllers and observers are synthesized to reduce the complexity of the controller design. One
very natural framework for deriving low-dimensional ODE
systems that accurately reproduce their solution is based on the
concept of inertial and approximate inertial manifolds (IM/
AIM). The IM appears as the dominant solution to a highdimensional input-output system under feedback when all but
very few modes are heavily damped. In such systems the
high-order dynamics dissipate and control only need to take
effect in a low dimensional space. For example, the dynamics
of a parabolic PDE system is restricted on an inertial manifold
described by a set of ODEs (Temam, 1988; Brown et al., 1990;
Smoller, 1996).10,11,12 Systematic methods for the derivation of
the inertial form has been proposed by Deane et al., 1991;
Christodes and Daoutides (1997), and Kazantzis and Good
The Lyapunov method for control system design is based on
the idea of dening an energy related function which captures
the dynamics of the state and choosing a controller to make the
energy function decrease. A review of this approach applied to
nite dimensional systems can be found in the books by Slotine
and Li (1992) and Sontag (1998).16,17 The method has been
applied to control for robust control of PDE systems (Christodes and Daoutidis, 1998, and Christodes, 1998).18,19 More

December 2005

recently El-Farra (2003)20 extended the method to cover systems with input constraints, whereas Kazanzis and Kravaris
(1999)21 related the energy based method to predictive control.
Sliding mode control (Palazoglu and Owens, 1987; Hanczyc
and Palazoglu, 1995; Sira-Ramirez, 1989)22,23,24 has been applied to control TRPs. The main idea behind this technique is
to let a discontinuous feedback controller dene a so-called
sliding surface on which the system has desirable behavior. A
Lyapunov-like stability condition guarantees that the distance
to the sliding surface decreases exponentially along all system
trajectories. Application of sliding mode control, a form of high
gain feedback (the feedback gain is innitely large), requires
that the input-output pairing (or transformations thereof) are
passive. Finally, Godasi et al. (2002)25 develop a symmetry
group representation which they use to develop control methods for PDE systems.
The inventory approach for stabilization of TRPs distinguishes itself from the methods mentioned previously in that
we develop stabilizing controllers using macroscopic (integrated) properties for feedback. The method is related to the
averaging methods used to study nonlinear systems and stochastic processes since the use of inventories leads to smearing of the detailed structure of the dynamics. However, in the
inventory approach we obtain the reduced order model by
averaging over space rather than time. This gives a coarse
grained view of the system dynamics and in this way the
method can be linked to the classical methods of irreversible
thermodynamics and statistical mechanics. The resulting manifold is described by a small number of ODEs (one for each
control objective) derived from the macroscopic conservation
equations. The nal control design is achieved using either
Lyapunov or passivity design techniques. A description of this
approach and its application to nite dimensional systems with
and without input constraints is given by Farschman et al.
(1998).5 Ydstie and Jiao (2004)26 applies inventory and ow
control to a oat glass production system and Ruszkowski et al.
(2005)27 applied inventory and ow control to real-time optimization of a silicon furnace.
Passivity, a concept derived from electrical circuit theory, is
encountered in many scientic disciplines (Desoer and Vidyasagar, 1975).28 The fundamental idea is based on dissipation
of useful work by irreversible transformations and input-output
pairings with phase shift no more than /2. Willems
(1972a,b)29,30 developed a systems perspective for dissipativity
and linked the concept to state space representations. Byrnes et
al. (1991)31 showed that passivity and Lyapunov stability is
equivalent for a class of feedback systems using geometric
methods. Van der Schaft (1996)32 developed control methods
linking passivity and L2 stability, while Slotine and Li (1992)16
and Ortega et al. (1998, 2001)33,34 stressed passivity based on
physical arguments following from the law of conservation of
energy. Krstic et al (1995)35 linked passivity and nonlinear
adaptive control. Ydstie and Alonso (1997)36 advanced the idea
of combining thermodynamics and passivity and showed that
passivity could be motivated using a storage function related to
the available work, and Gibbs tangent plane criteria for phase
stability. Alonso and Ydstie, (2001)37, Ydstie (2002)38, and
Coffey et al. (2000)39 highlighted the importance of concavity
of the entropy function, and used the Poincare inequality to
estimate the impact of the size on stability of the system.
The main objectives of the current article are: (1) to give a
Vol. 51, No. 12

AIChE Journal

tutorial overview of passivity based control, and (2) to show

that passivity based methods are effective in stabilizing innite
dimensional transport reaction systems. We develop this result
using a Lyapunov-like energy function, which exploits macroscopic, as well as microscopic system properties. An important technical contribution of the article is that we exploit the
homogeneity of the entropy function more fully than we have
achieved in previous articles.

Classical Thermodynamics
Let , a convex subset of R
, be called the space of extensive variables and let Z denote an arbitrary point. We can for
example have Z (U,V,M1,. . .,Mn,X)T where U is the internal
energy, V is the volume, Mi is mass or moles of chemical
component i and X can correspond to the charge, degree of
magnetization, area, momentum et cetera.
Assumption A1 (The structure of equilibrium thermodynamics): There exists a C2 function, S : R, called the
entropy so that1:
1. For any positive constant we have S(Z) S (Z) (S
is positively homogeneous of degree one).
2. For all points Z 1 , Z 2 and any positive constant
we have
S Z 1 1 Z 2 SZ 1 1 SZ 2
(S is concave).
3. T (U/S) 0(the temperature is positive).
The developments below do not depend on a particular
denition of S since we do not actually need to evaluate the
entropy function numerically. It sufces that there exists an
entropy function which satises the axioms given earlier. We
also want to note that the assumption of concavity is not needed
for passivity design. It sufces that the entropy is over-bounded
by a concave function. This is demonstrated in the rst part of
the article by Ydstie and Alonso (1997).36
To make the discussions precise, consider the example of
semiclassical statistic mechanics. The microcanonical partition
function for a single component system of N structureless
particles is dened so that (Ruelle, 1969)40

U, V, N


U H p 3N, q 3Ndp 3Ndq 3N

where H the Hamiltonian, is Plancks constant (divided by

2), p and q are the momenta and positions, is a scaling
constant with units of energy and ( )is the Dirac delta
function. The entropy is now dened using Boltzmans principle so that
S k Bln U, V, N
Together these expressions give an entropy function which is
(1) rst-order homogeneous, (2) concave, and (3) has positive
temperature. All these properties are due to the denition of the
logarithm and Assumption A1 is, therefore, satised (Kestin
We follow the approach espoused by Caratheodory and Callen. This approach to
thermodynamics is based on a limited number postulates expressed in mathematical
form. The requirement that the entropy is twice differentiable is convenient but not
necessary to obtain the main results of the paper. In general it sufces that S is C1.

AIChE Journal

December 2005

Figure 1. Equilibrium thermodynamics can be viewed as

an input-output theory of systems.
Three actions (change of volume, heat and mass) provide the
inputs to the system. The system response is observed by a
change in the conjugate variables (pressure, temperature and
chemical potential). By dening the corresponding input vector u and output vector y we can represent the thermodynamic
processes as a control system as shown in the gure on the

and Dorfman, 1971).43 Methods to evaluate (U, V, N) numerically have recently been developed by Wang and Landau
(2001) and Shell et al. (2002).41,42
The implied function relating Z (U,V,N) to S given by the
two equations earlier is called a fundamental relation. In the
case of an ideal gas with constant heat capacity the result is
very clear since we get an explicit formula for the entropy so

S 2 S 1 c Vln

R ln

In more general cases it can be quite difcult to evaluate S

numerically. In nonequilibrium systems it may not be possible
to dene a unique entropy function since the concept of an
equilibrium state loses its precise interpretation. It is possible to
make good approximations, and the use of classical thermodynamic variables in nonequilibrium settings gives excellent predictions in all but very extreme cases, however. We will follow
an approximate approach to nonequilibrium theory called
coarse graining via the hypothesis of local equilibrium
(HLE). In this approach we grid the system into sub-regions
and make the assumption that Boltzmans principle holds for
each sub-region. The results of the article are based on the
premise that Assumption A1 and HLE hold.
A mapping P: is called a thermodynamic process
(Figure 1). For control purposes, it is useful to represent the
thermodynamic process by the input output operator S so that
y Su
We may, for example, have u (V, Q, M) T and y
((1/T), (P/T), ( /T)) T . Thus, we are interested in modeling
how intensive variables respond to a nite change in volume,
heat and mass ow. The fundamental equation provides the key
for constructing the operator S since it contains all information
needed to characterize how the variables y change in response
to a given action u. In the coarse grained approach we model
systems distributed in time and space by cascading simple
systems while making the assumption that each unit cell still
obeys the assumptions A1.1 A1.3. A coarse grained system
with an input-output operator S and dynamics constrained by a
Vol. 51, No. 12


fundamental relation satisfying axioms A1.1 A1.3 is called a

process system (Ydstie and Alonso, 1997).36
Consider now two points Z1, Z2 and let Z3 Z1Z3. The
second law of thermodynamics for a simple system can now be
stated in the following, and maybe more familiar manner
(Callen, 1985; Sandler, 1999)44,45
The Second Law: SZ3 SZ1 SZ2
This expression states that the entropy of combined systems at
equilibrium is never smaller than the sum of the entropies of
the subsystems. The following result shows that the classical
denition of the second law and Assumption A1.2 are in fact
Result R1 (Equivalence of concavity and the second law):
The entropy S(Z) is concave if and only if for all Z 1 , Z 2 Z
SZ 1 Z 2 SZ 1 SZ 2

Figure 2. (a) Shows a projection of an entropy function

and a unique point of stability indicated at the
point Z1. The slope of the tangent line, w1 denes the intensive variables at that point, and
(b) shows a projection, as it might arise in a
cubic equation of state like the Van der Waals
In this case the EOS gives three points where the slope, and,
hence, the intensive variables are the same. The actual entropy
is dened to be the smallest concave relaxation as indicated
by the straight line segment. Entropy functions below the line
are in violation of the second law.

Proof: For any positive constant

S Z 1 1 Z 2 S Z 1 S1 Z 2
Using homogeneity gives


S Z 1 1 Z 2 SZ 1 1 SZ 2
It follows that S is concave. The reverse argument follows
We now introduce the set of conjugate variables



dU dV
dM 1, . . . , dMn

Eulers theorem for homogeneous functions gives2


MZ 0



M ij

where the negative signs are introduced to follow the standard

notation. The intensive variables therefore dene the slope of
the entropy as shown in Figure 2. Using these denitions we
can also generate the Legendre transform of S so that
S*w w TZ SZ
Z is now regarded as a function of w and S*. We see that the
variables Z and w are dual. The space * {w: Z } is
called the space of intensive variables. Alonso and Ydstie
(2001)37 showed that taking the Legendre transform of S(w)
does not yield unique results due to the homogeneity of S. This
simply means that we can scale size of the systemwithout
changing the intensive variables.
From Eq. 1 we now have the Gibbs equation dS w T dZ,
which for a simple system we write out term by term so that


Using Eq. 3, and the chain rule we can now write (S/Z)
(w/Z) Z w T . It follows from Eq. 1 that

called the vector of intensive variables. For a simple system,

we can make the assignment
1 P
, , ,...,


December 2005

2 S
Zi Zj

The matrix M of second derivatives of S is symmetric and

nonpositive, as indicated. The symmetry gives rise to the
Maxwell relations, and the nullspace of M denes invariant
sub-spaces 0 of (Figures 2 and 3). Gibbs showed that
dim(0) is equal the number of phases present is a nonreactive,
simple system.
Equation 4 is often written on the form
0 dw TZ


For a simple system this gives the Gibbs-Duhem equation

(Sandler, 1999)45
Eulers theorem asserts that if f( x) f( x) then we must have f( x) (f/ x) x.

Vol. 51, No. 12

AIChE Journal

That A1 (Z2) is non-negative is easy to verify geometrically as

shown in Figure 2. From Eq. 8 and the Euler identity (Eq. 3) we
A 1Z 2 w 1TZ 2 SZ 2
Thus, we see that A is closely related to S* only that we have
xed w1. From the Gibbs tangent plane condition we have that
two points Z1 and Z2 are in equilibrium if and only if A1 (Z2)
0. It is easy to to see that this condition is equivalent to w1
w2. The function A, therefore, measures distance from an
arbitrary, xed reference point w 1 *3.
Figure 3. The mapping w PZ from the space of extensive to the space of intensive variables is not
one to one.
The line 1, corresponds to uniform scaling of the extensive
variables. Such a line maps a point w1 in * , indicating that
pressure, temperature and chemical potential remain invariant. Subregions with positive measure (regions in with
phase or reaction equilibrium ) map to points with zero
measure (in *). Controlling intensive variables neither
guarantees stability or convergence in the space of the extensive variables.

0 dw TZ d


V d

1, . . . , d n M

AIChE Journal

v is here the vector of coarse grained variables (inventories)
corresponding to the total internal energy, volume and number
of moles (mass) of each species so that
v T V t, U t, M 1t, . . . , M nt


and M i1
M i denotes the total mass (moles).
Neither the Helmholtz function F UTS, which is based
on xing the micro-canonical ensemble (U,V,N) nor Gibbs
function G H TS, which is based on xing the canonical
ensemble (T,P,N), provide convenient starting points for the
development of process control theory. Both functions fail for
similar reasons. The Helmholtz function fails because here is
no way to guarantee that N and U are xed unless the system
already is well controlled. The Gibbs function fails because we
need to develop methods to stabilize pressure and temperature
before we can use the stability theory. In practice these problems show up as a problem of lower-boundedness of G and F
in unconstrained systems (Coleman and Owen, 1974).46
Ydstie and Alonso (1997)36 proposed to overcome the problem of lower-boundedness by dening a Lyapunov function
which measure distance to a xed reference point in a stationary, open system. They used the microcanonical ensemble
description, and the concavity of the entropy as their point of
departure. This theory led to denition a Lyapunov function
related to thermodynamic availability and exergy dened so
that for any pair of points Z1, Z2 in we have

A 1Z 2 SZ 1 w 1TZ 2 Z 1 SZ 2 0

In order to introduce the theory of transport reaction systems

we must deal with the following problem: Classical thermodynamics does not consider the spatial dependency of the system.
One way to overcome this problem is to use coarse graining. In
this approach, we tessellate physical space into N subregions.
Assumption A1 is assumed to be valid for each subregion
provided that the tessellation is sufciently ne.
The macroscopic balances for each sub-region, independent
of scale, can be written on the form (Hangos and Cameron,


which turns out to play a particularly important role in our

developments. The specic (or molar) properties are dened so
n T
, M
1, . . . , M
Z Z/M V , U

Coarse Graining, Transport and Reaction


December 2005

The subscript t refers to total in order to distinguish the

inventories from extensive variables. The state of the subregions, as well as the overall system may then change continuously. The ows between the subregions will be governed by
classical transport theory so that



where fi gives the net rate of transport. The vector p represents

the rate of production. We then get the following (very nice)
decomposition for the thermodynamic process

vt v0


Note that A does not dene a metric on the space * since we do not necessarily
have symmetry. Consider two arbitrary points Z1 and Z2. We may then have A1(Z2)
A2(Z1), indicating that the distance measured in one direction is not equal to the
distance measured in the opposite direction.

Vol. 51, No. 12


can furthermore calculate the gradients of intensive variables which are needed to estimate the rate of energy and
mass ow between the subregions. HLE is very well tested
and holds even at very small scales and quite sharp gradients.
According to the denitions above we can approximate the
coarse grained system locally (Eq. 9) by a system of partialdifferential equations on the form


Figure 4. Hypothesis of local equilibrium leads directly

to the separation of thermodynamics, transport and reaction engineering into distinct
subelds of process systems engineering.

where f(x,t) represents the ux density, and (x,t) represents

the density of production. These variables are connected to the
macroscopic balances via the relationships

t fL, t f0, t
As illustrated in Figure 4 we see that in order to predict the
evolution of a process from its initial condition we need to
evaluate a line integral (the action) involving three classical
branches of chemical engineering: thermodynamics, transport
and reaction engineering.
In the coarse-grained system we have

By letting N 3 we can then dene the density of the

inventory i so that

v it

x, t zi x, tdVt i 1, . . . , n 2

x, tdx

where the sign chosen so that it is positive for ow into the

system and negative for ow out. We normally divide the ux
density into orthogonal components (in the sense of GibbsDuhem) so that
f f conv fdiff


and pt



fconv uz


u is the center of mass velocity. The division into diffusive and

convective terms can be motivated on physical grounds and
corresponds to a separation into Eulerian and Lagrangian components.
Consider now a process system with rst-order reaction to
illustrate application of HLE, and the balances given earlier.
Following the notation of Sandler (1999),45 we can make the
following assignments in Eqs. 14 and 15


where is the molar or mass density, zi is the local state, and

dVt is the measure of Vt (Kestin and Dorfman, 1971).43 In order
to maintain clarity of presentation we consider processes distributed in one spatial dimension only. We let x represent the
spatial dimension, let dx represent dVt, and we assume furthermore that functions of x and t dened on the domain [0,
L] [0, ) exist and are continuous.
Assumption A2: (The hypothesis of Local Equilibrium
(HLE)): For every 0, there exists 0, so that if Vt
then z(x, t) Z(x, t) .
This is the essence of HLE (Kreuzer, 1981; Astarita, 1990;
Ottinger, 2005).46,49,50 The main importance of HLE is that we
can make the substitution
z x, t Z x, t


This means that the local state z(x,t), and the specic properties Z are equivalent on small enough scales. We can
then use statistical/quantum mechanics to develop fundamental equations that link the energy and mass conservation
laws with thermodynamic quantities like temperature, pressure and chemical potentials. In course grained systems we

December 2005

z ,

f conv m

fdiff ,



the heat
We note that m
denotes the center of mass ow, Q
ow FDi the rate of diffusion, is the viscosity as used in the
Navier-Stokes equations and i is the local rate of chemical
reaction. The conservation law corresponding to volume (the
second entry) does not add information and is only introduced
for convenience4. The term P(u/ x) is the generation of
internal energy due to compression. This term is reversible and
The corresponding conservation law is given by ( V / x) (m
V / x)
(u/ x) which is true but uninteresting.

Vol. 51, No. 12

AIChE Journal

cancels in the 1D entropy balance as shown in Appendix5. The

term 2 (u/ x) 2 0 represents the irreversible generation of
heat due to viscous dissipation.
The chemical potential of a mixture can be expressed so that

i x, t i0T, P k BT lnai zi
where ai is the activity coefcient. A continuous relationship,
therefore, exists among concentrations, pressures, temperatures
and chemical potentials. Reaction rates expressed in terms of
concentrations can, therefore, be related to the chemical potentials and more generally to the afnities as shown in the
Appendix. For example, the rst-order reaction A 3 B with
Arrhenius rate expression can be written so that

x, t k 0expE/RTx, tcA x, t


The HLE allows us to express concentrations as continuous

functions of the intensive variables except across phase boundaries. It follows that expression Eq. 17 is (Lipschitz) continuous in the intensive variables w. If instead we consider a
bioreactor with substrate inhibition model, then we have

x, t

w w 1 w 2
X X 1 X 2

where subscript 2 refers to the time invariant reference system.

We can analyze stability by studying the trajectories of the
deviation system derived from Eq. 14 dened so that

using the function A as a Lyapunov like storage function. This
theory can be extended using the idea of contraction analysis as
introduced by Lohmiller and Slotine (1998)51.
Lemma L1 (Evolution equation for A): Suppose that assumptions A1 and A2 are satised. We then have
A fTw

f diff
X wT
Proof: Since A1 and A2 are satised we can use Eq. 13
and set z( x, t) Z ( x, t). We can now write

mc S
K s c S c S2/K i

1 z1 2 z2
w1 w2 T

where m is the maximum growth rate, cs is the concentration

of substrate in the reactor, and Km and Ki are the growth
constants and continuity still holds. More generally, for continuous functions (w) we can write

By using the conservation laws Eq. 14, and the deviation

variables (Eq. 19) we get


w 1 w 2 k 0w 1 w 2

We now use the identity ((fT w )/ x) f T (w / x) w T (f/

x) and Eqs. 18 and 19 to give

It follows that
w 1 w 2 Tw 1 w 2 k0 w1 w2 2

A fTw

f TX w T

for some constant k0.

We now investigate the stability properties of the conservation laws (Eq. 14) using the potential A (dened by Eq. 8) as
an energy function. For this development is convenient to
dene the natural set of driving forces




f f 1 f 2
1 2

In the general case the right hand side of the energy balance consists of additional
terms P: u i1
f diff i Fi where the rst term represents the conversion of
kinetic energy to internal energy through compression and viscous dissipation, and the
second term represents the conversion potential energy to internal energy through

December 2005

A x, tdx 0


This gives the following key result.

Corollary C1 (Dissipation equality for process systems):
f Tw 0L

We see that according to HLE we have X(x,t) equal to zero if

the system is at equilibrium. It is also convenient to dene the
deviation variables

AIChE Journal


fT X wT dx

Proof: Follows from Lemma L2 by integration.

We see from Corollary C1 that the storage function W
decreases if the righthand side is negative. We furthermore
notice that the righthand side consists of three distinct terms.
1. The term f T w L0 is due to deviations in boundary conditions.
represents deviations due to convection
2. The term f TX
diffusion and heat conduction.
Vol. 51, No. 12


3. The term w T represents deviations due to chemical

reaction and power of compression.
We will now review very briey the substance of linear
irreversible thermodynamics (LIT). LIT is based on modeling
diffusion ow, heat conduction and chemical reaction using
linear relations so that
f LX



where L and K are symmetric with nonnegative and all positive

eigenvalues, respectively. The entropy production due to diffusion and heat-conduction can now be written as XTLX which
denes the Riemannian metric, often referred to as the Rayleigh-Onsager dissipation function. Using linearity we can
f diff

wT wT Kw 0


where 0 is smallest eigenvalue of L. From Corollary C1 we

then have
f Tw 0L 0

X TX dx

Using the Poincare inequality (Straughan, 1992; Struwe,

1992)52,53 we can write
f Tw 0L 0 2

f TX w T dx f Tw 0L 0w Tw

Then we have (dW(t)/dt) 0 W(t) for some positive

constant 0.
Proof: Follows from Corollary C1 by integration and using
the fact that according the Result R2 and HLE we a positive
constant 0. so that W(t) 0 w T w 0.
W(t) is radially bounded and positive denite on *. It
follows that W is a Lyapunov function for the intensive variables in the nite dimensional case. This theory is explored
more fully in the context of process networks by Jillson and
Ydstie (2005).59 We also note that if the conditions above are
satised then we use the theory to establish the existence of an
inertial manifold. We now introduce the following assumption.
Assumption A3 (Sufcient conditions for dissipativity):
There exists constants kL and 0 0 so that
1 w T k Lw Tw

and 2 fT X 0 XT X

We see that condition (1) of Assumption A3 is satised if

(w) is Lipschitz continuous in w. The conditions are satised
if we use the Arrhenius or substrate inhibition rate described
above and Fourier-Fick theory for heat conduction and diffusion.
Corollary C2 (Sufcient condition for dissipativity): Suppose that Assumption A3 is satised. The dissipation condition
of Result 3 holds true if 0 and kL dened in Assumption A3
satisfy the inequality

w Tw dx

kL 0

From Result R2, we can now conclude that there exists a

constant 0 so that w T w A(Z ).. Hence

where L is the length.

Proof: The Poincare inequality for homogeneous boundary
conditions gives

f Tw 0L 0 0 2 Wt

X TX dx 0


It follows that W(t) is a (control) Lyapunov function (Sontag,

1990)54 for the intensive variables w , and stability follows by
choosing the boundary conditions so that f T w L0 0.
Prigogine developed a similar stability result using the second variation of entropy as a Lypunov function. His approach
gives local results and it cannot be applied to phase equilibria
since the stability conditions obtained from the second-order
variation, S2 are not topologically equivalent to those obtained
from the Gibbs tangent-plane method across phase boundaries
due to discontinuities in S. A further discussion of the framework of classical equilibrium thermodynamics, linear laws and
stability is given in the Appendix.
The remainder of the article does not rely on linearity and
symmetry. Instead we derive a broader class of stability conditions which can be tried for nonlinear problems.
Result R3 (Stability of Process Systems): Suppose that
there exists positive 0 so that

December 2005

w Tw dx

and the result follows from Result R3 using

0 0


By rearranging the inequality in Result 4 we see that the

process system with homogeneous boundary conditions is stable if
k L 0


This expression shows that the process system is always stable

if kL is positive. Autocatalytic, coupled systems like the van der
Vusse system and exothermic reactions may give kL negative.
Vol. 51, No. 12

AIChE Journal

Wt W0


Figure 5. Trajectories of passive process systems with

different initial conditions, and the same
boundary conditions converge to a single trajectory in the mean square sense.

where W represents the stored electrical energy and product uy

represents the power supply. The inequality shows that the
electrical energy stored in circuit at time t is less or equal to
what initially stored plus the integrated power supply. The
difference has been dissipated as heat in the resistors. A linear
transfer function G(s) is passive if
1. G(s) is real for real s.
2. Re[G(s)] for all Re(s) 0.
We note that there is a close connection between passivity
and Lyapunov stability theory which can be expressed through
the Kalman-Yakubovich-Popov type stability theory. Let G(s)
have the controllable state space realization {C,A,B} with A
having all its eigenvalues in the closed left half plane. It then
follows that G(s) is passive if and only if there exists a positive
denite matrix P and vector d so that
PB C 0

In such cases it is possible to achieve stability in distributed

reaction diffusion systems by counterbalancing reaction by
diffusion and heat conduction if L is not too large (Alonso and
Ydstie, 2001).37 In order to develop spatiotemporal patterns in
reaction diffusion equations we, therefore, need to satisfy certain limiting size constraints as might be expected.
There are many ways to interpret deviation variables as
illustrated in Figure 5. One way is to view System 2 as a
stationary reference for System 1. Convergence results then
show that all trajectories converge to stationary behavior determined by the reference system. Another way is to view
deviations as spontaneous/stochastic uctuations that die out
after period of time.

The function V XT Px turns out to be a Lyapunov function for

the unforced system. More generally and assuming that W(t) is
differentiable we have the following denitions for nonlinear
Denition D1: Consider system S in Figure 1 with input u
and output y. Suppose that there exists a nonnegative C1
function W(t), so that
u Ty 22
The system S is

Passivity of Input Output Systems

It is quite easy to develop counter examples to control. For
example, the linear transfer function

s 3s 0.5
s 10s 0.5s 4


is impossible to stabilize because there is an unstable pole-zero

cancellation. The Bode plot of G(s) shown in Figure 6 reveals
one aspect of the input output behavior of the system however.
The plot shows that the phase shift of G(s) is never more than
90. Systems that display this kind of behavior are said to be
passive. Passivity provides a counterpart to small gain stability
theory. In the small gain theory we ignore phase and rely on
gain margin for stability. In passivity theory, we ignore the gain
and we use instead the phase margin for stability.
The classical example of a passive system is an electrical
circuit made op of resistors, capacitors, and inductors connected in parallel with resistors. In this case the input u in
Figure 1 represents the applied voltage and the output y represents the current. Independently of the complexity of the
circuit we can now write
AIChE Journal

December 2005

Figure 6. Bode diagram of the transfer function shown

in Eq. 22 generated by MATLAB.
Note the Bode stability criterion does not apply due to the
presence of the unstable pole.

Vol. 51, No. 12


1. passive if 0.
2. strictly input passive if u and 0.
3. strictly output passive if y and 0.
4. strictly state passive if represents the state and 0.
The notation 2 denotes the L2 norm for a function dened
on a domain so that for square integrable vectors we have


T d

In this article is dened by the semiclosed set [0, ) [0,

L], and subsets thereof.
Willems (1972b)30 proved that if W has a strong local
minimum then there is an intimate connection between dissipativity and Lyapunov stability. In this case, W is convex, and
it is also a Lyapunov function. Byrnes et al. (1991)31 developed
these ideas further and established conditions for stabilizability
of nonlinear, nite dimensional systems using passivity. It was
shown that a nite dimensional system is stabilizable if it is
feedback equivalent to a passive system. These results suggest
a close relation between passivity and the methods of irreversible thermodynamics.
Result R3 shows that process systems are strictly passive if
Assumptions A1-A3 are satised, and if L is sufciently small.
We can then write
u Ty 0

w t, x Tw t, xdx


where we made the assignments

f t, 0
f t, L

w t, 0
y w t, L

Thus, the input vector consists of ow deviations at the boundary and the measurement consists of intensive variable deviations at the boundary. Strict state passivity (in the space of
intensive variables) follows when we make the assignment w
. We note that passivity has a very nice intuitive interpretation which is evident from inequality (Eq. 23). The input and
the output in the average have the same sign. In other words,
if u is positive then y should be positive most of the time.
Conversely, if u is negative then y should be negative most of
the time. In the nonlinear case this translates into a sector
condition which plays a crucial role in nonlinear stability
One aspect concerning passive systems is that they are easy
to control. This can be observed by dening a proportional
u Ky,

Wt W0

Ky 2dt 0

w Tw dt

Thus, the stored energy never increases under negative output

feedback, and it is in fact decreasing if either of the vectors y
or w is not identically equal to zero for some nite period of
time. This gives convergence of y and w since W(t) cannot
decrease indenitely6.
The central result of passivity aimed at stability analysis of
interconnected systems is the passivity theorem. The following
result is most useful for out purposes.
Lemma L2 (Stability of feedback systems): Consider a passive system S and an input strictly passive controller C connected in feedback as shown in Figure 7. The respective
storage functions WS and WC have dissipation rates S 0 and
C 0. The interconnected feedback system with input d and
output e has L2 gain g 1/C.
Proof: The proof is instructive so we give it here for
completeness. First we note that we have
dW S
uT e
dW C
eT u CeT e
where u is dened in Figure 7. By adding the two storage
functions, dening W WS WC and setting u d u we
e Td CeT e
We now integrate over the interval [0,T] to give


Substituting this control law into expression Eq. 23 convinces

us that the storage function decreases according to

Figure 7. Feedback interconnection between a passive

system S, and a strictly passive controller C.

December 2005

Note that convergence is in the L2 sense, meaning that the error signals are square
integrable. This result can be strengthened to one of asymptotic stability of the signals
have uniformly bounded rst derivative. It turns out that it it quite easy to establish
that this is the case if the smallest eigenvalue of L is bounded away from zero.

Vol. 51, No. 12

AIChE Journal

Wt W0

e Td CeT edt

Since W is nonnegative we can re-arrange this so that

eT edt W0

eT ddt

Using the Schwartz-inequality we get

eT edt W0

dT ddt

eT edt

Figure 8. Inventory controller acts on transformed measurements and generates a set of synthetic
control outputs which need to be transformed
to real process inputs through control variable
synthesizers (CVs).



eT edt

eT edt

This structure is similar to the extensive variable controller

proposed by Georgakis (1986).4

dT ddt

is passive.
Proof: Let

By ignoring the effect of initial conditions we can write

e Tedt/

d Tddt


Inventory and Flow Control

Farschman et al. (1998)5 used the macroscopic balance
m, z, d pm, z, d


to derive control structures for chemical processes. We have

indicated that the ux and production variables depend on
vectors of manipulated variables m, state variables z, and
disturbances d. We furthermore note that (Eq. 24) is a positive
system since the state consists of elements that are nonnegative. All states of the system are stabilized (in the sense of
Lyapunov) if the total energy and total mass is bounded. We
now have the following result.
Result R4: Let v* denote a time-varying reference. The
synthetic input and output pair

AIChE Journal



That the vectors u and e are related in a passive manner is now

easily veried. By differentiating V we get

The result then is established by letting T 3 .

The passivity theorem, Lemma L2, shows that the gain from
the disturbance input d to the error e is nite and bounded by
the constant 1/C. Other results concerning the interconnection
of passive systems can be found in the books by Desoer and
Vidyasagar (1975) and Khalil (1996).28,55


v v* Tv v*

e v v*
December 2005

V e Te e Tv v * e Tu
Passivity follows.
The input output pair is called synthetic since it does not
necessarily correspond to what is directly measured and manipulated in the real process.
Denition D2 (Inventory Control): A control strategy
which ensures that an inventory asymptotically tracks a desired
set point is called inventory control.
It follows from Lemma L2 and Result R4 that inventory
control is input/output stable when we use feedback-feedforward control in the form
u m, z, d p m, z, d



In this expression denotes the estimate of the net ow, and p

denotes the estimate of the net production. Furthermore, let
and p p p . The operator C(e), which maps
errors into synthetic controls, should be strictly input passive.
We can then use Lemma L2 to show that the gain from the
error d p to set point errors is given by the inverse of
the L2 gain of the operator C. The structure of the control
system is shown in Figure 8. What we see here is that we
distinguish between two types of inputs and outputs to the
process system. One type of inputs and outputs concern the
physical ows. These ows connect with other process systems
Vol. 51, No. 12


is strictly input passive for any choice of Kc 0 and I , D

0. The dissipation rate is given by C Kc.
Proof: First dene the variables
s e
K c 2 K c D 2
where dot denotes time derivative. We now dene

u Kc e

s D e

Hence, by multiplying through with e

Figure 9. Process systems can be connected together

to form large networks.
Each node corresponds to a passive system and inventory
controllers may be used to control the node in a decentralized
fashion by acting on inputs u using measurements y. The
nodes themselves may consist of more devices connected
together as indicated.

eu Kc e2

m, z, d p z, d 0
has a unique inverse with respect to the manipulated variable
m. Farschman et al. (1998)5 refers to this property as
controllability. This condition fails close to maximum conversion point for component B in reaction scheme A 3 B 3 C. A
further discussion is provided by Ruszkowski at al (2005).27
There are many examples of control that at are strictly input
Result R5: The PID control

m, d p m, d

Kc e

edt D

es Kc D ee


We now differentiate the storage function to give


ss K c Dee


By using the fact that e s, and combining Eq. 27 and Eq. 28

we get
V eu Kc e2
It follows from Denition D1 that the PID controller is input
strictly passive with dissipation rate C Kc.
If I and D are equal to zero, then the closed-loop response
is given by the rst-order equation
et expKc te0
Thus, the closed loop time-constant is given by CL 1/Kc.
Setting Kc very large gives fast response whereas setting Kc
small gives slow response.
The method is not limited to PID control. For example, it is
often possible to write

m, z, d, t pm, z, d, t m, z, d, t T


where (m,z,d,t) is a vector which depends on observable

quantities, and is a vector of unknown parameters.
Result R6: The parameter adaptive controller
u m, z, d, t T


December 2005

eu K ce 2

and boundary conditions, and can be used to dene large and

very complex networks of process systems as indicated in
Figure 9. The beginnings of a theory for how such systems are
connected, and the stability can be analyzed is described by
Gilles (1998); Farschman et al. (1998), and Hangos et al.
(1999).56,5,57 Jillson and Ydstie (2005)51 recently developed a
process systems analog of Tellegens theorem which allows for
the very efcient analysis of complex thermodynamic networks. In such systems a very natural decomposition ensues
which allows one to develop distributed process control and
optimization methods using strong stability and optimality
results that have been developed for the analysis of electrical
circuits (Peusner, 1986; Ydstie, 2005).60,58
Inventory control is most useful for when the term p is
absent since the modeling requirements are minimal. This is
case for conserved properties like total mass, energy and components that do not undergo chemical transformation. Farschman et al. (1998)5 demonstrated that we get a unique
solution to the inventory control problem and stability provided
the equation

es D ee


gm, z, d, te
Vol. 51, No. 12

AIChE Journal

is strictly input passive for any choice of adaptive gain g 0

provided the feedback controller C is strictly input passive.
Proof: We rst note that from Eq. 29, and the denition of
u, we can write
u T t T t v *

Invariant Sets and Stabilizability

We, therefore, have by multiplying through with e

eu e T t e T t ev *
This we re-arrange to give

T te eu e T t ev *


We now dene the storage function


where Ci I 1,2,3,. . ...denotes a sequence of dynamic operak

tors so that i1
Ci is passive for all k. A recent exposition of
passivity theory, and its connections with nonlinear and H
robust control is given by van der Schaft (1996).32

t T t

In this section, we are interested in establishing the existence

of invariant sets when we apply inventory control to process
systems. A process system was dened by Ydstie and Alonso
(1997)36 to be system of conservation laws with a concave
extension which satises the laws of thermodynamics. In our
context this means that we can apply Assumptions A1 and A2
to the system of conservation laws (Eq. 14).
We have seen that application of ISP inventory control to a
process system guarantees according to the passivity theorem
that the inventories converge to their setpoints. We are now
interested in seeing what happens to the remainder of the
variables. We are in particular interested in seeing what happens if we control the total mass and energy of the system to
constant setpoints E* and M*. We then get
limM M* 0

By differentiating V we get


limE U* 0


V g 1 t T t


The total mass and energy are dened so that

By using the adaptive update law for (t) we have


V t T m, z, d, te

and E U K P


Using Eq. 30 we get

V eu e T t ev *
From the control expression we have (m, z, d, t) T (t)
v * C(e). It follows that we may write
V eu eCe
The input strict passivity property follows from Denition 1
since C is input strictly passive.
We can use nonlinear feedback. Sliding mode inventory
control is described by Wang and Ydstie (2004).61 As noted by
Shinskey (1967)3 it is often advantageous to use a nonlinear
feedback gain so that the controller takes the form

All these variables are bounded from below (by zero, with
exception of U which can be dened relative to an arbitrary
reference point). It follows that there exist invariant sets Z
so that limt3 Z Z. By continuity and concavity we conclude
that similar invariant sets are dened for the intensive variables. The sets are especially simple for the total mass and
energy since they are determined by the setpoints. It is well
known that the invariant sets corresponding to the remaining
variables may be quite complex, especially for the case of
chemical reactions.
We now have the following results
Lemma L3: The process system with input and output pair
dened so that

u f0 f L



m, z, t pm, z, t

y M M*

Kc v v*v v*

This gives the high gain for large errors and small gain for
small errors. Strict input passivity is easy to establish. Finally,
we can implement any kind of gain scheduling (hybrid) controller

is strictly state passive on the set * if there exist, and 0 0

so that

f TX w T dx f Tw 0L 0w Tw

C i e,
m, z, t pm, z, t
AIChE Journal

i 1, . . . , N.
December 2005


First dene the storage function

Vol. 51, No. 12


where W is dened by Eq. 21 and V by Eq. 25. We get, using
Corollary C1 and Result R4
f Tw 0L

Application to 1D TRPs

fTX w T dx

v v* T p


The result follows.

Lemma L3 gives sufcient conditions for strict state passivity on the set * for a process system with nonlinear reaction
and transport rates. In the limit this allows us to use the
classical contraction type analysis which is used in variational
calculus to establish tight bounds for stability using rst-order
conditions. This approach for stability analysis of nonlinear
systems can give tight bounds for nonlinear stability as long as
the trajectories are noncritical.51
Result R7: Suppose that we have following properties satised
1. L0 f T X w T dx f T w L0 0 w T w for some 0 0.
2. The total mass and energy is controlled using ISP inventory control.
Then there exist constants c0 and c1 so that we have

number and all intensive variables converge to stationary

points under inventory control. The theory can be extended to
study stability more complex and time-varying invariant sets
due to the introduction of the storage function R.

w Tw dt c 0


eT edt c1

In this section we illustrate the application of inventory

control to a TRP described by the equation
D 2
with t [0, ) and x (0, L), Hulburt boundary conditions
c0, t c 0t
cL, t
and undergoing a rst-order reaction of the form A 3 B with
reaction rate

where k is constant and c is the local concentration of chemical
specie A. In this example there is only one state variable.
Ydstie and Jiao (2004)26 and Ruszkowski et al. (2004)27 apply
the inventory control method to more complex examples.
Result R8: The transport-reaction system (Eq. 22) with
input and output dened so that

If in addition wi and y are uniformly continuous then it follows

that limt3 w 0 and limt3 y 0.
Proof: Follows directly from Lemma 3.
If set points and boundary conditions are constant and the
solutions to the steady-state equations
f i

u f0 f L

e M M*

are continuous then we can conclude that

x 0, L


w* denotes the solution to the steady-state problem, hence, we

limT1 T* 0

limP1 P* 0

limi1 *i 0




In this case the invariant set is very simple and corresponds to

a point in the space of intensive variables.
To summarize, the results in this section establish that chemical process systems are strictly passive and, hence, stabilizable
if the reaction rates are Lipshitz continuous L 0 and the
dimension of the system (L) is not too large. The stability
condition can be expressed in terms of a Dahmkohler type


and for all


is strictly passive if

i 1, . . . , n

limwx* wx, t 0

December 2005

k 0

Proof: Follows by application of Lemma L1 and Corollary

C2. Full developments are given by Ruszkowski (2003).27
We see that the system is stable since k is positive. The main
objective of the section is to look at a range of feedback control
structures that can be used to control the inventory.
In light of the motivation behind passivity theory for control
design, we now consider developing stable control structures
with passive input and output pairs as shown in Eq. 31. A
proportional feedback controller is the simplest type of controller which is strictly input passive. We get

f0 f L

dx Kc M M*


Kc 0

Vol. 51, No. 12

AIChE Journal

The control law is quite general in the sense that we are free to
choose manipulated variables including ow rate, inlet composition or the total reaction rate. In our particular case the inlet
ux f (0) admits the decomposition
f 0 vc0 v*c*0
We can choose to manipulate the transport velocity (v) or the
inlet concentration (c(0)). Regardless of the choice, for the
control law (Eq. 32) to be solvable, specic criteria must be
satised. The requirements are as follows:
1. The righthand side of control law (Eq. 32) must be well
dened and invertible with respect to the manipulated variable
so that the control action can be implemented.
2. The geometry of the system must be chosen so that the
system is strictly dissipative. This can always be achieved by
choosing L small.
The rst point ensures that a control action can actually be
calculated; this may set some limitations on selection of the
manipulated variables as we discuss a companion article (Ruszkowski et al., 2004).27 The second point ensures that all states
(controlled and uncontrolled) reach a stable stationary point.
In the following examples we do not include the term dv*/dt
in the control laws.
Example 1. In this example, we perform a servo control
study. A step change is induced to the inventory set point, and
the transport velocity (v) is chosen as the manipulated variable.
The control law (Eq. 32) takes the form
vc0 v*c*0 vcL v*c*L

*dx Kc M M*

Solving for velocity (v) we obtain ratio control

Kc (M M*)
c0 cL

dx b*

Figure 11. Initial (dashed) and nal (solid) concentration

proles for Example 1.

b* v*c*0 v*c*L


The control consists of proportional feedback (P-FB) term, a

feedforward (FF) signal from offset in the reaction rate and a
constant. This control law is well dened as long as c(0) is not
equal to c(L), which is the case if the reaction is present. If
there is no reaction present then the controllability is lost and
it is not possible to control the concentration of A using this or
any other method.
The response of the inventory will be given the rst-order
Mt M* e KctM0
Thus, 1/Kc gives the time constant of the closed-loop system.
Figure 10 shows the systems response to an inventory set point
change. The transport velocity converges to a stationary value
as the inventory (M) converges to its set point. The inlet
boundary condition is unchanged, but nonetheless a new concentration prole is achieved due to a change in space-time
velocity, as shown in Figure 11.
One disadvantage of the proportional control approach is the
necessity for a priori knowledge of steady-state offsets. One
way of overcoming this inconvenience is to drop the steady
state terms and use integral action. In this case, the new
proportional and integral action feedback-feedforward (PI-FB/
FF) control law becomes

where the constant b* is dened so that

Kc (M M*)
c0 cL

(M M*)dt


Figure 10. Servo control simulation results under proportional feedback and nonlinear feedforward control for Example 1.
The graph on the left shows the controlled output (Inventory
y MA). The graph on the right shows the manipulated
input (Velocity u v).

AIChE Journal

December 2005

The PI and PID controllers also represents a strictly passive

systems so the interconnection will be stable. The simulation is
now repeated without any a prior knowledge of the steady state
related terms. The response is comparable to previous results as
shown in Figure 12.
We now consider approximations that can be made while
maintaining passivity.
1. The term 1/(c(0)c(L)) is positive and can be ignored.
The response will be nonlinear instead of linear, but asymptotic
stability is maintained.
2. The term L0 dx can be omitted from the control law if
it is bounded. We can now use high-gain feedback like sliding
Vol. 51, No. 12


Figure 12. Servo control simulation results under proportional integral feedback and nonlinear
control for Example 1.

mode control. Passivity ensures that we converge to a sliding

surface characterized by zero error (Wang and Ydstie, 2004).61
In other words, the innite dimensional system is stabilized
PID type inventory control, and no knowledge of the process
Example 2: We now manipulate the inlet boundary condition (c(0)) instead of the transport velocity. Solving for this
quantity, the control law now becomes

c0 cL

Kc (M M*)

dx b*

where b* is constant as dened earlier. In this case we loose

controllability as the velocity decreases toward zero. This is
expected since we cannot control the concentration when there
is no ow. Figure 13 shows the response of the system under
this control strategy. The inlet concentration converges to a
stationary value as the mass inventory M exponentially converges to its set point.
In the examples we achieved the desired inventory levels
employing two different control approaches. Therefore, it is
possible and advantageous to choose the manipulated variable
that facilitates greatest controllability and stability. In practice,
however, the choice of a control strategy may be limited by the
availability of control variables and other equipment constraints.
Example 3: We perform a regulatory control study were
the transport velocity acts as a step disturbance. To maintain a
constant inventory set point, the inlet concentration (c(0)) is
manipulated. For this experiment, a proportional feedback/

Figure 13. Servo control simulation results under proportional feedback and nonlinear feedforward control for Example 2.
The graph on the left shows the controlled output (Inventory
y MA). The graph on the right shows the manipulated
input (Inlet concentration u c(0)).


December 2005

Figure 14. Disturbance rejection simulation results under proportional feedback and nonlinear
feedforward control for Example 3.
The graph on the top left shows the controlled output (Inventory y MA). The graph on the top right shows the
disturbance input (velocity d v). The graph on the bottom
left shows the manipulated input (inlet concentration u
c(0)). The graph on the bottom right shows the outlet concentration c(L).

feed-forward controller was tested. Figure 14 shows the results.

The concentration proles are shown in Figure 15. The prole
is tilted counterclockwise with respect to allow a velocity
change while maintaining the holdup of component A constant.
This gure shows that different internal conditions can give the
same macroscopic holdup.
In all three examples, whether the system was inuenced by
a proportional controller or a proportional integral controller, it
was able to converge to a stable stationary point.

In this article we have proposed a method to integrate
physics and process control by developing a thermodynamic
based approach to passivity based control of transport-reaction
systems. The class of systems we consider is quite broad and
encompasses a wide range of classical processes in chemical,
biochemical and metallurgical process engineering, as well as
control of small-scale systems, as long as the Boltzman principle holds and coarse graining can be applied. We provided a
review of the notion of passivity and related this to the second
law of thermodynamics. Using a local storage function con-

Figure 15. Initial (dashed) and nal (solid) concentration

proles for Example 3.
Note that the concentration prole adjusts so that molar
holdup of component A is kept constant.

Vol. 51, No. 12

AIChE Journal

sisting of microscopic and macroscopic elements, we derived

sufcient conditions for strict passivity and stability properties
by relating the storage function to thermodynamics using
Gibbs tangent plane method. This result also delineated general control structures for controlling convection-diffusionreaction systems. An array of feedback/feedforward control
structures were reviewed including PID, adaptive, nonlinear
and gain scheduling control. The theoretical results were illustrated with a numerical simulation of a 1-D convection-diffusion-reaction model. We showed the response for setpoint
changes for two types of control conguration. In one case we
controlled the inventory using ow rate as manipulated variable, in the other case we used the inlet concentration. In the
nal simulation we investigated the disturbance rejection property of the control system.

This research was generously supported by the National Science Foundation (CTS-9316572), Alcoa Inc. and ELKEM ASA. This article was
presented at the AIChE Annual Meeting in Austin TX, Nov. 2004.
In an interesting discussion in the poster session at the 2003 AIChE
meeting in Indianapolis, Professor Mike Doherty of UMass/UCSB pointed
out that stability conditions developed from the storage function A introduced by Alonso and Ydstie (199736), are going to be equivalent to those
using the Gibbs tangent plane condition.

Literature Cited
1. Christodes PD. Control of nonlinear distributed process systems:
Recent developments and challenges. AIChE J. 2001:514-518.
2. Buckley PS. Techniques of Process control. New York: John Wiley &
Sons, Inc; 1964.
3. Shinskey FG. Process Control Systems: Application, Design, Adjustment. New York: McGraw Hill; 1967.
4. Georgakis C. On the use of extensive variables in process dynamics
and control. Chem Eng Sci. 1986;41:1471-1484.
5. Farschman CA, Viswanath K, Ydstie BE. Process systems and inventory control. AIChE J. 1998;44:1841-1857.
6. Ray WH. Some Recent Application of distributed parameter systems
theory - A survey. Automatica. 1978;14:281-287.
7. Aling H, Benerjee S, Bangia A, Cole V, Ebert J, Emani-Naeini A,
Jensen K, Kevrekidis I, Shvartsman S. Nonlinear model reduction for
simulation and control of rapid thermal processing. Proceeding of
American Control Conference: Albuquerque, NM: 1997;2233-2238.
8. Balas MJ. Finite-dimensional control of distributed parameter systems
by Galerkin approximation of innite dimensional controllers. J of
Math Anal and Appls. 1986;114:17-36.
9. Christodes PD. Nonlinear and Robust Control of PDE Systems.
Berlin: Birkhauser; 2000.
10. Temam R. Innite-Dimensional Dynamical Systems in Mechanics and
Physics. New York: Springer-Verlag; 1988.
11. Brown HS, Jolly MS, Kevrikidis IG, Titi ES. Use of Approximate
Inertial Manifolds in Bifurcation Calculations. Continuation and Bifurcations: Numerical Techniques and Applications. (D. Roose, B. de
Dier, and A. Spence, eds.) Netherlands: Kluwer Acad. Publisher;
12. Smoller J. Shock Waves and Reaction Diffusion Equations. New York:
Springer Verlag; 1996.
13. Deane AE, Kevrikidis IG, Karniadakis GE, Orszag SA. Low dimensional models for complex geometry ows: Application to grooved
channels and circular cylinders. Phys Fuids A. 1991;3:2337-2354.
14. Christodes PD. Daoutidis P. Finite-dimensional control of parabolic
PDE systems using approximate inertial manifolds; J Math Anal Appl.
15. Kazantzis N, Good T. Invariant manifolds and the calculation of the
long-term asymptotic response of nonlinear processes using singular
PDEs. Comp and Chem Eng. 2003;29:999-1012.
16. Slotine JJ, Li S. Applied Nonlinear Control. Englewood Cliffs ;NJ:
Prentice Hall; 1992.

AIChE Journal

December 2005

17. Sontag ED. Mathematical Control Theory: Deterministic Finite Dimensional Systems. 2nd ed. New York: Springer; 1998.
18. Christodes PD, Daoutidis P. Robust Control of Hyperbolic PDE
Systems, Chem. Eng. Sci., 1998; 53:85-105.
19. Christodes PD, Robust control of parabolic PDE systems, Chem Eng
Sci. 1998;53:2449-2465.
20. El-Farra N, Armaou A, Christodes PD. Analysis and control of
parabolic PDE systems with input constraints. Automatica. 2003;39:
21. Kazantzis N, Kravaris C. Energy-predictive control: A new synthesis
approach for nonlinear process control. Chem Eng Sci. 1999;54:16971709.
22. Palazoglu A, Owens SE. Robustness Analysis of a xed-bed tubular
reactor: Impact of modeling decisions. Chem Eng Commun. 1987;59:
23. Hanczyc EM, Palazoglu A. Sliding mode control of nonlinear distributed parameter chemical processes. Ind. & Eng. Chem Res. 1995;34:
24. Sira-Ramirez H. Distributed sliding mode control in systems described
by quasilinear partial differential equations. System and Control Letts.
1989; 13:171-181.
25. Godasi S, Karakas A, Palazoglu A. Control of nonlinear distributed
parameter processes using symmetry groups and invariance conditions. Comp. & Chem. Eng. 2002;26:1023-1036.
26. Ydstie BE, Jiao J. Passivity based inventory and ow control in at
glass manufacture, Proc. IEEE-CDC, Bahamas; December 14-17,
27. Ruszkowski M, Read M, Kaiser RW, Richardson P, Kern T, Ydstie
BE. Passivity based control and optimization of a silicon process. Proc.
IFAC World Congress; (M. Simandl and P. Horacek, eds.) July 4-8,
2005, Praha, Czech Republic, paper TU-M06-TO/I.
28. Desoer CA, Vidyasagar M. Feedback Systems: Input-Output Properties. New York: Academic Press; 1975.
29. Willems JC. Dissipative dynamical systems, Part I: General theory.
Arch Rational Mech. Anal. 1972a;45:321-351.
30. Willems JC. Dissipative dynamical systems, Part II: Linear systems
with quadratic supply rates. Arch Rational Mech Anal. 1972b;45:352393.
31. Byrnes CI, Isidori AJ, Willems JC. Passivity, feedback equivalence,
and the global gtabilization of minimum phase nonlinear systems.
IEEE Transactions on Automatic Control. 1991;36:1228-1240.
32. van der Schaft A. L2-gain and passivity techniques in nonlinear
control. London U.K: Springer-Verlag Lecture Notes in Control and
Information Science; 1996:218.
33. Ortega R, Loria A, Nicklasson PJ, Sira-Ramirez H. Passivity based
control of EulerLagrange systems, Berlin: Communications and
Control Engineering. Springer-Verlag; 1998.
34. Ortega R. van der Schaft A, Mareels I, Maschke B. Putting energy
back in control. IEEE Control Systems Magazine; 2001;4:18-33.
35. Krstic M, Kanellakopoulos I, Kokotovic P. Nonlinear and adaptive
control design, New York: John Wiley & Sons, Inc; 1995.
36. Ydstie BE, Alonso AA. Process systems and passivity via the Clausius-Planck inequality. Systems and Control Letts. 1997;30:253-264.
37. Alonso AA, Ydstie BE. Stabilization of distributed systems using
irreversible thermodynamics. Automatica. 2001;37:1739-1755.
38. Ydstie BE. Passivity based control via the second law, Comp. and
Chem. Eng. 2002;26:1037-1048.
39. Coffey DP, Farschman CA, Ydstie BE A. Distillation Stability Using
Passivity and Thermodynamics. Comps Chem Eng. 2000;24:317-322
40. Ruelle D. Statistical Mechanics Rigorous Results, Amsterdam: WA.
Benjamin Inc; 1969.
41. Wang F, Landau DP. Determining the density of states for classical
statistical models: a random walk algorithm to produce a at histogram. Phys Rev E. 2001;64:1-16.
42. Shell SM, Debenedetti PG, Panagiotopoulos AZ. Generalization of the
Wang-Landau method for off lattice simulations. Phys Rev E. 2002:66.
43. Kestin J, Dorfman JR. A course in statistical mechanics, New York:
Academic Press; 1971.
44. Callen HB. Thermodynamics and an introduction to thermostatics. 2nd
ed. New York: JohnWiley & Sons Inc; 1985.
45. Sandler SI. Chemical and engineering thermodynamics. 3rd ed. New
York: JohnWiley & Sons Inc; 1999.
46. Coleman BD, Owen DR. A mathematical foundation for thermodynamics. Arch Rational Mechanics. 1974;54,1-104.

Vol. 51, No. 12


47. Hangos K, Cameron IT. Process Modeling and Model Analysis. New
York: Academic Press; 2001.
48. Kreuzer HJ. Nonequilibrium Thermodynamics and its Statistical Foundations. Oxford: Clarendon Press; 1981.
49. Astarita G. Thermodynamics: an advanced textbook for chemical
engineers. New York: Plenum Press; 1990.
50. Ottinger HC. Beyond equilibrium thermodynamics, Hoboken, NJ:
Wiley Interscience; 2005
51. Lohmiller W. JJ Slotine, On contraction analysis for nonlinear systems. Automatica. 1998;34:683-696.
52. Straughan B. The energy method, stability, and nonlinear convection.
Berlin: Springer-Verlag. Appl Math Sci Ser. 1992;91.
53. Struwe M. Variational methods. Applications to Nonlinear Partial
Differential Equations and Hamiltonian Systems. New York: Springer
Verlag; 1992.
54. Sontag ED. Mathematical Control Theory. Texts in Applied Mathematics 6. New York: Springer Verlag; 1990.
55. Khalil K. Nonlinear Systems. Englewood Cliffs, NJ: Prentice Hall
Professional Technical 1996.
56. Gilles ED. Network theory for chemical processes. Chem Eng Tech.
57. Hangos KM, Alonso AA, Perkins JD, Ydstie BE. Thermodynamic
approach to the structural stability of process plants. AIChE J. 1999;
58. Ydstie BE. Distributed decision making in complex networks: The
adaptive enterprise. Comp Chem Eng. 2005;24.
59. Jillson KR, Ydstie BE. Process networks. Paper to be presented in
IFAC World Congress: Czech Republic; Praha; 2005.
60. Peusner L. Studies in Network Thermodynamics. Amsterdam:
Elsevier; 1986.
61. Wang J, Ydstie BE. Robust inventory control of process systems.
Boston: American Control Conference; 2004.
62. Ruskowski M. Passivity Based Control of Transport-Reaction Systems: Application to Silicon Production. Chemical Engineering; Carnegie Mellon University; 2003. PhD.
63. Kondepudi DK, Prigogine I. Modern thermodynamics: From heat
engines to dissipative structures. New York : John Wiley & Sons, Inc;
64. Onsager L, Machlup S. Fluctuations and irreversible processes. Phys
Rev. 1953;91:1301-1308.
65. Gelfand IM, Fomin SV. Calculus of variations. Mineola; NY: Dover
Publications; 1963.
66. Glansdorff P, Prigogine I. Non-Equilibrium Stability Theory. Physica.
67. Wei J. Irreversible thermodynamics in engineering. Ind and Eng
Chem. 1966;58:55-60.

Appendix: Review of Classical Irreversible

Thermodynamics and Symmetry of the Linear
In this appendix we revisit two classical problems in
linear irreversible thermodynamics (LIT). The rst problem
concerns the issue of stability. The second problem concerns
the question of whether or not steady-state thermodynamics
can be formulated via a variational principle. Note that
while the constitutive laws for heat and mass transport, as
well as chemical reaction are assumed to be linear, this does
not imply linearity since we still allow for nonlinear phase
behavior as given by for example by the Peng-Robinson
equation of state.
Onsager and Prigogine were both awarded Nobel prizes for
contributing to LIT: Onsager for having established symmetry
of the conductance parameters using microscopic reversibility;
Prigogine for having established the principle of minimum
entropy production using Onsager symmetry as his point of
departure. The ideas are closely related to the principles of le
Chatelier-Braun and Maxwells theorem of minimum heat dis3164

December 2005

sipation. A very readable account of these ideas can be found

in the book by Kondepudi and Prigogine (1995).63
The entropy density is dened so that for any nonempty
volume V


We can now use HLE write

s s z

z t
By using the transport Eq. 14 we get


From which we get the conservation law for the entropy

w Tf

Entropy Flux


Entropy Generation

The second law of thermodynamics stated in the local form is

easily derived from Eq. A1. It states that the rate of entropy
generation is positive so that

S f TX w T 0


where the thermodynamic forces X are dened in Eq 18. By

combining Eq. A1 and inequality (Eq. A2), and integrating we
obtain the celebrated Clausius-Planck inequality which was
used as the starting point for the stability theory by Ydstie and
Alonso (1997).36
By using the ow decomposition (Eq. 15) we have from
Assumption A2
S f conv
X fdiff
X wT 0

Using the denition of the forces in Eq. 18 we have

f conv
X u

dx 1 /T
2 /T

By writing this expression out term by term and using the fact
we get
that u M

f conv

d 1
d P
d 1
dx T
dx T
dx T

Vol. 51, No. 12

d 2
dx T

AIChE Journal

Using the Gibbs-Duhem Eq. 3 we conclude that

f conv

Inequality (Eq. A2) can, therefore, be written

S f diffXT X wT 0


By using the expression Eq. 16 we have

w T

1 u 1 u

T x T x

T 2 T1 xu


The rst two terms represent reversible work and they cancel.
From Eq. A3 we, therefore, get

T 2 T1 xu 0

S f diffXT X


sufcient that 0 and that the eigenvalues of L and K are

nonnegative since T is positive7.
Result A1: Suppose that L has eigenvalues i 0 0.
The stationary state of a process system with linear laws (Eqs.
A5 and A6) and constant boundary conditions is stable.
Proof: This follows immediately from Result R3 since
Assumption A3 is satised.
Strict positivity of L implies that the resistance matrix R
L1 is bounded which it means that the medium has some
conductivity. Superconducting materials have the best stability
as might be expected, since gradients cannot be sustained at all.
We note that while the dissipation due to uid ow has been
included, the result does not say anything about uid ow
stability since the Lyaponov function(al) does not include the
We now want to address the issue of Onsager reciprocity
(symmetry of L) and see how symmetry is related to Prigogines principle of minimum entropy production. de Donder
(Kreuzer, 1981, Kondepudi and Prigogine, 1995)48,63 showed
that chemical reactions (not necessarily equilibrium reactions)
can be modeled using afnities. We review this theory very
briey here. Consider a system with n species and r chemical
reactions. We then have

We can, therefore, write the local form of the second law so


1 u
S f diffXT X TR m 2
T x


where ij is the stochiometric coefcient of species si as it
enters in reaction j. For example if we have n 3 and r 1,
and reaction

2A B 7 C

f diff LX


and for the reactive terms we get

R Km


where L and K are constant matrices. From expression (Eq.

A4) we can, therefore, write the entropy production as a quadratic form

1 u
T x

then we have 1 2, 1 2 and 3 1.

The afnity of reaction j is dened so that

v T



j 1, . . . , r



The afnity is equal to zero if reaction j is at equilibrium

(Sandler, 1996).45 de Donder, therefore, proposed to set the
reaction rate proportional to the afnity so that for reaction rj
kjAj. We can then combine Eqs. A8 and A9, and write the
expression for the reaction rate for component k so that

v k A v k v T


kj j


kj j



This we can write on the vector matrix form


In order to satisfy the second law for all possible variations in

forces and intensive variables it is, therefore, necessary and
AIChE Journal



The rst term in expression (Eq. A4) represents entropy generation due to heat conduction and diffusion. The second term
represents entropy generation due to chemical reaction, and the
third term represents entropy generation due to viscous dissipation.
LIT assumes that the constitutive equations for diffusive
ow and chemical reaction are linear. This gives

S X TLX mT Km 2

j 1, . . . , r

ji i


v s 0,

December 2005

Any square matrix, say B, can be decomposed into a symmetric matrix S and an
antisymmetric matrix A where A has zeros on the diagonal. We write this decomposition so that B A S. We therefore have xTBx xTSx for any vector x. It
follows that inequality (A7) doesnt imply symmetry.

Vol. 51, No. 12


kvv m

Using the steady-state equation A12 with A13 and A14, we get


j j j


d S

wT K
dx X


where vj (v 1j , . . . , v nj ) . We can now write


Using the symmetry of K and the steady-state Eq. A12 we get


j j j



Equation A10 now has the form given in Eq. A6. It follows that
K automatically is symmetric.
Lars Onsager (Onsager and Machlup, 1953),64 however,
argued that in LIT L must also be symmetric due to microscopic time reversibility (Kreuzer, 1984).48 This property
prompted Prigogine to propose the principle of minimum entropy production at steady state. He was able prove the generality of the principle using example problems. Later, we generalize the results due to Onsager and Prigogine, by showing
that in the linear case minimum entropy production and symmetry are equivalent.
Result A2: Consider an LIT system. The functional
L0 S (w, X, x, t)dx 0 is minimized at steady state if and
only if KL1 (L LT ) 0.
Proof: We have shown above that the rate of entropy
production is independent of the center of mass velocity since
the Gibbs-Duhem orthogonality allows us to write X T vz 0.
Without losing generality, we set u 0. At steady state we then


By differentiating (A7) with respect to X we have


d S

dx X


w TK KwT




December 2005

d S

w TKL 1 L LT
dx X


Using the assumption KL1 (L LT ) 0, we get

d S

dx X
However, these are the Euler-Lagrange equations for the functional and it follows that is at an extremum. Suppose now
that KL1 (L LT ) 0. Then there exists w so that the right
hand side of Eq. A15 is not equal to zero. The E L gives the
necessary and sufcient conditions for an extremum for this
problem (Gelfand and Fomin, 1963),65 and it follows, therefore, that the functional is not minimized for general w.
That the extremum indeed is a minimum follows from the
fact that the entropy production with linear laws is convex
according to constraint (Eq. A7).
We see that the condition of Result A2 is satised if L is
symmetric. It is also satised for nonreacting systems since we
then have K 0. Minimum entropy production, therefore, is a
more general principle than symmetry.
Prigogine developed a local stability result by using the
second-order variation of the entropy as a Lyapunov function
and looked at small deviations from equilibrium steady states
(Glansdorff and Prigogine, 1970; Kreuzer, 1981; Kondepudi
and Prigogine, 1998).66,48,63 This theory was quite successful
for diffusion, heat conduction and many other phenomena that
can be modeled using linear rate expressions.
The linear theory was not successful for the analysis of
systems far from equilibrium. The use of linear reaction is
especially troublesome and LIT, therefore, has limited scope
for practical application (Wei, 1966).67 Another important
problem with LIT is that it does not address input-output
behavior. In order to make use of the formalism for control it
is necessary, therefore, to develop theory that can (1) deal with
instabilities that arise due to nonlinear phenomena, and (2)
address the input output behavior as we have done is the main
body of this article.
Manuscript received July 30, 2004, and revision received Mar. 8, 2005.

Vol. 51, No. 12

AIChE Journal

Observer Design for Particle Size Distribution in

Emulsion Polymerization
David Edouard, Nida Sheibat-Othman, and Hassan Hammouri

Laboratoire dAutomatique et de Genie des Procedes (LAGEP), University Lyon I, Ba t 308,

69622 Villeurbanne Cedex, France
DOI 10.1002/aic.10559
Published online September 13, 2005 in Wiley InterScience (www.interscience.wiley.com).

Monitoring of the particle size distribution (PSD) in emulsion polymerization is a very

important task to improve the product quality. However, on-line sensors used to measure
PSD, such as capillary hydrodynamic fractionation, usually require taking samples and
the analysis time is still too long for control. This work proposes the combination of
mathematical models with existing on-line sensors to obtain a continuous estimation of the
PSD by constructing nonlinear estimators. Simulations of the observers show that they are
able to detect particle nucleation and to give an estimation of the whole PSD. 2005
American Institute of Chemical Engineers AIChE J, 51: 31673185, 2005

Keywords: emulsion polymerization, particle size distribution, observer, estimation, modeling, monitoring

Controlling the particle size distribution (PSD) is an important task in emulsion polymerization processing. The PSD is an
important parameter that affects the polymer rheology, adhesion, optical properties, mechanical strength, and the latex
stability and is an important parameter in control of process
productivity because it determines the reaction rate. The particle size is also an important parameter for lm formation (see,
for example, Geurts et al.1). In addition, by producing multimodal latices, Guyot et al.2 were able to maximize the latex
solids content and to reduce its viscosity.
One of the difculties in controlling the PSD lies in the lack
of on-line instruments that are able to measure it continuously
and on-line. Quasi-elastic light scattering (QELS) is usually
used to measure the particle size for monomodal latices. Online spectroscopic techniques such as near-infrared or Raman
spectroscopies have also been used to measure the particle size
in polymerization processes.3 However, with respect to lightscattering techniques, the QELS technique is more efcient for
monomodal latices. Multimodal latices can be measured by
off-line microscopic techniques, transmission electron microsCorrespondence concerning this article should be addressed to N. Sheibat-Othman
at nida.othman@lagep.univ-lyon1.fr.

2005 American Institute of Chemical Engineers

AIChE Journal

December 2005

copy (TEM), or scanning electron microscopy (SEM), all of

which require sample preparation. A disc centrifuge photosedimentometer is also used off-line to measure the PSD of multimodal latices with a wide range of diameters. For on-line
measuring the PSD of multimodal latices, separation techniques such as capillary hydrodynamic fractionation (CHDF)4
and ow-eld ow fractionation (f-FFF)5,6 give good performances. In CHDF, the particle size is obtained from the elution
time by calibration. The relative particle concentration is obtained by UV absorbance. In f-FFF, usually multiangle laser
light scattering is used as a detector, which allows absolute
determination of the particle size. In CHDF, small particles are
retained in the column for a longer time and thus larger particles exit rst, whereas in f-FFF smaller particles are the rst to
leave the column. Sample analysis takes about 12 min in
CHDF and a waiting period of 10 min is required between the
samples. In f-FFF, the separation phase takes 20 30 min.
Another factor retarding the development of PSD control lies
in the difculty in modeling and determining the kinetic parameters of the system. Gilbert7 proposed the 0 1 model and
identied the kinetics for styrene polymerization. This model
was then used by Coen et al.8 to model coagulation and
secondary nucleation. Using the same model, Sood9 used
CHDF and calorimetry to study the inuence of the monomer
addition protocol, emulsier, and initiator charges on the evolution of the PSD of a seeded styrene system.
Vol. 51, No. 12


Immanuel et al.10 used the pseudo-homopolymerization

model, initially proposed by Storti et al.,11 to study the PSD for
vinyl acetate and butyl acrylate polymerizaion and in a later
work4 to study particularly the effect of surfactant type and
coagulation on the PSD using CHDF and densimetry. This
model was also used by Ginsburger et al.12 to model and
identify styrene/butyl acrylate copolymerization.
The difculties in controlling the PSD also arise from the
fact that the PSD is represented by partial differential equations
that can be solved by different approximation methods (discretization methods, method of moments). Ramkrishna13 gave
a good presentation of the population balance equations and the
mathematical methods that can be used to solve them. Immanuel and Doyle14 presented an efcient solution technique for
population balance equations that is not based on approximating the population balance equations.
Finally, applying techniques of estimation and control to
these models requires theoretical validation. Semino and
Ray15,16 proposed a theoretical study of the controllability of
the PSD of continuous-emulsion polymerization reactors. The
theoretical controllability of the system was demonstrated using a linearized process model and validated by simulation to
the nonlinear model. The authors showed that the PSD is
controllable using the manipulating variables: surfactant, initiator and inhibitor concentrations, and assuming monomer conversion and total number of particles in the reactor to be
measured. However, the authors outline that the process is
easier to be controlled using the surfactant ow rate that acts
directly on the concentration of micelles. An experimental
study of the evolution of the PSD in continuous reactors was
done by Ohmura et al.17
Because of these difculties, only few papers in the literature
have treated controlling the whole PSD. Crowley et al.18 used
the 0 1 model to obtain a bimodal target distribution of styrene
polymerization. The authors compared both the surfactant feed
rate and free concentration as control variables that were calculated by an optimization algorithm. To control the PSD,
Zeaiter et al.19 used the 0 1 model for styrene polymerization.
The authors applied an optimal control strategy that calculates
on-line the desired monomer ow rate required to obtain the
desired PSD. Immanuel and Doyle20 proposed an open-loop
control strategy to control the PSD for vinyl acetate and butyl
acrylate polymerization. The generated algorithm was based on
the optimization of proles of the feed rate of monomer and
surfactant to obtain a target PSD. Later on, Immanuel and
Doyle21 controlled the PSD using a multiobjective optimization
algorithm by both open-loop and feedback algorithms. Doyle et
al.22 developed a hybrid model based approach to control the
PSD of styrene polymerization. The algorithm uses the partial
least-square model to optimize the feed rates of surfactant and
Using estimators in the control algorithm is interesting to
obtain information about the nonmeasured properties and to
overcome modeling errors. Actually, because the model is used
in a closed-loop manner by using the process outputs, this
allows detection of process mismodeling. In this study, we
propose estimators of the PSD and the nucleation rate. The
estimators are developed based on a new model derived from
the 0 1 model. After a presentation of the 0 1 model, the new
model is developed, validated, and compared to the original

December 2005

0 1 model. Then estimators of the PSD and nucleation rate are

constructed using this model.

Population balance equations
The general population balance equation is given by the
following hyperbolic differential equation:

, t

( , t)


generation , t


where (, t) is the population density function, generation is

the rate of generation of new members in the population, and
d/dt accounts for variation in the property with time.
In emulsion polymerization, the PSD can be described by
two types of models. Both models couple the population balance equations and the reaction kinetics in the aqueous phase
and in the polymer particles: initiation, propagation, terminations, transfer, nucleation, growth, and coagulation. The rst
model is the pseudobulk model that is available for large
particles in which more than one free radical can coexist for a
signicant period. To represent these kinetics, two differential
equations are necessary. The rst equation represents the evolution of the PSD for particles that have zero, one, or more
radicals and the second equation represents the number of
radicals in those particles as follows (for more details see
Immanuel et al.10):

n(r, t) [n(r, t)G(r)]

nuc (r r nuc) coagulation

n (r, t)
desorption 2


The particle density n(r, t)dr is dened as the number of

moles of particles of size between r and r r at time t; nuc
is the rate of nucleation; (r rnuc) is the Dirac delta function,
which is unity at r rnuc and zero elsewhere; G(r) is the

particle growth [G(r) dr/dt]; and

entry , desorption , and

termination are the total rates of entry, desorption, and termination within a single particle of size between r and r r,
The second model is the zero one model.7 This model is
adapted only for systems with the number of radicals per
particle equal to 0 or 1. Therefore, this model is generally
available for small particles, except for monomers that propagate rapidly, such as acrylates, where small particles exhibit
pseudobulk kinetics.23 This model distinguishes particles that
have a polymeric radical (n1p), particles that have no radicals
(n0), and particles that have a monomeric radical (n1m), as
shown in Figure 1 Nucleation produces particles with one
radical (n1p). A particle of type (n1p) becomes a particle of
category n0 if a new radical enters in or becomes a particle of
the category n1m if the radical it contains becomes a monomeric
radical by transfer to monomer or small chains. The distinction
Vol. 51, No. 12

AIChE Journal

between particles containing one radical and particles containing a monomeric radical is necessary in the 0 1 model because
desorption concerns only monomeric radicals.
Because the objective of this work is to monitor the PSD
starting from the beginning of the reaction, where the particles are small, the 0 1 model will be used throughout

this work to monitor the process. The particles were

assumed to be colloidally stable, and therefore coagulation
is not represented but can, however, be added to the
0 1 model. (For more details on the model see Crowley et

n 0(r, t)
(r, t)[n 1p(r, t) n 1m(r, t) n 0(r, t)] k des(r)n 1m(r, t)
[G(r)n 1p(r, t)]
n 1p(r, t)
entr(r, t)n 0(r, t) [ r, t) k tr[M] P]n 1p(r, t) k pe[M] Pn 1m(r, t)
nuc (r r nuc)
n 1m(r, t)
entm(r, t)n 0(r, t) k tr[M] Pn 1p(r, t) n 1m(r, t)[ (r, t) k pe[M] P k des(r)]

where entr(r, t) represents the rate of entry of a radical of

length i (z i jcrit1) into particles; entm(r, t) represents the
rate of entry of a monomeric radical into particles; and (r, t)
entr(r, t) entm(r, t) is the overall rate of radical entry into
The total number of particles of size between r and r dr
is given by
nr, t n 0r, t n 1pr, t n 1mr, t


Micellar (nucmic) and homogeneous (nuchom) nucleations

cause formation of very small particles containing a radical.
The total nucleation rate (nuc nuchom nucmic) thus
acts as a boundary condition of Eq. 3. This results in a Dirichlet
condition as follows:
n 1pr nuc, t

Gr nuc






[Micelle][IM i]V aq


nuchom(t) k

p,( jcrit1)

[M] aq[IM jcrit1]V




The average number of radicals in particles of size r at time

t [n (r, t)] is the ratio of particles containing either a radical or
a monomeric radical to the total number of particles, which
theoretically cannot exceed 1:
n r, t

n 1pr, t n 1mr, t
nr, t


Aqueous-phase reactions
Particle nucleation and radical entry/exit from the particles
depend on the concentration of micelles and radicals in the
aqueous phase. Therefore, to solve the population balance
equations, material balances in the aqueous phase should be
The polymerization process starts by the initiator decomposition (I) in the aqueous phase producing primary radicals (I)
that react with monomer molecules (M) to generate oligomeric
radicals (IMi). The main reactions taking place in the aqueous
phase are illustrated in Table 1.
Radical entry and desorption from particles is governed by
the coefcients (keE, ke,i, and kdes) that are assumed to be
diffusion controlled (Dw) and depend on the particle radius as
shown in Table 2. [E] is the concentration of monomeric
radicals that have a lipophilic nature and can therefore diffuse
easily into particles and desorb out of them. [T] is the total
concentration of radicals in the aqueous phase, dened by




where z is the critical chain length at which polymer radicals

can enter polymer particles or micelles (micellar nucleation)
and jcrit is the chain length at which the radicals become
insoluble in water and thus precipitate as new particles (homogeneous nucleation).
Writing the material balances of the species in the aqueous
phase requires the integration of the whole PSD:
1 dIV aq
aq k dI
V aq

Figure 1. 0 1 model.
AIChE Journal

December 2005

Vol. 51, No. 12


Table 1. Aqueous-Phase Reactions


Primary radical initiation

I 2I

Polymer radical initiation









inactive polymer chain

E inactive polymer chain


Radical transfer to monomer


IMi Micelle O
new particle z i jcrit 1

Micellar nucleation

crit 1

Homogeneous nucleation

new particle
IMj crit1 M O

Radical entry into particle

IMi particlen O



E particlen O

Entry of a monomeric radical into particles

1 dI V aq
2k dI k pI
M aq k taqTI
V aq


1 dIM iV aqt
k p,i1
M aqIM i1 k p,i
M aq k taqT
V aq
k traqM aq k em,iMicelleIM i

1 dIM 1V aq
V aq
k pI
M aqI k p,1
M aq k taqT k traqM aqIM 1


1 dIM iV aqt
k p,i1
M aqIM i1 k p,i
M aq k taqT
V aq
k traqM aqIM i, i 2, z 1

V aq


kp MP MWm
4r2 dp NA

k des

3Dw Msat
rs2 MPsat

k e,i



k eE
k em,i



k e,inrdrIM i

i z, j crit 1


1 dEV aqt

k desrn 1mrdr k traqM aqT



V aq

k eErEnrdr k taqET



4 r s NaD w

where Vaq is the volume of the aqueous phase, QI is the initiator

ow rate (mol/s), and [I] is the initiator concentration (mol/
It is important to note that species in the aqueous phase are
difcult to measure precisely. Therefore, determination of the
kinetics of a system is not easy. Moreover, these species are
usually affected by the inhibition phenomena. Therefore, attention should be made when using these equations in an
open-loop algorithm.

Monomer material balance and partitioning between the




Table 2. Description of Parameters


z i jcrit 1

z i jcrit 1

dm MPsat MWm


December 2005

The amount of residual monomer can be calculated either by

doing a balance on the monomer, taking into account the
reaction rate,
Vol. 51, No. 12

AIChE Journal

dN mt
Q m k p M P



or by doing a mass balance on the polymer ([M]P),

m p(t) d pN A

Surfactant material balances and micelle concentration

The surfactant is assumed to be introduced at a ow rate QS
(mol/s) that is smoothed by a rst-order model of reference:

r 3n(r)dr


maintain the polymer particles under saturation in all the simulations ([M]P [M Psat ]). Therefore, the model representing the
concentration of monomer in the polymer particles in interval
III is not necessary.


dN St
Q St

m P(t)
N m(t) N (t)
MW m

where Nm(t) and N m
(t) are the number of moles of residual
monomer and the total number of moles of monomer introduced to the reactor at time t, respectively; Qm is the monomer
ow rate (mol/s); mP is the polymer mass; and NA is Avogadros number. Both ways require the integration of the PSD
because , the number of moles of radicals in the polymer
particles, depends on the average number of radicals in the
polymer particles of any size at time t [n (t)] and the total
number of particles in the reactor (N PT ):

n t N TPt


where NS is the total number of moles of surfactant in the

The concentration of free surfactant in the aqueous phase
([Saq]) can be calculated from the following balance equation:

S aqV aq N S N SP N Sd

where N PS and N dS are the number of moles of surfactant

adsorbed on the surface of particles and droplets as given by
the following thermodynamic relationships:


nr, t N Adr



n t

S parb sS aq
a spN A1 b sS aq


S par 4 N A

nrr 2dr



n r, tnr, tdr



Micelle max 0,

During interval II (saturation of polymer particles), the concentration of monomer in the polymer particles {[M]Psat (mol/
dm3)} is calculated using the partitioning coefcients18:

M aq

and Vd is the droplets volume (dm3).

The concentration of micelles ([Micelle]) is then calculated
using the following expression:

nr, tdr


K daqMW m

where [M]sat
aq is the concentration of monomer in the aqueous
phase under saturation (mol/dm3):
M sat
P M aq K paq

[Saq ] CMC


where CMC is the critical micelle concentration and nagg is the

aggregation number for the surfactant. This signies that micelles exit only if the concentration of surfactant in the reactor
exceeds the CMC. Particle nucleation occurs at a minimum
size (rnuc) as a result of the presence of micelles that are formed
under sufcient concentration of surfactant. The rate coefcient of particle nucleation (kem,i) is diffusion controlled and
calculated using the radius of micelles (rmic) (Table 2).

Mathematical Treatment

Because the 0 1 model used during this entire studyis

adapted to small particles in the absence of gel effect (therefore, it is rarely available during interval III), we intended to
AIChE Journal


where Spar is the total particle surface:

and n (t) can be related to n (r, t) using Eq. 7, which gives the

3V d
a sdr dN A

N Sd

where N PT can be obtained by integrating the function representing the PSD [n(r, t)] over r:



December 2005

Experimental measurement of the PSD, by CHDF for example, gives the total number of particles of size r [n(r)] and does
not give n1p(r), n1m(r), and n0(r) individually. Therefore, constructing an estimator using this measurement would be easier
Vol. 51, No. 12


using a model that gives n(r) directly rather than using the 0 1
model that describes the three states individually. Therefore,
developing a model for small particles that have the same form
as the pseudobulk model would be of interest.
For this reason, the 0 1 model presented above is used in
this section to develop a new model that has the form of the
pseudobulk model.

By combining Eqs. 24 and 25 with the approximation given
by Eq. 26, the following model is obtained for n(r, t):
Grn r, tnr, t
nr, t

nuc r rnuc

Development of a bulk-like model for the 0 1 system

The new model is developed using equation system 3. The
change of variables to obtain the new model consists of the
equation of n (r, t) (Eq. 7) that can be written under the
following form:
Nr, t
n r, t
nr, t

Grn1p r, t n1m r, t
nr, t

nuc r rnuc

and the following equation can be obtained for n (r, t) by

deriving Eq. 24 with respect to time:
n r, t


Grn1p r, t
nr, t

nuc r rnuc


Grn r, tnr, t n (r, t) 1

k desrn 1mr, t

n(r, t)
nr, t


In this model, the particles n1m are assumed not to be

growing, that is,

Applying the quasi-steady-state assumption for [n1m(r, t)], as

suggested by Crowley et al.,18 the following algebraic equation
is obtained from Eq. 3:
n 1mr, t

Grn 1mr, t
This is explained by the fact that the lapse of time during which
a particle is of type n1m is very small. Actually a particle of type
n1m becomes very quickly either an n1p (if the monomeric
radical propagates), in which case it grows normally, or becomes an n0 (if the monomeric radical desorbs or terminates),
in which case it does not grow. For this reason, assuming that
the particles (n1m) grow as the particles of type n1p does not
affect the evolution of the PSD. In this case, Eq. 25 can be
written under the following form:

entmr, tn 0r, t k trM Pn 1pr, t

r, t k peM P k desr

Using the denitions of N(r, t), n(r, t), and Eq. 24, the
expression of n1m(r, t) can be rewritten under the following
n 1mr, t

entmr, tnr, t1 n r, t
r, t M Pk pe k tr k desr

k trM Pn r, tnr, t
r, t M Pk pe k tr k desr


Finally, the new model (bulk-like model) representing the

PSD of a 0 1 system is therefore the following:


n(r, t)
[G(r)n (r, t)n(r, t)]

nuc (r rnuc )


n (r, t)
[G(r)n (r, t)n(r, t)] n (r, t) 1
(r, t)[1 2n (r, t)]
n(r, t)

kdes r

nr, t
Nr, t
Nr, t
n r, t

n r, t
r, t1 2n r, t


nr, t

In a similar way, by combining Eqs. 3, 24, 26, and 28, the

following equation can be derived:

with N(r, t) n1p(r, t) n1m(r, t).

The sum of the three equations in the 0 1 model (Eq. 3)
gives a balance on the total number of particles in the reactor
[n(r, t)]:



ktr MP n r, t
entm r, tn r, t 1

r, t MP kpe ktr kdes r r, t MP kpe ktr kdes r

December 2005

Vol. 51, No. 12

AIChE Journal

Comparisons between the original and bulk-like models

for the 0 1 system

Numerical solution

The bulk-like model still describes the PSD of a 0 1 system.

The equation of n (r, t) contains the terms related to absorption/
desorption and transfer that are dispersed between the equations of the 0 1 model. To calculate n (r, t), the PSD has to be
known and integrated as in the 0 1 model.
The main reason that encouraged us to write the model under
this form is the facility of applying the observers, as will be
seen in the next section. This is because the state of the system
[n(r, t)] is the direct measurement and property we are interested in controlling. A second advantage of this form is that the
bulk-like model allows us to study easily the inuence of each
parameter on n (r, t) and therefore on n(r, t). For instance, if
desorption or transfer is neglected, then only the equation of
n (r, t) is modied and the equation of n(r, t) is maintained
without any change.
Finally, the bulk-like model simplies the transition between
small and large particles. Actually, to calculate the PSD of
small particles, the bulk-like model will be used and then, for
large particles, the pseudobulk model can be applied. This will
require changing the equation of n (r, t) only for large particles,
whereas the same equation of n(r, t) continues to represent the
PSD. When the 0 1 model is used to describe the PSD of small
particles, the transition to large particles represented by the
pseudobulk model that is completely different from the 0 1
model would be difcult.23
To validate the bulk-like model, it was compared to the 0 1
model by simulation under the same conditions. Two cases
seemed of interest to be studied: In the rst case, no desorption
is assumed [kdes(r) 0] and in the second case, the desorption
is calculated using the equation represented in Table 2.
Figures 25 represent the PSD and n (r, t) without and with
desorption, respectively. The gures show that the bulk-like
model gives very similar results as the 0 1 model. It can also
be seen that when no desorption takes place n (r, t) is equal to
1 for most particle sizes, which is the maximum allowable
value in this model. However, when desorption takes place,
n (r, t) has smaller values and increases with the particle size.

The population balance equations for the 0 1 system given

by Eq. 3 and the bulk-like model (Eq. 31) are solved by the
nite-difference method of order 1 applied on the derivative
with respect to r (particle radius). This approximates the integro-partial differential equations by ordinary differential equations with respect to time and discrete radius increment. An
equal increment of radius is used (r 0.4 nm). So, a large but
manageable number of coupled ordinary equations is obtained.
The nite-difference method is one of the discretization
techniques applied to solve the population balance equations
that is easy to use. However, approximating the derivatives
using this method might result in erroneous oscillations in the
distribution. In this work, no effort was made to improve the
approximation method. The nite-element method was found
to be sufcient to compare the two models and to test the
observers robustness. It remains yet important to develop
more robust techniques to solve the population balance equations under hard conditions.
The parameters used for simulation correspond to those of
styrene polymerization at 50C that could be found in the

Figure 3. Average number of radicals per particle at

1700 s (without desorption).

Figure 2. PSD at 1700 s (without desorption).

with the following boundary conditions:

nr nuc, t

Gr nuc


n rnuc , t 1

As in the 0 1 model or even in the pseudobulk model, n (r,

t) has no meaning if n(r, t) 0, that is, if there are no particles
of size r at time t. Actually, the equation that allows us to
calculate n (r, t) in the 0 1 model is singular if n(r, t) 0 (Eq.
7). This singularity is maintained in the new representation of
the 0 1 model derived above (bulk-like model). Therefore, a
test has to be done on the value of the PSD before calculating
n (r, t). If n(r, t) 0, then n (r, t) is not calculated, or taken to
be zero.

AIChE Journal

December 2005

Vol. 51, No. 12


assuming rst a known model of n (r, t) and then with errors in

the model of n (r, t), which is more usual.

Estimation of
Excitement of equation system 31 is attributed to particle
formation either by micellar or homogeneous nucleations
[nuc(t)]. The rate of nucleation is directly related to the
concentrations of monomer ([M]Psat ), radicals ([IM i ] for z i
jcrit 1) and micelles ([Micelle]) in the aqueous phase. However, the concentrations of radicals in the aqueous phase is
sensitive to impurities and inhibitors and the concentration of
micelles depends on a number of coefcients that are not well
known (adsorption and desorption). This makes the process
irreproducible and necessitates on-line measurement of the
reaction rate and the kinetics in the aqueous phase.
In this work, the monomer conversion
Xt 1

Figure 4. PSD at 1700 s (with desorption).

Figures 6 and 7 show the ow rate of surfactant used in the

simulations and the resultant concentration of micelles when
desorption is neglected or included in the model, respectively.
It can be seen that the same ow rate of surfactant does not give
the same evolution of the concentration of micelles in both
cases. Under desorption, the concentration of radicals in the
aqueous phase is higher and therefore the nucleation rate is
higher. The total number of particles produced is therefore
smaller without desorption than with desorption (see Figures 8
and 9).

Observer Design
A continuous stream information about the evolution of the
PSD is necessary for process control. However, the PSD is
difcult to measure on-line and the process model involves a
number of parameters that are sensitive to impurities and are
difcult to identify. This necessitates the construction of an
observer to monitor the PSD.
An observer combines the process model describing the
chemical reactions and physical behavior with the existing
on-line measurements to estimate the unknown parameters or
states of the system on-line. Among these algorithms, the
Kalman lter is often proposed in the literature. This lter
requires solving the dynamic Riccati equation, which is nonlinear. For a system of dimension k, the dimension of the space
of the Riccati equations is (k2 k)/2. Moreover, the choice of
the weighting matrices of the Riccati equations is difcult.
Therefore, the Kalaman lter becomes difcult to adjust and is
time consuming.
Our approach consists in designing an observer that derives
from the high-gain techniques (see, for instance, Bornard and
Hammouri24; Deza et al.25; Gauthier et al.26; Gauthier and
Kupka27; Farza et al.28; Hammouri et al.29; and Edouard et
al.30). The gain of the proposed observer does not require
solving any differential equation and its tuning is easier.
In the following, two estimators are constructed. The rst
one allows the estimation of the number of moles of radicals in
the particles () from the measured overall conversion. Then
this estimate is used in a second observer to estimate the PSD

December 2005

N mt
N m0

is assumed to be measured on-line. This can be done by calorimetry,31-33 on-line gas chromatography, or by spectroscopy.
Based on this measurement, Othman et al.34 developed an
estimator of the parameter . The estimation of allows
detection of inhibition and determination of the real value of
the reaction rate. The estimation is based on the material
balance of monomer given by Eq. 14. In this nonlinear equation, the sole unknown parameter is . This is similar to saying,
is observable from the model 14 if the residual number of
moles of monomer is measured. A high-gain nonlinear estimator could therefore be constructed to monitor :

dN m(t)
Q m k p (t)[M] P 2 (N m N m)
d (t)
(N N m)

k p[M] P m


where is a positive tuning parameter of the observer.

Figure 5. Average number of radicals per particle at

1700 s (with desorption).
Vol. 51, No. 12

AIChE Journal

Figure 6. Surfactant ow rate and concentration of micelles (without desorption).

The PSD and the average number of radicals per particle

will be estimated using the bulk-like model and based on the
estimate of . The parameter contains information about
the particles number and the average number of radicals per
particle. However, the PSD and the average number of
radicals per particle are not observable using the unique
information of .

To dissociate the inuence of the PSD and the average

number per particle on a model of n (r, t) will be used. If
this model is very well known, then n (r, t) can be simulated
in an open-loop manner and, in this case, n(r, t) becomes
observable because the unique unknown parameter. However, the known models of n (r, t) involve the kinetics in the
aqueous phase that are not always reproducible, and the

Figure 7. Surfactant ow rate and concentration of micelles (with desorption).

AIChE Journal

December 2005

Vol. 51, No. 12


Figure 8. Total number of particles (without desorption).

parameters of absorption and desorption that might contain

some irreproducibility. The model of n (r, t) must thus be
adjusted on-line. This requires the use of a new measurement that decouples n (r, t) and n(r, t). In this case, the

average number of particles is supposed to be measured

from time to time. This ensures the correction of the model
of n (r, t) and thus allows estimation of the PSD.
Both cases will be studied in the following sections:

Figure 9. Total number of particles (with desorption).


December 2005

Vol. 51, No. 12

AIChE Journal

First, the model of n (r, t) is considered known.

Second, uncertainties are considered in the model of
n (r, t).

where x(t) is the state of the system of dimension k 3, where

k corresponds to the size of the nite-element approximation:

n(r nuc k r, t)
x 1(t)

n(r nuc r, t)
x k(t)

x k1(t)
N pT(t)
x k2(t)
x k3(t)

Estimation of the PSD with a well-known model of

n (r, t)
To estimate the nucleation rate and the PSD, the bulk-like
model developed for the 0 1 system is used with the on-line
measurement of the monomer conversion and the estimator
of .
If is estimated on-line and if the model of n (r, t) is
assumed to be accurate, then the dynamic of N Tp (t) can be
estimated from the value of (t). Deriving Eq. 17 with respect
to time, and using Eq. 31 gives the following expression:
dN pTt


nr, tdr Gr nucn r nuc, tnr nuc, t


nuct Gr n r , tnr , t


Note that, the unknown dynamic of nuc(t) is represented by

(t), and A(r) is given by

ak ak1

ak1 ak2

Ar 0




However, if the domain of r is large enough such that the

particles size does not exceed it, then the following assumption can be made:

The ai expressions are given by the following equation:

Gr n r , tnr , t 0

The dynamic structure of Eq. 35 is under a canonical form of

observability, which allows us to construct a high-gain observer (see, for example, Farza et al.35) as given by the following system:

Numerically, if the nite-elements method is used to solve

the PSD, then this assumption can be ensured by using a high
nal particle size or by using a moving horizon of nite
elements that is always higher than the largest particles.
This gives
dN pTt


Equation 34 shows that the nucleation rate is observable

from the measurement of N Tp (t) that can be obtained from if
n (t) is obtained by open-loop simulation. Because the nucleation rate is the sole parameter affecting the PSD, then the PSD
can be reconstructed using the estimated value of nuc(t) and
Eq. 31a. This assumes that the kinetics of growth are well
identied. In the case of well-known kinetics of coagulation,
the estimation of the PSD based on the estimated rate of
nucleation would also be possible using the same observer.
Actually, the rate of coagulation is added to the rate of nucleation but with some dependency on the PSD. This would not
affect the observability of the system. The nucleation rate
remains observable from Eq. 33 where the known coagulation
term is added.
The global extended model needed to construct the observer
takes the following form:

x (t) A(r) x(t)

y(t) Cx(t) [0 0100] x(t)

(t) N A
N TP(t)
n (t)

AIChE Journal

December 2005

a i Gr nuc i rn r nuc i r, t/ r

x (t) A(r) x (t) [ x k1(t) y(t)]

(t) N A
n (t)





The positive parameter allows us to adjust the observer

convergence rate. Using a high value of allows the observer
to converge quickly to the set point but amplies measurement
noise. In our case, the best value obtained by simulation corresponds to 7. It can be seen that the rst k states,
representing the PSD, do not have a corrective term. In fact, the
observer allows us to lter N Tp (t) and gives an estimate of the
nucleation rate rnuc(t). The nucleation rate is then introduced in
Eq. 31 to calculate the entire PSD.
Figure 10 gives the surfactant feed rate used in the simulation.
The resulting free surfactant concentration resulting from the
addition of surfactant is presented in Figure 11. The estimation
results using equation system 35 combined with Eq. 31b are
shown in Figures 12 to 14. It can be seen that the estimator
converges well to the process model. Figure 12 shows that the
Vol. 51, No. 12


Figure 10. Feed rate of surfactant.

nucleation rate increases when the concentration of surfactant

exceeds the CMC, which results in two nucleation times. The
second time is saturated by the concentration of radicals present in
the aqueous phase. Figure 13 shows that the average number of
radicals per particle is 0.5 during homogeneous nucleation and
then decreases to 0.1 during micellar nucleation and is maintained
very small during the growth of particles. Finally, the PSD at
1500 s is shown in Figure 14. The estimator reconstructs the
estimation of the PSD very well.

The observer developed in this section thus allows us to estimate the PSD from the measurement of the monomer conversion
even under disturbances if the model of n (r, t) is supposed to be
accurate. However, the model of n (r, t) contains uncertainties
arising from inhibition and possible errors of parameter estimation. For this reason, the observer must be adapted in a way that
allows the estimation of the PSD even under uncertainties in the
model of n (r, t). Figures 15 and 16 show the inuence of n (r, t) on
the PSD. Actually, n (r, t) and therefore n (t) have a direct inuence

Figure 11. Concentration of free surfactant.


December 2005

Vol. 51, No. 12

AIChE Journal

Figure 12. Estimation of the nucleation rate.

on the nucleation rate and on the rate of polymerization by

inuencing the parameter (t) in Eq. 16. To illustrate this importance, simulation of the PSD was done with a constant value of
n (r, t). These results were then compared to the simulation using
the value of n (r, t) obtained from the equation model 31a. The two
following values were used in the simulations and are called
estimate (1) and estimate (2), whereas the real value obtained
from the model is called model:

Estimate (1)

n r



and thus n 0.2.

Figure 13. Estimation of n (t).

AIChE Journal

December 2005

Vol. 51, No. 12


Figure 14. Estimated PSD at 1500 s.

Figure 15 shows that n (r, t) substantially affects the nucleation rate and therefore the PSD. Figures 16a, 16b, and 16c
represent the real PSD and the PSD obtained by the constant
values of n (r, t), estimate (1) and estimate (2), respectively. This shows that in-depth knowledge of n (r, t) is necessary to obtain the real PSD. The open-loop simulation of the
model of n (r, t) that might undergo uncertainties is not ef-

Estimate (2)

n r

and thus n 1.

Figure 15. Inuence of the value of n (r, t) on the nucleation rate.


December 2005

Vol. 51, No. 12

AIChE Journal

Figure 16. Inuence of the value of n (r, t) on the PSD.

cient. In the following section, an observer is developed to take

into account any uncertainties in the model of n (r, t).

Estimation of the PSD assuming uncertainties on the

model of n (r, t)
To estimate n(r, t) and n (r, t) simultaneously, the observer
should be able to overcome inaccuracies in the model used for
the estimation, especially concerning the model of n (r, t) that is
sensitive to impurities.
If the model of n (r, t) undergoes any uncertainties, then it
cannot be used in an open-loop manner to estimate the PSD.
The unique measurement of the monomer conversion does not
permit one to update the model of n (r, t) and to estimate N Tp (t).
To take into account uncertainties in the model of n (r, t), a new
measurement has to be done to distinguish the inuence of n (t)
and N Tp (t) on the parameter (t).
This can be done either by measuring n (r, t) or n(r, t) at
discrete intervals, which allows one to update the model of n (r,

t). However, a discrete measurement of the PSD is not usually

available and requires sophisticated apparatus such as CHDF.
Nevertheless, to decouple the inuence of n (t) and N Tp (t) on the
parameter (t), measuring n (t) or N Tp (t) at discrete intervals
might be sufcient. This would allow to update the model of
n (r, t). Experimentally, the average size of particles can be
measured easily and quickly by light scattering, which allows
us to calculate N Tp (t).
The monitoring strategy thus consists of measuring the
monomer conversion continuously and measuring the average
particle size (therefore the total number of particles) at discrete
intervals. Based on the monomer conversion measurement, the
parameter is estimated using the observer presented by
equation system 32. The PSD observer (Eq. 37) is then applied
to estimate the PSD using a simplied model of n (r, t), which
is derived from Eq. 31a by maintaining the most important
terms numerically.
Let us consider Eq. 31a again:

Grn r, tnr, t n (r, t) 1

n r, t
r, t1 2n r, t
n(r, t)

k des(r)

entm(r, t)[n (r, t) 1]

k tr[M] Pn (r, t)

(r, t) [M] P(k pe k tr) k des(r) (r, t) [M] P(k pe k tr) k des(r)


Br, t

In this equation, A(r, t) and B(r, t) are functions of the

concentration of radicals in the aqueous phase (IM3 and IM4)
and the concentration of micelles ([Micelle]). To calculate
AIChE Journal

December 2005

these terms exactly, the desorption and adsorption phenomena

should be accounted for. These terms thus imply an important
calculation time. However, a rapid numerical analysis of these
Vol. 51, No. 12


Figure 17. Inuence of the parameter Dw on the parameters kdes, keE, and ke,3.

terms showed that the terms A(r, t) and B(r, t) have a slow
dynamic and do not change abruptly even under nucleation.
The second term is the most sensitive to nucleation and undergoes a rapid dynamic. This allows us to neglect changes
coming from the terms A(r, t) and B(r, t) if they are well

initialized. In this way, we ensure that we take into account any

nucleation that would guarantee having the good PSD by the
continuous observer presented in the previous section.
The following model of n (r, t) can therefore be derived from
Eq. 38:

Figure 18. Estimation of the PSD by the continuous-discrete observer.


December 2005

Vol. 51, No. 12

AIChE Journal

Figure 19. Estimation of the nucleation rate by the continuous-discrete observer.

n r, t Grn r, tnr, t n (r, t) 1

n(r, t)


Initialization of this equation becomes the crucial step of the

estimation. Using the discrete measurement of N Tp , n (t) is
estimated and the model of n (r, t) (Eq. 39) is reinitialized at this
estimated value.

The estimation results are presented in Figures 1719. Error in the

model of n (r, t) was mainly supposed to become from the rate of
diffusion (Dw) that is supposed to inuence the rate of absorption of
monomeric radicals (keE) and the rate of desorption (kdes). Figure 17
presents the inuence of the parameter Dw on the parameters kdes, keE,
and ke,3. It can be seen that Dw primarily inuences the parameter keE.
Figure 18 presents the estimation of the PSD using the

Figure 20. Estimation of the particles number by the continuous-discrete observer.

AIChE Journal

December 2005

Vol. 51, No. 12


continuous-discrete observer. Figure 18a represents the real

PSD (Dw 1.5 107). Figures 18b and 18c represent the
estimated PSD with an error on Dw (Dw 1.2 107 and Dw
1.8 107) with a discrete correction every 300 s. In the
second part of the gure, it can be seen that n (t) is well
estimated. An error of estimation can be seen on the estimation
of the nucleation rate, presented in the rst part of Figure 19.
However, it can be observed that both populations could be
detected. This is because the nucleation could be detected by
the growth term in Eq. 38. This is shown in Figure 19 representing the estimation of the nucleation rate and n (t). As in the
previous gures, Dw 1.5 107 represents the real process
and the other curves correspond to the observer with an error
on Dw. It can be seen that these curves tend to converge to the
real process at each correction (every 300 s) and then return
quickly to the normal model dynamic, although these corrections ensure the best detection of nucleation.
Corrections can also be seen in Figure 20, representing the
estimation of the number of particles by the continuous-discrete observer. At each step of correction, a jump of the
estimated curve of N PT is remarked, which allows us to stay
close to the real PSD. The rate of change in the number of
particles is detected in both cases (Dw 1.2 107 and Dw
1.8 107). The estimated number of particles increases at the
middle, although no nucleation is taking place. This arises from
the error in estimating the nucleation rate. However, the rate of
increase is smaller than that detected during nucleation.
It is worthy noting, however, that these corrections can be
ameliorated if the real PSD is measured from time to time,
rather than an average diameter value. Even without any supplementary measurements, this method can be used for feedback control of the PSD. Actually, nucleation could be detected
and therefore the entire PSD could be reconstructed based on
this estimate.

In this work, a new bulk-like model was developed to
represent the PSD in the 0 1 system. The new model is easier
to handle and makes possible the continuity between the 0 1
system (small particles) and the pseudobulk system (large
Based on this new model, a continuous-discrete observer
was developed to estimate the PSD and the average number of
radicals per particle simultaneously.
The observer is able to undergo modeling errors. An error
was supposed in the rate of desorption and absorption of
radicals, although the observer was able to detect nucleation
and to reconstruct the PSD.
This strategy should be a good tool for the control of the
PSD. Actually, the observer can be included in a control
scheme to manipulate the control variables: surfactant or initiator ow rates to obtain a target PSD.

asd surface area of droplets covered by a single surfactant molecule (4.2 1017), dm2
asp surface area of particles covered by a single surfactant molecule (4.2 1017), dm2
bs Langmuir adsorption isotherm parameter (related to the surfactant coverage area ratio) (2.1 103), dm3/mol
CMC critical micellar concentration (3 103), mol/dm3


December 2005

dm monomer density (0.878), kg/dm3

dp polymer density (1.044), kg/dm3
Dw diffusion coefcient for monomer in water (1.5 107),
[E] concentration of monomeric radicals, mol/dm3
G(r) particle growth, dm/s
[IMi] aqueous phase concentration of oligomeric radicals of degree i, mol/dm3
[I] concentration of primary initiated radicals, mol/dm3
jcrit chain length at which the radicals become insoluble in water
and principate (5)
Kdaq coefcient of monomer partitioning between monomer droplets and the aqueous phase (1966)
Kpaq coefcient of monomer partitioning between polymer partisat
cles and the aqueous phase ([M]sat
aq /[M]P 1348)
kd coefcient of initiator decomposition rate (7.4 107), s1
kem,i coefcient of the rate of radical entry into micelles, dm3
mol1 s1
ke,i coefcient of the rate of radical entry into particles of size r,
dm3 mol1 s1
keE coefcient of the rate of entry of monomeric radicals into
particles of size r, dm3 mol1 s1
kdes coefcient of the desorption rate of monomeric radicals, s1
kp coefcient of monomer propagation into particles (260), dm3
mol1 s1
k pI coefcient of initiation rate (100 kp), dm3 mol1 s1
k aq
p,1 coefcient of propagation of monomer with a one-unit radical (1200), dm3 mol1 s1
k aq

coefcient of propagation of monomer with a two-unit radp,2

ical (280), dm3 mol1 s1
k p,i
coefcient of propagation of monomer with an i-unit radical
(i 3, 4) (kp), dm3 mol1 s1
kpe coefcient of propagation of monomeric radicals (260), dm3
mol1 s1
k aq
coefcient of termination in the aqueous phase (1.16 109),
dm3 mol1 s1
ktr coefcient of transfer to monomer in the polymer particles
(9.3 103), dm3 mol1 s1
k tr
coefcient of transfer to monomer in the aqueous phase
(ktr), dm3 mol1 s1
[M]aq concentration of monomer in the aqueous phase, mol/dm3
aq concentration of monomer in the aqueous phase under saturation, mol/dm3
[M]P concentration of monomer in the polymer particles, mol/dm3
[M]Psat concentration of monomer in the polymer particles under
saturation, mol/dm3
[Micelle] concentration of micelles, mol/dm3
mp polymer mass, kg
MWm monomer molecular weight (0.104), kg/mol
N(r, t) total number of particles
NA Avogadros number, mol1
Nm number of moles of free monomer in the reactor, mol
NS total number of moles of surfactant in the reactor, mol
N PS number of moles of surfactant adsorbed on the surface of
particles, mol
N dS number of moles of surfactant adsorbed on the monomer
droplets, mol
N aq
S number of moles of surfactant in the aqueous phase, mol
n(r, t) number of moles of particles of size between r and r dr at
time t, mol
n0(r, t) number of moles of particles of size between r and r dr
containing no radicals at time t, mol
n1m(r, t) number of moles of particles of size between r and r dr
containing a monomeric radical at time t, mol
n1p(r, t) number of moles of particles of size between r and r dr
containing a polymeric radical at time t, mol
nagg micellar aggregation number (average number of surfactant
molecules in micelles) (60)
n (r, t) average number of radicals in particles of size r at time t
n (t) average number of radicals in the polymer particles of any
size at time t
QI ow rate of initiator, mol/s
Qm ow rate of monomer, mol/s
rmic radius of micelles (2.6 108), dm

Vol. 51, No. 12

AIChE Journal


nuc hom

(r, t)
entr(r, t)
entm(r, t)

nucleation radius (rmic), dm

radius of the monomer droplets (1 104), dm
swollen radius, dm
polymerization rate, mol/s
nucleation rate, mol/s
rate of homogeneous nucleation, mol/s
rate of micellar nucleation, mol/s
concentration of surfactant in the aqueous phase, mol/dm3
total particle surface, dm2
total concentration of radicals in the aqueous phase, mol/dm3
aqueous phase volume, dm3
volume of the monomer droplets, dm3
critical chain length at which polymer radicals can enter
polymer particles or micelles causing micellar nucleation (3)
number of moles of radicals in the polymer particles, mol
overall rate of radical entry into particles, mol/s
rate of entry of a radical of length i (z i jcrit1) into
particles, mol/s
rate of entry of a monomeric radical into particles, mol/s

Literature Cited




1. Geurts JM, Lammers M, German AL. The effect of bimodality of the

particle size distribution on lm formation of latices. Colloids Surf A
Physicochem Eng Aspects. 1996;108:295-303.
2. Guyot A, Chu F, Schneider M, Graillat C, McKenna TF. High solid
content latexes. Prog Polym Sci. 2002;27:1573-1615.
3. Santos AF, Lima EL, Pinto JC. In-line evaluation of average particle
size in styrene suspension polymerizations using near-infrared spectroscopy. J Appl Polym Sci. 1998;70:1737-1745.
4. Immanuel ChD, Crowley TJ, Meadows ES, Cordeiro CF, Doyle FJ III.
Evolution of multimodal particle size distribution in vinly acetate/
butyl acrylate emulsion copolymeriztions. J Polym Sci Part A: Polym
Chem. 2003;41:2332-2249.
5. Schneider M, Claverie J, Graillat C, McKenna TF. High solids content
emulsions. I. A study of the inuence of the particle size distribution
and polymer concentration on viscosity. J Appl Polym Sci. 2002;84:
6. Bartsch S, Kuliche WM, Fresen I, Moritz U. Seeded emulsion polymerization of styrene: Determination of particle size by ow eld-ow
fractionation coupled with multi-angle laser light scattering. Acta
Polym. 1999;50:373-380.
7. Gilbert RG. Emulsion Polymerization, A Mechanistic Approach. San
Diego, CA: Academic Press; 1995.
8. Coen EM, Gilbert RG, Morrison BR, Leube H, Peach S. Modeling
particle size distribution and secondary particle formation in emulsion
polymerization. Polymer. 1998;39:7099.
9. Sood A. Particle size distribution control in emulsion polymerization.
J Appl Polym Sci. 2004;92:2884-2902.
10. Immanuel ChD, Cordeiro CF, Sundaram SS, Meadows ES, Crowley
TJ, Doyle FJ III. Modeling of particle size distribution in emulsion
co-polymerization: Comparison with experimental data and parametric
sensitivity studies. Comput Chem Eng. 2002;26:1133-1152.
11. Storti G, Carra S, Morbidelli M. Kinetics of multi monomer emulsion
polymerization: The pseudo-homopolymerization approach. J Appl
Polym Sci. 1989;37:2443-2467.
12. Ginsburger E, Pla F, Fonteix Ch. Modeling and simulation of batch

AIChE Journal


December 2005


and semi-batch emulsion copolymerization of styrene and butyl acrylate. Chem Eng Sci. 2003;58:4493-4514.
Ramkrishna D. Population Balances: Theory and Applications to
Particulate Systems in Engineering. San Diego, CA: Academic Press;
Immanuel ChD, Doyle FJ III. Computationally efcient solution of
population balance models incorporating nucleation growth and coagulation: Application to emulsion polymerization. Chem Eng Sci. 2003;
Semino D, Ray WH. Control of systems described by population
balance equationsI. Controllability analysis. Chem Eng Sci. 1995;
Semino D, Ray WH. Control of systems described by population
balance equationsII. Emulsion polymerization with constrained control action. Chem Eng Sci. 1995;50:1825-1839.
Ohmura N, Kataoka K, Watanabe Sh, Okubo M. Controlling particle
size by self-sustained oscillations in continuous emulsion polymerization of vinyl acetate. Chem Eng Sci. 1998;53:2135.
Crowley T, Meadows ES, Kostoulas E, Doyle FJ III. Control of
particle size distribution described by a population balance model of
semibatch emulsion polymerization. J Process Control. 2000;10:419432.
Zeaiter J, Romagnoli JA, Barton GW, Gomes VG, Hawkett BS,
Gilbert RG. Operation of semi-batch emulsion polymerization reactors: Modeling, validation and effect of operating conditions. Chem
Eng Sci. 2002;57:2955-2969.
Immanuel ChD, Doyle FJ III. Open-loop control of particle size
distribution in semi-batch emulsion copolymerization using a genetic
algorithm. Chem Eng Sci. 2002;57:4415-4427.
Immanuel ChD, Doyle FJ III. Hierarchical multiobjective strategy for
particle-size distribution control. AIChE J. 2003;49:2383-2399.
Doyle FJ III, Harrison ChA, Crowley TJ. Hybrid model-based approach to batch-to-batch control of particle size distribution in emulsion polymerization. Comput Chem Eng. 2003;27:1153-1163.
Coen EM, Peach S, Morrison BR, Gilbert RG. First-principles calculation of particle formation in emulsion polymerization: Pseudo-bulk
systems. Polymer. 2004;45:3595-3608.
Bornard G, Hammouri H. A high gain observer for a class of uniformly
observable systems. Proc of the 30th IEEE Conference on Decision
and Control, Brighton, UK; 1991:1494.
Deza F, Busvelle E, Gauthier JP. High gain estimation for nonlinear
systems. Syst Control Lett. 1992;18:295.
Gauthier JP, Hammouri H, Othman S. A simple observer for nonlinear
systems: Application to bioreactors. IEEE Trans Autom Control. 1992;
Gauthier JP, Kupka IAK. Observability and observers for nonlinear
systems. SIAM J Control Optim. 1994;32:975.
Hammouri M, Farza H, Busavon K. A simple observer for a class of
nonlinear systems. Appl Mater Lett. 1998;11:27.
Hammouri H, Targui B, Armanet F. High gain observer based on a
triangular structure. Int J Robust Nonlinear Control. 2002;12:497.
Edouard D, Schweich D, Hammouri H. Observer design for reverse
ow reactor. AIChE J. 2004;50:2155-2166.
Arzamendi G, Asua J. Copolymer composition control of emulsion
copolymer in reactors with limited capacity for heat removal. Ind Eng
Chem. 1991;30:1342-1350.

Manuscript received Jan. 5, 2005, and revision received Mar. 22, 2005.

Vol. 51, No. 12



Process Intensication in a Trickle-Bed Reactor:

Experimental Studies
S. K. Dhiman, Vishal Verma, D. P. Rao, and M. S. Rao
Dept. of Chemical Engineering, Indian Institute of Technology, Kanpur 208 016, India
DOI 10.1002/aic.10560
Published online August 11, 2005 in Wiley InterScience (www.interscience.wiley.com).

Process intensication has been the focus of increasing attention in recent years. A
volume reduction of distillation and absorption units by a factor of 100 or more using
rotating beds has been claimed. Industrial trickle-bed reactors are bulky because the
liquid ow and interphase transport of the reacting species are governed by the earths
gravity, as in the case of distillation columns. To explore the possible process intensication in these reactors, we have measured the reaction rates of hydrogenation of
-methyl styrene with palladium as catalyst in the rotating beds of spherical particles and
metal foam, which acted as catalyst support. To quantify the intensication achieved, the
reaction rates were compared with those of conventional trickle beds. The enhancement
in the reaction rates was in the range of 30 40 times in a centrifugal force eld of about
450 times the gravitational force eld. An industrial reactor of 60 m3 could be replaced
with a rotating bed 1.5 m3 in volume. It appears that there is a possibility of further
process intensication. However, the volume reduction is possible only for mass-transfer
limited reactions. 2005 American Institute of Chemical Engineers AIChE J, 51: 3186 3192,

Keywords: process intensication, Higee (high gravity), trickle-bed reactor, rotating

packed bed, rotating trickle-bed reactor

Process intensication in distillation and absorption equipment has been the focus of increasing attention in recent years.
In conventional columns, the achievable vapor and liquid
throughputs and interphase mass transport are inherently governed by the gravitational eld. These can be enhanced in a
rotating annular packed bed [also known as Higee (high gravity)] because the liquid ows under the inuence of a centrifugal force eld of 350 1000 times the gravitational force eld.
It has been claimed that such process intensication can lead to
a volume reduction of the equipment by a factor of 100 to
1000. This in turn leads to lower capital costs and inherently
safer design because the liquid holdup is smaller than that in
conventional equipment. Ramshaw,1 at Imperial Chemical InCorrespondence concerning this article should be addressed to D. P. Rao at

2005 American Institute of Chemical Engineers


December 2005

dustries (ICI), UK and recently Trent et al.,2 at Dow Chemicals,

reported the commercial use of rotating packed beds. An appraisal of the process intensication in the rotating packed beds
has been presented by Rao et al.3 The centrifugal eld permits
the use of a packing with high specic area in the range of
1000 4000 m2/m3, which is about ve to ten times larger than
the conventional packing. The liquid ows as thin lms because of the centrifugal force and this leads to a reduction in the
liquid-side resistance to mass transfer. Sivalingam et al.4 carried out studies on mass transfer from gas to solid through the
liquid lm owing over a rotating string of spheres. They
found that the enhancement in the transfer rate is ve to seven
times on the rotating string of spheres compared to that with the
gravity ow. Therefore, the enhancement in the volumetric
liquid-side transfer coefcient could be as high as 2570 in
rotating packed beds of spheres. Therefore, for the reaction
systems in which the rate-limiting step is the liquid-side mass
transfer, a similar volume reduction of the reactor could be
achieved. Rao et al.5 proposed a novel rotating packed bed in
Vol. 51, No. 12

AIChE Journal

Figure 2. Arrangement of inert catalyst layers and

metal-foam (top view).
Radial width of layers (cm): (a) outer 2.5, active 1.5, inner
1.0, innermost 1.0; (b) outer 3.1, metal-foam 2.2, inner 0.7.

dium. To quantify the process intensication, we have compared the rates in rotating trickle-bed reactors with those in
conventional trickle-bed reactors reported in the literature.
Figure 1. Model trickle-bed reactor.
1: Manometer tap; 2: gas outlet; 3: liquid inlet; 4: motor; 5:
gas inlet; 6: glass wall; 7: annular packed bed; 8: alumina
catalyst beads; 9: stand; 10: liquid outlet; 11: distributor; 12:
AC inverter drive; 13: cylindrical side plate with rectangular
slots; 14: cup for collecting splash liquid; 15: elliptical dish;
16: splashed liquid outlet; 17: cylindrical bafe.

which the slip velocity between gas and liquid is as high as

10 30 m/s in contrast to 13 m/s in the conventional columns.
A similar volume reduction can be achieved if the rate-limiting
step is the gas-side mass transfer.
Trickle-bed reactors are used for hydrodesulfurization, hydrodemetallization, hydrodenitrication, hydrogenation, and
oxidation. These units are bulky because the ow of the liquid
in these reactors and interphase transport of reactants are inherently governed by the gravitational eld, as in the case of
distillation and absorption columns. The volumes of some of
these reactors are in the range of 150 200 m3.6 Van Hasselt et
al.7 showed that the volume of a trickle-bed reactor would be in
the range of 430 to 600 m3 for a throughput of 20,000 barrels
per day for deep hydrodesulfurization of vacuum gas oil
required to meet future emission standards. Further, they suggest the use of special packing and countercurrent ow of the
gas and liquid to reduce the volume. The rotating packed beds
permit countercurrent ow, or even crosscurrent ow, which
may be superior to the former in that the reverse reaction can
be reduced because the H2S is removed as it is formed. They
may nd application for the FisherTropsch synthesis of wax,
liquid-phase hydrodesulfurization of kerosene and naphtha,
and hydrogenation of aromatics and olens in solvent manufacture. The rotating beds are not new; the rotating-basket
reactors have been used for the determination of intrinsic
The objective of the present study is to explore the possible
process intensication in rotating trickle-bed reactors. Toward
this end, we have chosen the hydrogenation of -methyl styrene to cumene using palladium as catalyst, which is known to
be a mass-transferlimited reaction above 40C. We have
carried out the experimental studies on a rotating trickle-bed
reactor using a bed of spherical alumina particles impregnated
with palladium as well as with metal foams plated with pallaAIChE Journal

December 2005

Rotating trickle bed
Figure 1 shows the experimental rotating trickle-bed reactor.
It consisted of an outer casing, housing for the bed, a liquid
distributor, and a cup for collecting the liquid that was splashed
at the inner periphery of the bed housing. The side of the casing
was a cylindrical glass tube sandwiched between two metal
plates. To facilitate the liquid withdrawal, a stainless steel
elliptical dish was attached to the lower end of the glass wall as
shown in the gure. The housing was attached to the shaft of
the motor mounted on the top plate. The rotational speed of this
induction motor was varied continuously between 500 and
2500 rpm using an inverter drive.
The housing of the bed was made of two annular disks
separated by two inner and outer slotted cylindrical shells and
were bolted to provide a rigid housing for the bed. The dimensions of the housing were: height, 4.5 cm; inner diameter, 9 cm;
and outer diameter, 21 cm. The free area for the ow of gas and
liquid on the inner and outer plates was 50% of the total.
To ensure uniform ow of liquid over the active catalyst, the
bed of catalyst particles was sandwiched between the layers of
inert particles as shown in Figure 2a. Similarly, the metal-foam
rings were sandwiched between the layers of inert alumina
particles as shown in Figure 2b.

Figure 3. Liquid distributor and bed assembly.

Vol. 51, No. 12


Table 1. Physical Properties of -Alumina Particles

Mean diameter, mm
Packing density
External porosity
Internal porosity
Particle density
Pore volume
External surface area

2.2, 1.5
797.6 kg/m3
1.17 g/cm3
0.5366 cm3/g
1690, 2489 m2/m3

The liquid distributor and the bed assembly are shown in

Figure 3. The liquid distributor was made of a disk-shaped
header. The liquid inlet tube was attached to the header at the
center. Four tubes were attached near the outer periphery as
shown in Figure 3. These were provided with eight equispaced
circular holes of 0.3 mm diameter. The spacing of holes on the
four tubes was staggered such that the liquid issued as 32 jets
directed toward the inner periphery of the housing. A clearance
of 0.5 cm between the tubes and the inner periphery was kept
to ensure that the jets reached the latter almost horizontally. At
very low liquid ow rates, some of the alternate holes were
blocked by winding a Teon tape over the holes to achieve
such a condition.
If the bottom plate of the housing was a circular disk, a part
of the liquid reaching the inner periphery gets splashed and
falls onto the bottom disk. It ows on the disk as a thick layer
without proper contact with the gas phase.9 To prevent the
bypassing of liquid, an annular disk and a cup arrangement
were provided to draw the splashed liquid separately and to
measure its ow rate. A cylindrical bafe was tted to the
bottom disk to prevent the ow of splashed liquid into the
casing. The clearance between the bafe and the cup was large
enough to allow free ow of hydrogen from the casing to the
inner periphery of the bed.

Catalyst preparation
The alumina catalyst particles were prepared by impregnating them with palladium, as suggested by Herskowitz et al.10
The visual observations indicated that the palladium was uniformly distributed throughout the particles. We have used

particles of 2.2 and 1.5 mm in diameter. The physical characteristics of particles supplied by the vendor (Norton Co., Philadelphia, PA) are given in Table 1.
The metal-foam rings were supplied by Recemat International B.V. (The Netherlands). Their structure and characteristics as supplied by the vendor are given in Table 2. The
metal-foam bed was made by stacking ve annular rings: four
of 1 cm height and one of 0.5 cm height. To deposit palladium
on these rings, we followed the electroless-deposition method
suggested by Lowenheim11 and Henry.12 First, the metal-foam
rings were given an alkaline soak for 30 min in a boiling
solution with a composition of 15 g/L NaOH and 15 g/L
Na2CO3. After soaking, the rings were thoroughly rinsed with
distilled water. These rings were then immersed in a plating
bath containing a solution of 2 g/L PdCl2, 160 mL/L NH4OH
(27%), 26 g/L NH4Cl, and 10 g/L NaH2PO4H2O at 50C. The
solution was stirred with a magnetic stirrer to eliminate gas
bubbles and to maintain uniform composition in the holes and
crevices. Upon completion of the electroless deposition, the
solution, which was dark in color, became clear (colorless).
After the deposition, the foam rings were rinsed thoroughly
with distilled water and dried in a stream of hot air. The
difference in the weight of metal rings before and after the
plating gave the exact amount of palladium deposited. The
visual observation of the metal foam indicated that it was black
in color and the palladium was uniformly deposited. The catalyst particles were activated by heating to 370C in a stream
of hydrogen for 24 h. However, the metal foams were used as
such after washing with distilled water.

The experimental setup is shown in Figure 4. Keeping the
bed rotating, rst nitrogen and then hydrogen were passed
through the casing for about 1 h each to ensure the casing was
lled with hydrogen before the start of the run. Then, the liquid
was circulated through a water bath and the bed using a
peristaltic pump (ColeParmer Instrument Co., Vernon Hills,
IL). A surge tank was used to dampen the uctuations in the
ow. The rotational speed and the liquid ow issuing out of the

Table 2. Properties of RECEMAT MetalFoam


RECEMAT (2733)

RECEMAT (1723)

Estimated avg. pore diameter, mm

Specic surface area, m2/m3
Dimensions of the rings
ID, mm
OD, mm
Density, g/cm3





Metal foams


December 2005

Vol. 51, No. 12

AIChE Journal

Figure 4. Experimental setup.

1: Reactor; 2: peristaltic pump; 3: A/C motor; 4: feed tank-cum-presaturator; 5: surge tank; 6: manometer tapping; 7: digital thermometer; 8:
gas ow control valves; 9: liquid inlet; 10: gas outlets; 11: liquid outlet; 12: H2 gas cylinder; 13: N2 gas cylinder; 14: gas inlet; 15: sampling
port; 16: magnetic stirrer; 17: cylindrical bafe; 18: water bath; 19: rotameter; 20: water bath.

bed were monitored using a digital stroboscope (Nova-Strobe

DB plus, ColeParmer) to observe the nature of the ow. We
carried out studies at 2200 and 1900 rpm and have reported the
centrifugal acceleration, computed at the inner periphery of the
catalyst layer. The liquid came out of the bed as ne droplets
and they appeared to be uniformly distributed throughout the
outer periphery. No special provision was made to circulate the
hydrogen through the bed. Because of the centrifugal force
eld the gas is drawn in through the inner periphery and forced
out through the outer periphery.
Liquid samples were drawn through the sampling port at
regular intervals and analyzed for the cumene concentration. A
gas chromatograph (Model 5765; Nucon Engineers, Delhi,
India) with a 2 m long bentone-34 column and a dual-ame
ionization detector was used for the analysis. A stream of
hydrogen was passed through the coil immersed in a water bath
before it entered the casing and the excess hydrogen was
vented from the reactor. In all the runs, the -methyl styrene
was presaturated with hydrogen. The liquid was circulated
through the bed with a peristaltic pump (Model 78022-10,
ColeParmer). The recirculation rate and the splashed liquid
ow rate were monitored. The rotational speed was measured
with the digital stroboscope and cross-checked with the inverter drive readings. All the experiments were carried out at a
temperature of 40C and 1 atm pressure.

V dC
V b dt

where V is volume of the liquid charge to the setup (cm3), Vb

is volume of the active catalyst bed (cm3), C is the concentration of cumene (mol/cm3), and t is time (s).
Figure 5 shows the reaction rates with the alumina particles
in the rotating beds. The liquid supercial velocity was computed based on the inner radius of the active layer. The rate is
higher for the smaller size particles. Under the conditions used,

Figure 5. Reaction rates with -alumina particles.

Results and Discussion

The reaction rate was determined from the rate of change of
the concentration of cumene in the charge with time using the
AIChE Journal

December 2005

, 1.5-mm beads: 485 g (2200 rpm), 1.0 mass % Pd; f,

1.5-mm beads: 374 g (1900 rpm), 1.0 mass % Pd; , 2.2-mm
beads: 483 g, 1.0 mass % Pd; F, 2.2-mm beads: 374 g, 1.0
mass % Pd.

Vol. 51, No. 12


Figure 6. Reaction rates with metal foams at 360 g (1900

F, RECEMAT (1723): 0.5 mass % Pd; f, RECEMAT
(1723): 1.0 mass % Pd; , RECEMAT (1723): 1.5 mass %
Pd; , RECEMAT (2733): 2.5 mass % Pd.

the kinetics of hydrogenation of -methyl styrene to cumene on

Pd is known to be irreversible, rst order with respect to
dissolved hydrogen and zero order with respect to -methyl
styrene. Further, it was in a mass-transfer controlled regime.
Therefore, the hydrogen transfer to the surface was the limiting
step. Because the smaller particles have larger specic area, the
rate was higher for smaller size. The rate increased with increasing centrifugal force and decreased with increasing liquid
velocity as the liquid lm became thinner.
Figure 6 shows the reaction rates with the foam metal. Note
that the reaction can occur only on the surface of the catalyst
because it is nonporous, unlike the alumina particles. Even with
alumina particles, the reaction occurs in a thin outer shell of the
particles. Therefore, the trend in the reaction rate was similar to
that with the alumina particles as the reaction took place. The
rates increased as the liquid lm became thinner with increasing centrifugal force eld. However, the rate was greater for the
smaller specic area for the metal foams. It is unlikely that
there is a liquid maldistribution. Even if it was present, the
extent of liquid maldistribution may not differ signicantly
between the two foam packings. Burns and Ramshaw13 observed the changing nature of the liquid ow with rotational
speed over a reticulated wire-mesh packing. At low rotational
speeds, the liquid ow covered the pore openings of the
network of the laments as lms, which they termed pore ow.
With increasing rotational speed, the pore ow turned into a
droplet ow in which the tiny droplets jumped from one
lament to the other. If the lm ow was laminar, there would
be concentration gradients within the lm, as is the case with
the pore ow. However, the periodic mixing of liquid within
the lm leads to higher mass-transfer rates. Such a mixing is
possible with the drop ow. The nature of ow may depend on
the size of the pore openings of the foam metal. There are

December 2005

indications that higher specic area does not always lead to

higher mass-transfer rates.3 The understanding of the nature of
liquid ow and of possible liquid maldistribution could help to
enhance the process intensication.
Further, the rates with the alumina particles were substantially lower than those of the metal foams, even though their
specic surface areas were comparable. In the case of the
former, the area covered by the liquid pendular rings, held
between the adjacent particles, would be very small, unlike that
for the gravity ow. However, it cannot account for the rate
being half that of the foam. Munjal et al.14 reported the gas
liquid interfacial area in a rotating bed of spherical particles.
Their data indicate that the gasliquid interfacial area could be
less than half of the specic packing area. There could be a
partial wetting of the bed, which may depend on the type of
packing. There is a need for deeper understanding of the nature
of liquid ow and its texture in the rotating beds.
To quantify the process intensication, that is, the volume
reduction that could be obtained with the rotating beds, we
require the reaction rate data for the conventional trickle-bed
reactor. Herskowitz et al.10 and McManus et al.15 reported the
reaction rates in trickle beds under similar conditions based on
the mass of catalyst. We have converted their rate data to
volume basis using their particle density and bed porosity; their
maximum reaction rates were 4.9 107 and 1.44 107
mol/(cm3 of bed s1). The former used catalyst of 0.75 mass %
Pd and 1.34 diameter, whereas the latter used 1.5 mass % Pd
and 3.3 mm diameter. The rates depend on the mass % Pd and
the size of the particle. The experimental rate data for this
reaction system under similar conditions have been reported in
trickle beds, on strings of beads, and single pellets with different percentages of Pd loading and particle diameters by several
groups of investigators. Therefore, these data had to be reduced
to a common basis for the use as a reference scale for quantifying the process intensication. It is reasonable to use the
surface area rather than the volume of bed or the mass of the
particles as the basis for the reaction rates, given that the
reaction occurs within a thin outer shell of the particles. The

Figure 7. Comparison of reaction rates of different investigators on specic area basis.

E, Herskowitz et al.10: bed, 0.75 mass % Pd; , Herskowitz
et al.10: bed, 2.5 mass % Pd; , Subramaniam18: string, 2.5
mass % Pd; f, Sivalingam4: string, 2.5 mass % Pd; ,
Sattereld et al.17: string, 1.0 mass % Pd; , Funk et al.19:
single pellet, 1.0 mass % Pd; , McManus et al.15: bed, 1.25
mass % Pd.

Vol. 51, No. 12

AIChE Journal

that of the conventional trickle bed reactors. Because the reaction rate would be greater with larger beds and higher rpm, the
volume reduction of similar magnitude or more could be
achieved. The intrinsic kinetic rate for this reaction is 596
107 mol/(cm3 s1) at 50C with 1.0 mass % Pd and 1 atm
hydrogen pressure. This rate is about ve times the highest rate
observed. At these conditions the intrinsic kinetic regime is
being approached, and thus further intensication may not be
In the present study, the reaction rate was controlled by the
liquid-side mass transfer. For the case in which the reaction
rate is controlled by the gas-side transport, the novel rotor
design proposed by Rao et al.5 can be used to achieve process
intensication. If the rate is controlled by the reaction kinetics,
it is unlikely that the process intensication can be achieved
with a rotating trickle-bed reactor.

Figure 8. Comparison of reaction rates for different substrates.
, RECEMAT (1733): 1.5 mass % Pd, 360 g; , 1.5-mm
alumina beads, 1.0 mass % Pd, 485 g; F, conventional trickle
bed: 1.0 mass % Pd; - - -, Herskowitz et al.10

data reported by Morita and Smith16 and Sattereld et al.17

indicate that the intrinsic kinetic constant varies linearly with
mass percentage of palladium up to 5%. Therefore, one would
expect that the reaction rate data based on the surface area and
palladium mass % of different investigators should be in agreement. Sivalingam et al.4 computed the reaction rate data reported by several workers in terms of mol/(cm2 mass % Pd1).
The liquid rate per string was computed by considering that the
bed is composed of several strings of spheres. The number of
strings was set as the number of spheres that can be packed in
a circular cross section of the bed perpendicular to the direction
of ow in a triangular pitch with spheres touching each other.
Figure 7 shows the comparison of the rates reported by different groups on a common basis. There was a considerable
deviation in the rates reported by different investigators at very
low ow rate, but the rate was in good agreement for most
cases and it was nearly constant beyond the liquid ow rate of
0.05 cm3/s. Therefore, we may consider the reaction rate to be
0.22 107 mol/(cm2 mass % Pd1 s1).
Because the process intensication requires the quantication in terms of volume reduction, we have converted the
reaction rate from unit area basis to unit volume basis for the
bed for 2.2-mm particles. The rate is 3.14 107 mol/(cm3
s1) for 1.5 mass % Pd.
Figure 8 shows the comparison of the reaction rates in
rotating beds with alumina beads and metal foam and in conventional trickle beds computed as mentioned above. To show
direct comparison, we have plotted the data of Herkowitz et
al.10 and have presented only the highest rates for the alumina
particles and the metal foam. It can be seen that the rate with
the RECEMAT (1723) metal foam was the highest. The rate
in the conventional trickle-bed reactor was in the range of
3.57 107 to 2.85 107 mol/(cm3 mass % Pd1 s1).The
highest rate with metal foam was in the range of 120 107
to 110 107 mol/(cm3 mass % Pd1 s1). The enhancement
in the rate with rotating beds was 33 to 39 times higher than
AIChE Journal

December 2005

The hydrogenation rates of -methyl styrene were measured

in a rotating trickle bed using alumina particles and metal foam
as catalyst supports. The rates with metal foam were nearly
twice as great as those with spherical particles of the same
specic area. Further, the metal foam with smaller specic area
gave higher rates. These trends in the reaction rates were
attributed to the nature of liquid ow over the packing. It was
shown that a volume reduction by a factor of 30 to 40 is
attainable with moderate centrifugal acceleration of 500 g
(4900 m/s2), if the liquid-side mass transfer is the controlling

We gratefully acknowledge Recemat International B.V., The Netherlands (http://www.recemat.com) for donating the metal-foam rings; Norton
Company, Philadelphia, PA for donating the alumina beads; and the partial
nancial support from Department of Science and Technology, Government of India.

Literature Cited
1. Ramshaw C. Higee distillationAn example of process intensication. Chem Engineer. 1983;389(Feb.):13-14.
2. Trent D, Tirtowidjojo D, Quarderer G. Reactive stripping in a rotating
packed bed for the production of hypochlorous acid. Process Intensication for the Chemical IndustrySmaller, Cheaper and Safer Production (Proceedings of the 3rd International Conference, Antwerp,
Belgium). Suffolk, UK: BHR Group Limited/PEP; 1999:217.
3. Rao DP, Bhowal A, Goswami PS. Process intensication in rotating
packed bed (HIGEE): An appraisal. Ind Eng Chem Res. 2004;43:11501162.
4. Sivalingam G, Radhika M, Rao DP, Rao MS. Process intensication in
a model trickle-bed reactor. Ind Eng Chem Res. 2002;41:3139-3144.
5. Rao DP, Goswami P, Bhowal A. A novel rotor design for intensication of gas-phase mass transfer in a Higee. Proc of AIChE Spring
Meeting, New Orleans, LA, April 2529; 2004.
6. Gianetto A, Specchia V. Trickle-bed reactors: State of art and perspectives. Chem Eng Sci. 1992;47:3197-3213.
7. Van Hasselt BW, Lebens PJM, Calis HPA, Kapteijn F, Sie ST,
Moulijn JA, van den Bleek CM. A numerical comparison of alternative
three-phase reactors with a conventional trickle-bed reactor. The advantages of counter-current ow for hydrodesulfurization. Chem Eng
Sci. 1999;54:4791-4799.
8. Smith JM. Chemical Engineering Kinetics. 3rd Edition. New York,
NY: McGraw-Hill; 1981:536.
9. Singh SP, Wilson JH, Counce RH, Villier-Fisher JF, Jennings HL,
Lucero AJ, Reed GD, Ashworth RA, Elliot MG. Removal of volatile

Vol. 51, No. 12



organic compounds from groundwater using a rotary of air stripper.

Ind Eng Chem Res. 1992;31:574-580.
Herskowitz M, Carbonell RG, Smith JM. Effectiveness factors and
mass transfer in trickle-bed reactors. AIChE J. 1979;25:272-283.
Lowenheim FA. Electroplating. New York, NY: McGraw-Hill; 1978.
Henry JR. Electroless (autocatalytic) plating. Met Finish. 2000;98:
Burns JR, Ramshaw C. Process intensication: Visual study of liquid
maldistribution in rotating packed beds. Chem Eng Sci. 1996;51:13471352.
Munjal S, Dudukovic MP, Ramachandran P. Mass-transfer in rotating
packed bedsII. Experimental results and comparison with theory and
gravity ow. Chem Eng Sci. 1989;44:2251-2268.
McManus RL, Funk GA, Harold MP, Ng KaM. Experimental study of


December 2005


reaction in trickle-bed reactors with liquid maldistribution. Ind Eng

Chem Res. 1993;32:570-574.
Morita S, Smith JM. Mass transfer and contacting efciency in a
trickle-bed reactor. Ind Eng Chem Fundam. 1978;17:113-120.
Sattereld CN, Pelossof AA, Sherwood TK. Mass transfer limitation
in a trickle bed reactor. AIChE J. 1969;15:226-234.
Subramaniam S. Reaction Rate Studies in a Model Trickle-Bed Reactor. MTech Thesis. Kanpur, India: Indian Institute of Technology;
Funk GA, Harold MP, Ng KaM. Experimental study of reaction in a
partially wetted catalytic pellet. AIChE J. 1991;37:202-214.

Manuscript received Jul. 2, 2004, and revision received Mar. 29, 2005.

Vol. 51, No. 12

AIChE Journal

Novel Micromixers Driven by Flow Instabilities:

Application to Post-Reactors
S. R. Deshmukh and D. G. Vlachos
Dept. of Chemical Engineering and Center for Catalytic Science and Technology (CCST), University of Delaware,
Newark, DE 19716
DOI 10.1002/aic.10591
Published online August 22, 2005 in Wiley InterScience (www.interscience.wiley.com).

Flow-driven instabilites in microdevices, consisting of small features (posts), are

studied using 2-D computational uid dynamics simulations as a means of enhanching
mixing at the microscale. It is found that oscillatory ow developed above a critical
Reynolds number in the wake of a post is weakly affected by volumetric and surface
(catalytic) reactions. The observed reactivity, on the other hand, is strongly affected by
ow-driven instabilities. Based on these results, a novel post micromixer that capitalizes
on this oscillatory ow while exhibiting a low pressure drop is proposed. Finally, different
feature congurations are investigated to gain insights into the mixing performance and
to develop design guidelines. 2005 American Institute of Chemical Engineers AIChE J, 51:
31933204, 2005

Keywords: microreactors, oscillations, micromixers, Hopf bifurcation, bluff bodies, von

Karman vortex

Microscale devices have recently been explored for numerous applications. Examples include in situ production of toxic
and specialty chemicals, mixing, high throughput screening of
catalysts, chemical and biological agents, energy production,1-5
in situ microfabrication of structures of sizes 5 m within
small capillaries,6 chemical and biological analysis and synthesis,7-10 DNA separation and amplication using the polymerase
chain reaction,11 and even medical diagnostics. Because of
their small size, the ow in these devices is laminar and mixing
is driven only by molecular diffusion. Their small size also
translates into low surface areas for catalytic reactions.
To overcome some of the challenges stemming from device
miniaturization, ideas reminiscent of microstructured catalysts12-14 are used to fabricate microscale features in the form of
cylindrical posts (or other shapes) within a reactor to create
large catalytic surface area and possibly enhance mixing. Such
microscale posts have recently been implemented experimenCorrespondence concerning this article should be addressed to D. G. Vlachos at

2005 American Institute of Chemical Engineers

AIChE Journal

December 2005

tally for enhanced chemical reactivity15,16 and for ow uniformity.17

It is well known that under suitable laminar ow conditions,
ow past bluff bodies gives rise to hydrodynamic instabilities.18 These ow-driven instabilities generate periodic wakes
behind the bluff body that lead to the formation of the von
Karman vortex street. Such vortices can prove useful in enhancing mixing and thus the effective reactivity in microdevices. Reacting turbulent wakes past bluff bodies have been
well studied in the combustion literature (see, for example,
Mavridis et al.19); however, not many studies have been reported for laminar ows. Mass transfer enhancement observed
for multiphase ow arising from bubble wakes20 and in spacerlled narrow channels of spiral-wound membranes21 and preliminary investigations for a catalytic reaction occurring at the
surface of an oscillating cylinder22 constitute a few studies
reported for laminar ow conditions, although a systematic
investigation of this subject is lacking.
In this article, we rst present a brief overview of laminar
ow patterns in ow past bluff bodies. Time-dependent computational uid dynamics (CFD) simulations are performed and
our results are compared to the existing literature at the macroscale. Next, instabilities driven by the ow past a single
Vol. 51, No. 12


micropost are investigated with a focus on understanding the

coupling between reactions and ow dynamics, a relatively
poorly understood topic, for both volumetric and surface reactions. These investigations are nally extended to an array of
posts, which is representative of an actual microdevice, to
explore the practicality of using post micromixers and to develop design principles for them.

Laminar Flow Past Bluff Bodies and Relevance to

In this section we present a brief overview of the existing
literature on the ow past bluff bodies. Most of these studies
have been carried out at the macroscale for ow past a single
cylindrical post (that is, a cylinder) with a diameter ranging
from 1 mm23 to 50 mm.24 The ow patterns were identied
using both experimental ow visualization techniques23-27 and
simulations28-33 and are typically characterized based on the
Reynolds (Re) number using the post diameter, dened as



where is the uid viscosity (kg m1 s1), is the uid

density (kg/m3), d is the post diameter (m), and U is the free
stream velocity (m/s).
Figure 1a shows typical ow patterns in terms of the boundary layer behind a post (ow patterns are based on our simulations detailed below). At very low Re number (Re 1), the
ow is creeping, and no boundary layer separation is observed
behind the post.34 In the range of 2 Re 6, a recirculation
zone appears behind the post with two symmetrical and xed
eddies, although the ow is still steady. The eddy size increases
with increasing Re.35 At a critical Re number, 40 Rec 50
(typically around 47), the ow becomes unsteady, eddies behind the post lose their symmetry, and an asymmetric timedependent wake appears behind the post. Vortices, parallel to
the post, are shed alternately from the upper and lower sides of
the post, giving rise to a von Karman vortex street. This
transition from a steady to a periodic unsteady ow corresponds to a Hopf bifurcation and is described by the Stuart
Landau equation.36 The periodic ow is characterized by the
Strouhal (St) number, that is, the dimensionless frequency of
the oscillations, dened as



where f is the frequency of oscillations (s1). With further

increase in the Re number, the onset of oblique vortex shedding
takes place around Re 190, giving a three-dimensional (3-D)
character to the wake.37,38 As the Re number increases further,
the wake becomes irregular in space and time and ultimately
the ow becomes turbulent at Re 2100.
The onset of time-dependent asymmetric wakes, that is, the
value of critical velocity, depends on the shape as well as the
size of the post. The effect of the post size is obvious. A larger
post diameter translates to a higher Re number, which implies
that the onset of oscillatory behavior shifts to slower ows. The
effect of post shape has been investigated by Jackson.39 It was

December 2005

Figure 1. Flow patterns for different values of Re number.

Panel (a) indicates the boundary layer behind a post for low to
moderate Re numbers. Panels (b)(d) show schematics for
inline or tandem, side-by-side, and staggered
conguration of two posts, respectively, along with the velocity contours at Re number of 123 and a separation between
the centers of two posts of 2d. [Color gure can be
viewed in the online issue, which is available at www.

reported that the shape of the bluff body affects the critical Re
number. For a given length perpendicular to the ow, the
critical Re number increases with an increase of the size
parallel to the ow (such as Rec 27.8 for a at plate and
45.5 for a cylinder).
Residence times in microchemical devices range from milliseconds to a few seconds. Larger residence times are common
for slow chemistries and low temperatures, such as thermal
decomposition and reforming reactions, whereas millisecond
residence times are characteristic of fast chemistries, such as
combustion and catalytic partial oxidation reactions. Typical
post sizes in structured microdevices are of the order of hundreds of microns. For a device of length of L 1 cm and a
residence time of 1 ms, the Re number, based on a post
diameter of 300 m, (considering uid properties of air) is 75.
These back of the envelope calculations clearly show that
Vol. 51, No. 12

AIChE Journal

ow-driven oscillatory instabilities at the microscale are entirely feasible and could play an important role in enhancing
mixing, at least for the fastest reactions. When liquids are
considered, the transition to oscillatory ow occurs for slower
A limited number of experimental and theoretical studies are
available for multiple post congurations. For the case of two
cylinders, the experimental studies40-43 have focused on the
high Re number (103104) regime, whereas numerical studies44,45 have focused on the low Re number (102103) regime.
Different ow patterns have been reported depending on the
conguration. They are briey summarized here and the reader
is referred to the above references for details. A conguration
of two posts with an angle 0 or 180 is termed as inline or
tandem conguration (see Figure 1b). For a separation 3d,
no distinct vortices are seen behind the rst post. For larger
separations, vortices are seen behind both posts. The shedding
frequency is initially lower and then reaches the single post
value for larger separations. In a side-by-side conguration
(angle of 90 between the two posts; see Figure 1c), the two
posts act like a single body for spacing 1.4d. For spacing
larger than this but 2d, a bistable vortex shedding, which is
unique to this conguration, occurs with two different frequencies observed behind each of the posts. This biased ow (the
wake behind one post is larger than that behind the other)
changes intermittently. For all other orientations of the posts
(staggered conguration; see Figure 1d), the wake behind
the upstream post is found to be narrower and sheds with a
higher frequency than that of a single post. Similar but more
complex ow patterns have been observed experimentally for
three and four posts.46-49 A simple superposition principle does
not provide an adequate explanation of these ow patterns. A
complete understanding of the highly nonlinear multiple post
vortex shedding still remains elusive.

CFD Modeling, Numerics, and Validation

Laminar ow is rst studied past an innitely long cylindrical post. In a typical microdevice, the height of the device is
about 510 mm,16 whereas the post diameter is in the 0.1- to
0.5 mm range. Because of this large aspect ratio (post diameter
to height), a 2-D model is used in this study. A 200 m
diameter cylindrical post is investigated. The size of the ow
domain simulated around a post is known to affect the solution.
Typical domains used in literature are 10 diameters (wide)
20 diameters (long).28,32,50-52 A recent study investigated the
effect of nite domain on the ow past a cylinder albeit under
steady-state conditions.53 With increasing domain size, the
eddies become fully developed, resulting in an increase in the
time period (that is, decrease in the St number). Thus, a larger
domain of 20 diameters (wide) 25 diameters (long) has been
used here to minimize nite domain effects. A nonuniform
mesh was generated with 200 nodes along the ow direction
and 160 nodes perpendicular to it. A ner mesh is generated
around the cylindrical post with about 104 nodes on the post
surface, resulting in a total of nearly 52,000 nodes within the
domain. The geometry and the mesh are shown in Figure 2.
Simulations with ner meshes were performed, and it was
found that this mesh provides an adequate balance between
accuracy and computational effort. For example, the change in
the St number between the domain size and mesh used here and
AIChE Journal

December 2005

Figure 2. Flow domain and computational grid.

The simulated domain size is 4 5 mm and the post diameter
is 200 m. The blowup shows the ner mesh near the post
surface. The solid arrows indicate the ow direction. [Color
gure can be viewed in the online issue, which is available at

the next larger domain (an increase by 5d along the width and
by 10d in the length) or a ner mesh (resulting from doubling
the number of nodes) was found to be 1%. The optimization
of computational effort is especially important because timedependent calculations, involving thousands of time steps, are
In most simulations, no symmetry is assumed. However, for
a limited number of cases the effect of symmetry has been
explored. Specically, for simulations with a single post where
symmetry was imposed, half of the above domain was simulated.
For multiple-post simulations, the width of the domain is
kept at 20d and the length is increased to 37.5d (this translates
to a domain of 4 7.5 mm, which is representative of a
microdevice). The nonuniform mesh discretization in this case
is 120 nodes (width) 225 nodes (length) 60 nodes (on the
surface of each post), which results in a mesh consisting of
nearly 200,000 nodes.
The commercial CFD software, Fluent (version 6.1.22)54 is
used to obtain time-dependent solutions using a nite-volume
method. The time-dependent 2-D continuity, momentum, energy, and species conservation equations are solved using a
second-order, implicit time-stepping method. A rst-order upwind discretization scheme is used for density, momentum,
species, and energy equations and a Standard pressure
interpolation scheme is used along with the Simple algorithm for the pressurevelocity coupling.54 A segregated solver
with an underrelaxation scheme is used to obtain the solution.
The equations are integrated with a time step of 0.5 s,
typically, for a total of 5000 10,000 time steps and up to
65,000 time steps. Time-dependent statistics are obtained after
3000 initial time steps to eliminate initial transients. A single
simulation, as described above, takes about 1 week of CPU
time on a single Pentium Xeon (2.4 GHz with 2GB of RAM)
The inlet gas (A), having properties of air (that is, molecular
weight of 29 and a diffusivity of 2.9 105 m2/s) is assumed
to ow in at a uniform velocity and a xed inlet temperature of
923 K. These conditions come from our previous work on
ammonia decomposition in post microreactors that ensure complete decomposition of ammonia.55 Re 100 is a standard
benchmark problem in the uid dynamics community,56 and is
also chosen here, unless otherwise stated. These reactive ow
Vol. 51, No. 12


simulations are carried out under isothermal conditions, unless

otherwise stated. The outlet is at a xed pressure of 1 atm and
the diffusive ux is set to zero. No-slip boundary condition is
applied at the post surface(s). The physical properties are the
same for all gaseous species.
To investigate the interplay of uid ow and catalytic reactions, a reaction
A 3 nB


of variable order is allowed to occur at the post surface(s) with

a rate given by

kmol m2 s1 148CA


where is the reaction order in A and Ci is the concentration

of species i (kgmol/m3). The reaction rate constant is chosen so
that the surface concentration of the reactant is close to zero.
To investigate the interplay of uid ow and volumetric
reactions, a reaction
AB 3 C


which is rst order in both A and B, is considered. Pure

components A and B enter the domain with equal ow velocities in a nonpremixed manner from the top and bottom half of
the channel, respectively (see the inset of Figure 6a below). It
should be noted that this reaction is highly dependent on
mixing and occurs at the interface between the two uids. The
rate of this volumetric reaction is given by

kmol m3 s1 2.7 105 CA CB


The reaction rate constant is chosen so that signicant exit

conversions (20%) occur.
To conrm the existence of oscillations at the microscale and
validate our approach, simulations were carried out for a single
post in the absence of chemical reactions over a range of Re
numbers. A comparison to available experimental and simulation data, albeit at the macroscale, is presented in the Appendix. Studies involving either surface or volumetric reactions in
the presence of a single post are undertaken next.

Effect of Chemistry on Flow-Driven Instabilities

To investigate the effect of chemistry on ow-driven instabilities, the catalytic reaction (Eq. 3) is allowed to occur at the
post surface for various reaction orders and values of n. Figure
3a shows typical velocity magnitude contours and Figure 3b
the mass fraction contours of species A for the rst-order
reaction A 3 B. It is clearly seen that asymmetric wakes
develop behind the micropost. The developing wake near the
post surface gradually increases in size and is nally shed into
the ow. The alternate shedding of wakes gives rise to a von
Karman vortex street (clearly seen in Figure 3b).
To gain further understanding, a point 3d downstream from
the post, termed as a reference point, is selected (see Figure 3a)
to monitor local variables. Figure 3c shows the phase portrait
of local mass fraction of species A and the transverse velocity
(Y-velocity) component for 5000 time steps. The retracing of

December 2005

Figure 3. Contours of (a) velocity magnitude (m/s) and

(b) mass fraction of species A for ow past a
catalytic post with a rst-order reaction A 3 B
occurring at its surface.
The phase portrait at the reference point shown in (a), indicating the periodicity of the ow, is depicted in (c). [Color
gure can be viewed in the online issue, which is available at

the curve clearly indicates the periodicity of the ow but also

the asymmetry of the two lobes.
The Y-velocity at the reference point as a function of time is
shown in Figure 4a (crosses). The corresponding mass fraction
of species A is shown in Figure 4b (bottom solid line). The
distance E in Figure 4b constitutes one period of the oscillation that is about 59 s. The intermediate peak splits the
period into two unequal parts, consistent with the phase portrait
in Figure 3c. During one set of alternately shed wakes, the
Y-velocity goes through a complete cycle but the mass fraction
(being a scalar quantity as opposed to the Y-velocity) goes
through two cycles, causing the splitting of the peak.
Figure 4a shows also the local Y-velocity as a function of
time for the cases without reaction (solid line) and with a
surface reaction A 3 B that is second-order in A (circles). All
trajectories overlap, indicating that the effect of surface reaction, with respect to its presence and reaction order, on oscillations is negligible.
Next, a rst-order reaction involving a change in the number
of moles, A 3 2B, is studied (dashed lines in Figures 4a and
Vol. 51, No. 12

AIChE Journal

affects the oscillatory ow (an increase in the amplitude of the

oscillations by about 50% and a decrease in the period of
oscillations by about 15% have been observed). Finally, simulations involving nonisothermal reactions, where the change
in density is driven by temperature variation only, were also
found to affect the oscillatory instabilities, that is, reaction
exothermicity and endothermicity caused an increase and a
decrease in time period of the oscillations, respectively (data
not shown).

Effect of Flow-Driven Instabilities on Chemistry

Here we investigate the effect of ow-driven instabilities on
chemistry. To this end, the ow past a single catalytic post
where a surface reaction A 3 B (rst-order in A) occurs on its
surface was rst solved in a half of the domain by imposing
symmetry. The corresponding velocity magnitude and mass
fraction contours are shown in Figures 5a and 5b, respectively.
The symmetry constraint eliminates the formation of asymmetric wakes and results in a stable stationary solution (a single

Figure 4. Effect of a catalytic post on oscillatory behavior for various types of surface reactions.
Trajectories of (a) the transverse velocity component (m/s)
and (b) mass fraction of species A, for the catalytic surface
reaction A 3 nB for various values of n and reaction orders
indicated. (c) Snapshot of the density contours (kg/m3) for a
rst-order reaction A 3 2B. [Color gure can be viewed in
the online issue, which is available at www.interscience.

4b). In this case the amplitude of the oscillations decreases by

about 10% and the period of the oscillations increases slightly
by about 2.5%. The density change caused by the difference in
the molecular weights of the reactant and product (such as in
the reaction A 3 2B, where the molecular weight of A is twice
that of B) affects the ow because of its compressibility. The
actual density contours observed are shown in Figure 4c. A
maximum change in the density of almost 100% occurs as a
result of the change in the number of moles.
Next, volumetric reactions were studied. A simple reaction
A 3 B has no effect on the oscillations, as has also been
observed for surface reactions. However, the uid phase reaction (Eq. 5) involving a change in the number of moles again
AIChE Journal

December 2005

Figure 5. Effect of symmetry of a catalytic post on mixing and conversion.

Contours of (a) velocity magnitude (m/s) and (b) mass fraction of species A for symmetric ow past a catalytic post with
a rst-order surface reaction A 3 B. (c) Comparison of the
time-averaged conversion at the reference point with and
without symmetry imposed for a catalytic surface reaction A
3 nB for various values of n and reaction orders. [Color
gure can be viewed in the online issue, which is available at

Vol. 51, No. 12


absence of a post, the reactants mix only by molecular diffusion and the reaction is conned to a very small zone. In the
presence of a post, on the other hand, the ow-driven oscillations create an enlarged reaction zone. Figure 6a shows the
cup-averaged mass fraction of product C at the outlet as a
function of time. Signicantly higher conversions are found in
the oscillatory case, which is explained by the larger interfacial
area created by the wakes behind the post.
In summary, it is found that ow-driven instabilities significantly affect the observed reactivity. The absence of the vortex
shedding (symmetry enforced) would lead to the formation of
a single wake and higher conversions for non productinhibited
positive order surface reactions. For volumetric reactions between non-premixed reactants, the wakes enhance the contact
area and can result in much higher conversions.

Arrays of Posts

Figure 6. Effect of a noncatalytic post on reactivity for a

volumetric reaction A B 3 C with non-premixed reactants.
(a) Conversion vs. time. Contours of mass fraction of product
C without a post (b) and with a post (c). The inset is a
schematic of the ow geometry. [Color gure can be viewed
in the online issue, which is available at www.interscience.

Investigations of a post array, which is more representative

of an actual experimental microdevice, are considered in this
section. In the following simulations, the length of the domain
has been increased to 7.5 mm. As alluded to earlier in the
overview of the previous work on ow past bluff bodies, the
arrangement of the posts can play a role in mixing. For example, constructive interference between the wakes behind successive posts could result in increased wake amplitude, causing
enhanced mixing. Regarding the transverse post spacing, as
mentioned in the literature overview, when the post separation
is small, multiple posts act like a single one, resulting in a
single wake. On the other hand, when the transverse spacing is
large enough, all the wakes are independent of each other and
in-phase. Finally, for intermediate spacings a bistable wake is
expected. Independent wakes behind each post allow tuning of
the mixing and, thus, the corresponding separation of 3d in the
transverse direction was used in the simulations that follow.
To quantify the extent of mixing, a mixing parameter is

mix 1 2
elongated wake behind the post) without oscillations. This
nding clearly indicates that symmetry breaking is a necessary
condition for the formation of oscillatory wakes.
Figure 5c shows the effect of symmetry on the time-averaged conversion calculated at the reference point for various
values of n and reaction orders of the surface reaction (Eq. 3).
In all cases the time-averaged conversion in the oscillatory ow
is lower than the corresponding steady-state conversion for the
symmetric case. For the latter case, the single wake behind the
post creates a stationary region of lower velocity and higher
residence time and thus of higher conversion. The time-averaged exit conversion is also found to be higher for the symmetric case but the difference is marginal (in the decimal
places) because of the low surface area of a single post.
The case of volumetric reactions of non-premixed reactants
is studied next. Equal ow rates of non-premixed components
A and B enter the domain (as shown in the schematic in Figure
6a) and react according to the volumetric reaction (Eq. 5).
Figures 6b and 6c depict the mass fraction contours of product
C in the absence and presence of a post, respectively. In the

December 2005

y i 0.5VdS



where yi is the mass fraction of species A at the transverse node

i, V is the magnitude of the local velocity, and S is the
cross-sectional area of the exit (device height for a 2D case).
The quantity within the brackets represents the cup-mixing
average of the local deviation from the perfectly mixed case.
The value of mix 0 implies a totally unmixed case, whereas
mix 1 indicates complete mixing.
First, the effect of the number of posts on mixing is studied
for two non-premixed streams. Figures 7a7d show the mass
fraction contours of species A for the case of no posts and for
arrays with one, two, or ve rows of posts each, respectively,
with a row spacing of 3d. Even with a single row of posts,
wakes in the ow develop behind all posts (not shown). However, only the post located at the interface of the two uids
plays an active role in mixing (see Figure 7b).
The corresponding mixing parameter at the exit as a function
of time is shown in Figures 8a 8d. The creation of chaotic ow
is one possible means of increasing the extent of mixing
Vol. 51, No. 12

AIChE Journal

Figure 7. Contours of mass fraction of species A indicating the effect of different number of post rows and spacing
between posts on mixing.
[Color gure can be viewed in the online issue, which is available at www.interscience.wiley.com.]

achieved. Although the ow over a single post is relatively

simple and periodic (as reported in Noack and Eckelmann57
and seen in Figures 3c and 4a), the dynamics in an array of
multiple posts is rather complex, as shown in Figure 9. It has
been previously hypothesized that multiple cylinders in a owing stream may not produce chaotic turbulence.58 Over the
relatively short time window (12 ms) we have been able to
simulate with long simulations (of the orders of 3 4 weeks),
we do not observe periodic ow. Obviously this time window
is too short to denitively conclude about long-term dynamics.

The attractor appears complex but longer simulations with

more efcient numerical schemes or parallelization are needed
to study the long-term dynamics. We leave this task for future
research. (Preliminary simulations over 40 ms indicate periodicity with a period of 5.3 ms.) Finally, one has to keep in
mind that the reported values of mixing parameter are the
corresponding time averages over these relatively short time
intervals, so the results should be valuable to understand trends.
It is interesting that an increase in the number of posts does
not always lead to increased mixing. In fact, the actual mixing

Figure 8. Effect of the number of post rows and row spacing on mixing.
Panels (a)(f) depict the average mixing parameter at the exit for conditions corresponding to the contours in Figure 7. [Color gure can be
viewed in the online issue, which is available at www.interscience.wiley.com.]

AIChE Journal

December 2005

Vol. 51, No. 12


Figure 9. Effect of feature staggering.

Contours of species mass fraction for ow in a 5 4 post array arranged in different congurations shown in (a)(c). (d) Corresponding
average mixing parameter at the exit. [Color gure can be viewed in the online issue, which is available at www.interscience.wiley.com.]

is slightly reduced when more than one row of posts is used for
a relatively small spacing between rows. For this spacing,
wakes do not have room between successive rows of posts
to develop. Apparently, this post spacing does not lead to
constructive interference. As indicated in Figures 7c and 7d,
the small separation between the rows of posts (of 3d)
renders the multiple rows effectively acting as a single row.
The pressure drop in the micromixer is another important
performance criterion. Because of the relatively open conguration, a small pressure drop (2 Pa) is observed for a single
row of posts. For ve rows of posts, the pressure drop is still
small (250 Pa), indicating no additional pumping devices
will be required to overcome the loss of pressure head resulting
from the enhanced mixing achieved in these post micromixers.
Figures 7e and 7f show the species contours for increased
separation of values of 4.5d and 10d, respectively, between
two rows of posts and Figures 8e and 8f show the corresponding mixing parameter. With an increase in row interspacing
( 4.5d), a wake develops behind the rst row that causes a
constructive interference with the wake from the second row of
posts, resulting in an increase in the wake amplitude. This
further results in enhanced mixing, as seen from the mean
values in Figure 8e. Because of the larger amplitude of the
oscillations, the standard deviation, however, increases from a
value of 0.005 to 0.0184 with this increase in row spacing.
With a further increase in spacing ( 10d), the wakes behind
the rst row of posts reach the second row with a very large
amplitude, causing a destructive interference and resulting in
worse performance but still better than that of the 3d
separation case. Given the variation in the mixing performance
as a function of row spacing, the separation between the two
rows is systematically varied in integer multiples of post diameter (at Re 60 and 100). The results shown in Figure 10
indicate that mixing is a highly nonlinear function of row

December 2005

spacing and the optimum mixing occurs for separations in the

range of about 4d to 5d and at Re 60, in the range of about
5d to 7d. Due to the large number of posts present in the
system, a signicant number of time steps (30,000 to 35,000)
is required to ensure that the change in the cumulative average
mixing parameter with every 2,000 more time steps is less than
1%. A comparison of the wakes for the two Re numbers
indicates that the wake length is longer at Re 60. Also, the

Figure 10. Effect of interrow spacing on mixing.

The mixing enhancement in an array of two rows (of ve
posts each) is found to be a highly nonlinear function of the
inter row spacing and of the Reynolds number. Averages
reported here are obtained over 18 ms. [Color gure can be
viewed in the online issue, which is available at www.

Vol. 51, No. 12

AIChE Journal

slower ow at Re 60 ensures that the wake behind the rst

post extends to the second post for larger post separations, and
hence, the optimum in spacing for this slower ow occurs
further apart and over a longer range of spacings (see Figure
10). A more detailed investigation on the optimal conguration
of the post micromixer will be the subject of a forthcoming
The conguration whose post row spacing is 4.5d is used to
determine the effect of post arrangement on mixing. Three
different congurations are explored as shown in Figures 9a
9c. Figure 9d shows the corresponding mixing parameter at the
exit. In the (100) type of conguration, an increase in wake
amplitude is seen at the very end of the post array. The same
conguration rotated through 45 (Figure 9b) causes the wake
amplitude to increase from the second row of posts. The wakes
are shed symmetrically and thus the standard deviation is
smaller (0.0184 vs. 0.0347 in the previous case). The conguration of Figure 9c, on the other hand, is skewed and the
shedding of wakes is also skewed, resulting in a larger standard
deviation despite a slightly better mixing performance. Our
simulations indicate that increasing the amplitude of the wake
holds the key to enhancing the mixing and the post staggering
is not as critical, at least for the congurations explored herein.
Enhanced mixing can be used to increase the conversion of
mixing-sensitive reactions. For example, for the volumetric
reaction A B 3 C, oscillations using an array of posts
enhance the conversion by 100% (data not shown).

Comparison to Other Micromixers and Potential

Micromixers aimed at enhancing the mixing under laminar
ow conditions can be characterized into two types.59,60 The
rst one entails active micromixers that use variable-frequency
pumps61,62 or ultrasonics63 and are more complex and prone to
break down. On the other hand, passive or static micromixers have no moving parts and rely on laminar shear ow (which
causes the stretching of uid elements)64,65or distributive mixing (physical splitting of uid streams into smaller segments)66,67 or hydrodynamic focusing68,69 or vortices.70-73 With
the exception of interdigital mixers, the above micromixers are
designed for low throughput because they typically use channels 100 m.
Micromixers based on the ow-driven oscillatory instability
have been exploited here. Figure 11a shows the average mixing
parameter at the exit as a function of the Re number for a 5
4 post array with 4.5d (conguration of Figure 9b) and the
associated pressure drop is shown in Figure 11b. Below the
critical Re number, mixing is comparable to that without posts
because diffusion is the only mixing mechanism in both cases.
In fact, the posts show a slightly worse performance as the
tortuous path lowers the effective diffusivity in the post array
(see Deshmukh et al.74for the effective diffusivity). As the Re
number increases, mixing becomes worse as a consequence of
decreasing residence time. Above the critical Re number, the
oscillations in ow provide an additional mechanism that enhances mixing. Typically an enhancement in the mixing is
accompanied by an increase in pressure drop, which necessitates additional power input. The pressure drop in the post
array is about an order of magnitude higher than that in the
absence of posts (see Figure 11b). Yet, the absolute values of
AIChE Journal

December 2005

Figure 11. Effect of Reynolds number on mixing and

pressure drop.
Above a critical Re number, enhanced mixing performance
is observed because of the onset of oscillatory instabilities.
The conguration used is that depicted in Figure 9b.

pressure drop are less than a few percent of the operating

pressure because of the relatively open post conguration. A
comparison of the two gures indicates a maximum in mixing
enhancement, over that of a reactor without posts, per unit
pressure drop between Rec and 2Rec.
The effect of ow rate on mixing is typically quantied in
terms of the Peclet (Pe) number



where L is the length of the device and D is the molecular

diffusivity (m2/s). Generally, mixing deteriorates rapidly at
large Pe numbers. For the ow conditions corresponding to
Re 100, the Pe number for this device is about 6 103. The
mixing achieved in this device is considerably lower (mix
30%) compared to a value of 80%75 and 90%71 for some of the
best micromixers reported in the literature but comparable to
some of the micromixers reported in Hardt and Schonfeld68 and
Johnson et al.70 The pressure drop in the post micromixer is
comparable to the micromixers reported in Johnson et al.,70
Vol. 51, No. 12


Figure 12. Potential improvements of the post micromixer.

Contours of mass fraction of species A indicating mixing enhancements. (a) Doubling the mixer length to 1.5 cm increases the mixing from
about 26% (Figure 9b) to about 50%. (b) Distributing the feed increases the mixing from about 20% (Figure 7a) to about 64%. [Color gure
can be viewed in the online issue, which is available at www.interscience.wiley.com.]

Strook et al.,71 and Eneld et al.,75but lower than the superfocus micromixer of Hardt and Schonfeld.68
The post micromixers studied here will be well suited for
high-throughput applications. In contrast to the additional
secondary vortices used by most mixers, these mixers rely
on the primary vortices. Given the length scales of mixing
indicated in the previous graphs, a distributive feed over a
millimeter length scale (each substream size should be of the
order of the transverse post spacing or smaller) and an
increased mixer length, will substantially improve the performance of the post micromixers. This is indeed the case as
shown in Figure 12.

In this article, ow-driven instabilities at the microscale
were explored using 2-D CFD simulations. Flow past a single,
cylindrical post at the microscale was found to exhibit the same
nonlinear behavior as that at the macroscale and to be associated with symmetry breaking. Chemical reactions were found
to have an effect on oscillatory instabilities only when a change
in the uid density occurs. Such a density change may arise
when the number of moles is not conserved or under nonisothermal conditions. For isothermal ows the effect of chemistry on oscillations is weak. On the other hand, the observed
reactivity was found to be strongly affected by ow instabilities. As an example, increased conversions were found for
volumetric reactions as a result of enhanced mixing, and reduced conversions were observed for nonproduct-inhibited
positive order surface reactions arising from an effective reduction of the mass transfer coefcient, compared to a symmetric ow, caused by the fast removal of reactants from the
back of a post.
Aside from the ow chemistry interplay, ow-driven oscillations were found to enhance mixing at the microscale, above
a critical Reynolds number, with a relatively small pressure
drop. The performance of the micromixer was quantied in
terms of a mixing parameter at the exit. It appears that inter row
spacing can be tuned to enhance the mixing, and staggering
affects the variance in the exit more than the net mixing.
Potential improvements were suggested. However, further
studies are needed to fully optimize these mixers. Given that
posts are often fabricated for other reasons, such as increased
surface area, one could take advantage of them to enhance
mixing by working in the proper Re number regime. Advantages of this new method of mixing include the ease of fabrication and handling of particulate solutions compared to most
existing micromixer designs, the ease of scalability, and the
ability to handle relatively fast ows.

December 2005

This work was supported by the Army Research Ofce under contract
DAAD19-01-1-0582. Any opinions, ndings, and conclusions or recommendations expressed are those of the authors and do not necessarily reect
the views of the Army Research Ofce.

Literature Cited
1. Jensen KF. Microreaction engineeringIs small better? Chem Eng
Sci. 2001;56:293-303.
2. Ehrfeld W. Design guidelines and manufacturing methods for microreaction devices. Chimia. 2002;56:598-604.
3. Schouten JC, Rebrov EV, de Croon MHJM. Miniaturization of heterogeneous catalytic reactors: Prospects for new developments in catalysis and process engineering. Chimia. 2002;56:627-635.
4. Schwalbe T, Autze V, Wille G. Chemical synthesis in microreactors.
Chimia. 2002;56:636-646.
5. Kursawe A, Dietzsch E, Kah S, Honicke D, Fichtner M, Schubert K,
Weimeier G. Selective reactions in microchannel reactors. Proceedings of the Third International Conference on Microreaction Technology. Berlin/New York: Springer-Verlag; 1999:213-223.
6. Kenis PJA, Ismagilov RF, Whitesides GM. Microfabrication inside
capillaries using multiphase laminar ow patterning. Science. 1999;
7. Min J, Kim J-H, Kim S. Microuidic device for bioanalytical systems.
Biotechnol Bioprocess Eng. 2004;9:100-106.
8. Yun K-S, Yoon E. Microuidic components and bio-reactors for
miniaturized bio-chip applications. Biotechnol Bioprocess Eng. 2004;
9. Makhijani VB, Raghavan J, Przekwas A, Przekwas AJ. Simulation of
biochemical reaction kinetics in microuidic systems. Proceedings of
the Third International Conference on Microreaction Technology.
Berlin/New York: Springer-Verlag; 1999:441-450.
10. Gunther A, Khan SA, Thalmann M, Trachsel F, Jensen KF. Transport
and reaction in microscale segmented gasliquid ow. Lab Chip.
11. Przekwas A, Wang D, Makhijani VB, Przekwas AJ. Microuidic
ltration chip for DNA extraction and concentration. Proceedings of
the Third International Conference on Microreaction Technology.
Berlin/New York: Springer-Verlag; 1999:488-497.
12. Fichtner M, Mayer J, Wolf D, Schubert K. Microstructured rhodium
catalysts for partial oxidation of methane to syngas. Ind Eng Chem
Res. 2001;40:3475-3483.
13. Trubac RE, Dautzenberg FM, Grifn TA, Paikert B, Schmidt VR,
Overbeek RA. Micro-engineered catalyst systems: ABBs advancement in structured catalytic packings. Catal Today. 2001;69:17-24.
14. Lyubovsky M, Karim H, Menacherry P, Boorse S, LaPierre R, Pfefferle WC, Roychoudhury S. Complete and partial catalytic oxidation
of methane over substrates with enhanced transport properties. Catal
Today. 2003;83:183-197.
15. Losey MW, Jackman RJ, Firebaugh SL, Schmidt MA, Jensen KF.
Design and fabrication of microuidic devices for multiphase mixing
and reaction. J Microelectromech Syst. 2002;11:709-717.
16. Ganley JC, Seebauer EG, Masel RI. Porous anodic alumina microreactors for production of hydrogen from ammonia. AIChE J. 2004;50:
17. Norton DG, Wetzel ED, Vlachos DG. Fabrication of single-channel

Vol. 51, No. 12

AIChE Journal


















catalytic microburners: Effect of connement on the oxidation of

hydrogen/air mixtures. Ind Eng Chem Res. 2004;43:4833-4840.
Batchelor GK. An Introduction to Fluid Dynamics. Cambridge, UK:
Cambridge Univ. Press, 1967.
Mavridis C, Bakrozis A, Koutmos PDP. Isothermal and non-premixed
turbulent reacting wake ows past a two-dimensional square cylinder.
Exp Therm Fluid Sci. 1998;17:90-99.
Khinast JG, Koynov AA, Leib TM. Reactive mass transfer at gas
liquid interfaces: Impact of micro-scale uid dynamics on yield and
selectivity of liquid-phase cyclohexane oxidation. Chem Eng Sci.
Schwinge J, Wiley DE, Fletcher DF. Simulation of the ow around
spacer laments between channel walls. 2. Mass-transfer enhancement. Ind Eng Chem Res. 2002;41:4879-4888.
Nguyen HD, Paik S, Douglass RW. Convective transport about cylinder with surface reaction of arbitrary order. AIChE J. 1996;42:15141524.
Roshko A. On the development of turbulent wakes from vortex streets.
NACA Report 1191. National Advisory Committee for Aeronautics
(predecessor organization to NASA) 1954:1-71.
Coutanceau M, Bouard R. Experimental determination of the main
features of the viscous ow in the wake of a circular cylinder in
uniform translation. Part 1. Steady ow. J Fluid Mech. 1977;79:231256.
Mathis C, Provansal M, Boyer L. The Bernardvon Karman instability: An experimental study near the threshold. J Phys Lett. 1984;45:
Wen C-Y, Lin C-Y. Two-dimensional vortex shedding of a circular
cylinder. Phys Fluid. 2001;13:557-560.
Van Atta CW, Gharib M. Ordered and chaotic vortex streets behind
circular cylinders at low Reynolds numbers. J Fluid Mech. 1987;174:
Li J, Chambarel A, Donneaud M, Martin R. Numerical study of
laminar ow past one and two circular cylinders. Comput Fluids.
Park J, Kwon K, Choi H. Numerical solutions of ow past a circular
cylinder at Reynolds numbers up to 160. KSME Int J. 1998;12:12001205.
Gresho PM, Chan ST, Lee RL, Upson CD. A modied nite element
method for solving the time-dependent, incompressible NavierStokes
equations. Part 2: Applications. Int J Numer Methods Fluids. 1984;4:
Braza M, Chassing P, Ha Minh H. Numerical study and physical
analysis of the pressure and velocity elds in the near wake of a
circular cylinder. J Fluid Mech. 1986;165:79-130.
Posdziech O, Grundmann R. Numerical simulation of the ow around
an innitely long circular cylinder in the transition regime. Theor
Comput Fluid Dyn. 2001;15:121-141.
Yang HH, Seymour BR, Shizgal BD. A Chebyshev pseudospectral
multi-domain method for steady ow past a cylinder, up to Re 150.
Comput Fluids. 1994;23:829-851.
Schlichting H, Gersten K. Boundary Layer Theory. 8th Edition. Berlin,
Germany: Springer-Verlag; 2000.
Kawaguti M, Jain P. Numerical study of a viscous uid ow past a
circular cylinder. J Phys Soc Jpn. 1966;21:2055-2062.
Provansal M, Mathis C, Boyer L. Bernardvon Karman instability:
Transient and forced regimes. J Fluid Mech. 1987;182:1-22.
Williamson CHK. Vortex dynamics in the cylinder wake. Ann Rev
Fluid Mech. 1996;28:477-539.
Barkley D, Henderson RD. Three-dimensional Floquet stability analysis of the wake of a circular cylinder. J Fluid Mech. 1996;322:215241.
Jackson CP. A nite-element study of the onset of vortex shedding in
the ow past variously shaped bodies. J Fluid Mech. 1987;182:23-45.
Kiya M, Arie M, Tamura H, Mori H. Vortex shedding from two
circular cylinders in staggered arrangement. J Fluids Eng. 1980;102:
Zdravkovich MM. Review of ow interference between two cylinders
in various arrangements. J Fluids Eng. 1977;99:618-633.
Zdravkovich MM. Flow induced oscillations of two interfering circular cylinders. J Sound Vib. 1985;101:511-521.
Kiya M, Mochizuki O, Ido Y, Suzuki T, Arai T. Flip-opping ow
around two bluff bodies in tandem arrangement. IUTAM Symposium:

AIChE Journal

December 2005



Bluff-Body Wakes, Dynamics and Instabilities. Gottingen, Germany:

Springer-Verlag; 1992:15-18.
Mittal S, Kumar V, Raghuvanshi A. Unsteady incompressible ows
past two cylinders in tandem and staggered arrangements. Int J Numer
Methods Fluids. 1997;25:1315-1344.
Johnson AA, Tezduyar TE, Liou J. Numerical simulation of ows past
periodic arrays of cylinders. Comput Mech. 1993;11:371-383.
Guillaume DW, LaRue JC. Investigation of the opping regime with
two-, three- and four-cylinder arrays. Exp Fluids. 1999;27:145-156.
Lam K, Cheung WC. Phenomena of vortex shedding and ow interference of three cylinders in different equilateral arrangements. J Fluid
Mech. 1988;196:1-26.
Lam K, Li JY, So RMC. Force coefcients and Strouhal numbers of
four cylinders in cross ow. J Fluid Struct. 2003;18:305-324.
Sayers AT. Vortex shedding from groups of three and four equispaced
cylinders situated in a cross ow. J Wind Eng Ind Aero. 1990;34:213221.
Farrant T, Tan M, Price WG. A cell boundary element method applied
to laminar vortex-shedding from arrays of cylinders in various arrangements. J Fluid Struct. 2000;14:375-402.
Young DL, Huang JL, Eldho TI. Simulation of laminar vortex shedding ow past cylinders using a coupled BEM and FEM model.
Comput Method Appl Mech Eng. 2001;190:5975-5998.
Behr M, Hastreiter D, Mittal S, Tezduyar TE. Incompressible ow past
a circular cylinder: Dependence of the computed ow eld on the
location of the lateral boundaries. Comput Method Appl Mech Eng.
Chakraborty J, Verma N, Chhabra RP. Wall effects in oe past a
circular cylinder in a plane channel: A numerical study. Chem Eng
Process. 2004;43:1529-1537.
Fluent, version 6.1.22. Lebanon, NH: Fluent Inc.; 2004.
Deshmukh SR, Mhadeshwar AB, Vlachos DG. Microreactor modeling
for hydrogen production from ammonia decomposition on ruthenium.
Ind Eng Chem Res. 2004;43:2986-2999.
Freitas CJ. Perspective: Selected benchmarks from commercial CFD
codes. J Fluids Eng. 1995;117:208-219.
Noack BR, Eckelmann H. On chaos in wakes. Physica D. 1992;56:
Thomas DG, Kraus KA. Interaction of vortex streets. J Appl Phys.
Campbell CJ, Grzybowski BA. Microuidic mixers: From microfabricated to self-assembling devices. Philos Trans R Soc Lond A Phys
Sci. 2004;362:1069-1086.
Stremler MA, Haselton FR, Aref H. Designing for chaos: Applications
of chaotic advection at the microscale. Philos Trans R Soc Lond A
Phys Sci. 2004;362:1019-1036.
Kim MC, Kim S, Park JS, Park HD. Numerical simulation of micromixing by pulsatile micropump. J Ind Eng Chem. 2003;9:602-606.
Tabeling P, Chabert M, Dodge A, Jullien C, Okkels F. Chaotic mixing
in cross-channel micromixers. Philos Trans R Soc Lond A Phys Sci.
Yang Z, Matsumoto S, Goto H, Matsumoto M, Maeda R. Ultrasonic
micromixer for microuidic systems. Sens Actuators A Phys. 2001;93:
Jen C-P, Wu C-Y, Lin Y-C, Wu C-Y. Design and simulation of the
micromixer with chaotic advection in twisted microchannels. Lab
Chip. 2003;3: 77-81.
Ehrfeld W, Golbig K, Hessel V, Lowe H, Richter T. Characterization
of mixing in micromixers by a test reaction: Single mixing units and
mixer arrays. Ind Eng Chem Res. 1999;38:1975-1082.
Schonfeld F, Hessel V, Hofmann C. An optimized split-and-recombine
micro-mixer with uniform chaotic mixing. Lab Chip. 2004;4:65-69.
Schonfeld F, Hardt S, Hessel V, Hofmann C. Optimized distributive
mixing by chaotic multilamellation. Chem-Ing-Tech. 2004;76:614618.
Hardt S, Schonfeld F. Laminar mixing in different interdigital micromixers: II. Numerical simulations. AIChE J. 2003;49:578-584.
Drese KS. Optimization of interdigital micromixers via analytical
modelingExemplied with the superfocus micromixer. Chem Eng J.
Johnson TJ, Ross D, Locascio LE. Rapid microuidic mixing. Anal
Chem. 2002;74:45-51.
Strook AD, Dertinger SKW, Ajdari A, Mezic I, Stone HA, Whitesides
GM. Chaotic mixer for microchannels. Science. 2002;295:647-651.

Vol. 51, No. 12


72. Schoenfeld F, Hardt S. Simulation of helical ows in micro channels.

AIChE J. 2004;50:771-778.
73. Liu RH, Stremler MA, Sharp KV, Olsen MG, Santiago JG, Adrian RJ,
Aref H, Beebe DJ. Passive mixing in a three-dimensional serpentine
channel. J Microelectromech Syst. 2000;9:190-197.
74. Deshmukh SR, Mhadeshwar AB, Lebedeva MI, Vlachos DG. From
density functional theory to microchemical device homogenization:
Model prediction of hydrogen production for portable fuel cells. Int J
Multiscale Comput Eng. 2004;2:221-238.
75. Eneld KE, Siekas JJ, Pence DV. Laminate mixing in micro-scale
fractal-like merging channel networks. Proc of First International
Conference on Microchannels and Minichannels, Rochester, NY,
ASME; 2003:859-867.
76. Fey U, Konig M, Eckelmann H. A new StrouhalReynolds-number
relationship for the circular cylinder in the range 47 Re 2 105.
Phys Fluids. 1998;10:1547-1549.
77. Williamson CHK. 2-D and 3-D aspects of the wake of a cylinder, and
their relation to the wake computations. Lect Appl Math. 1991;28:719751.
78. Friehe C. Vortex-shedding from circular cylinders at low Reynolds
numbers. J Fluid Mech. 1980;100:237.
79. Gerrard J. The wakes of cylindrical bluff bodies at low Reynolds
numbers. Philos Trans R Soc Lond A Phys Sci. 1978;A288:351.
80. Berger E, Wille R. Periodic ow phenomena. Ann Rev Fluid Mech.

Appendix: Validation for Nonreactive Laminar

Flow Past a Single Bluff Body
The laminar ow past a cylinder of diameter 200 m, a
benchmark problem to test our numerical work, is simulated at
various Reynolds numbers. It is observed that oscillatory ow
patterns, similar to those at the macroscale, develop in microdevices. Our results are compared with previous studies in
Figure A1 in terms of the Strouhal number. The St vs. Re
number relationship at the microscale obtained from our calculations is similar to that seen at the macroscale in the experiments of Roshko23 and Fey et al.76 At Re 100, corresponding to a uid free stream velocity of 23.4 m/s, the St number
obtained with the rst-order upwind scheme is 0.159. A value
of St 0.164 was recently proposed based on the universal
St vs. Re curve.77 Using higher-order discretization schemes
(second-order and QUICK), the value of St number calculated
was 0.169. Using an adaptive meshing strategy, that is, renement in the grid along the wake, also gave results slightly
different from 0.164. Even though it has been suggested that
the experimentally obtained universal St vs. Re curve is accurate to within 1%, signicant differences, by as much as 15%,
have been reported in the literature among simulations.77 The
source of this deviation is still a topic of discussion and several
reasons have been suggested, such as different boundary con-


December 2005

Figure A1. Comparison of Strouhal number vs. Reynolds

number for ow past a single micropost with
available literature data at the macroscale.
[Color gure can be viewed in the online issue, which is
available at www.interscience.wiley.com.]

ditions, mesh and domain sizes, and the possibility of solution

multiplicity of the NavierStokes equation. The boundary conditions used in this work are representative of the experimental
conditions for a typical post microreactor. Our values of St
obtained using different numerical schemes are within about
3% of the universal value. Typical CPU times for 100 iterations
using the rst-order, second-order, and QUICK schemes are
3704, 5921, and 6082 s, respectively. Given the large number
of time steps required for a solution and nearly the same
deviation from the value of 0.164, use of rst-order upwind
scheme is a reasonable trade-off between solution accuracy and
computational effort. Our value of St 0.16 using the rstorder upwind scheme lies in the range of values for St (0.15
0.18) found by either experiments78-80 or simulations,28,30,57
indicating that only the dimensionless groups, and not the
actual device dimensions, play a role. This is an expected result
given that the continuum approximation is still valid at these
scales. The transition from a steady wake to a periodic unsteady one at the critical Re number is characterized by a Hopf
bifurcation. Our results for nonreactive ow simulations indicate that the StuartLandau equation is satised near the critical Re number (not shown), as reported in Provansal et al.36
Manuscript received Dec. 2, 2004, and revision received Apr. 28, 2005.

Vol. 51, No. 12

AIChE Journal

Mixing Effects on Performance and Stability of

Low-Density Polyethylene Reactors
Gary J. Wells and W. Harmon Ray
Dept. of Chemical and Biological Engineering, University of Wisconsin-Madison, Madison, WI 53706
DOI 10.1002/aic.10544
Published online August 31, 2005 in Wiley InterScience (www.interscience.wiley.com).

Imperfect initiator mixing greatly affects the stability and efciency of low density
polyethylene (LDPE) autoclave reactors. A combined simulation technique utilizing
compartment models and computational uid dynamics extends previous work in the
literature by providing a physically detailed picture of imperfect mixing. Analysis indicates that the effective volume for chain propagation in the autoclave reactor can expand
and contract in a continuous fashion as operating conditions change. As mixing becomes
poor, the effective reactive volume decreases, causing a reduction in initiator efciency,
but an expansion in the stable operation region. Examples demonstrate that accurate
prediction of the effective reaction volume is crucial for predicting LDPE autoclave
reactor behavior. A new mixing model that represents the feed plume by a series of
interconnected tanks with geometrically increasing volumes provides a favorable tradeoff
between accuracy and model complexity. 2005 American Institute of Chemical Engineers
AIChE J, 51: 32053218, 2005

Keywords: mixing, stability, computational uid dynamics, low-density polyethylene,

autoclave reactor

Low-density polyethylene (LDPE) accounts for about a
quarter of U.S. polyethylene production, and is an important
commodity polymer used in plastic bags, lms and coatings.1
The physics involved in the production of LDPE can be quite
complex, and the reaction kinetics, polymer properties, and
reactor stability in LDPE processes have, therefore, been studied for over 50 years.23
LDPE processes employ two types of reactors: (1) tubular
reactors, or (2) stirred autoclave reactors, and each type of
process is used extensively for LDPE production.4 Each of
these processes operates at high pressures (1,500 3,000 atmospheres), and with reactor exit temperatures ranging from about
200 to 300C.
In this work, we model a single compact type autoclave
reactor, such as the DuPont type (c.f. Figure 1). Typical processes have reactor volumes ranging between 250 and 2,000 L
and residence times on the order of 10 s to one min. The heat
Correspondence concerning this article should be addressed to W. H. Ray at

AIChE Journal

December 2005

generated by polymerization is quite high, and causes a temperature rise of about 12C13C for each percent of monomer
conversion.5 Due to relatively short residence times and thick
reactor walls, the capability of removing heat through wall
cooling is quite low.6 Thus, autoclave reactors can be considered adiabatic, and the exit temperature is attenuated by feeding the monomer in a cold condition.7 Due to safety limits on
the maximum temperature rise and the low residence times,
single pass conversion in autoclave reactors is usually below
20%. Despite this relatively low conversion to polymer, stability issues can be a major operational concern.
At temperatures above about 300C, a violently exothermic
reaction can decompose ethylene, leading to localized dark
spots in the polymer or global reactor runaway. Such ethylene
decomposition reactions are common in LDPE plant operation,
and occurrences of this reaction mechanism are known in the
industry as a decomp. Decomp episodes can be quite costly
due to environmental and safety concerns, and can ultimately
lead to signicant process downtimes. LDPE autoclaves can
also lose stability and migrate away from the desirable operating branch if internal temperatures drop too low and ignition
Vol. 51, No. 12


Figure 1. LDPE autoclave reactor.

of the chemical initiator is lost. This behavior leads to reactor

extinction whereby the polymerization ceases. Such reactor
extinction is a common industrial occurrence that causes signicant production losses.
To achieve desired polymer properties and avoid extinction
or runaway of the autoclave reactor, feed initiator fraction is
manipulated to control a desired internal temperature. Despite
the use of such controllers and high amounts of agitation
energy input to the vessel, reaction conditions may not be well
mixed due to the similarity of reaction and mixing time scales.8
In two studies led by van der Molen,9,10 the efciency of
numerous peroxide initiators for producing LDPE in a stirred
autoclave reactor was studied. It was determined that the specic initiator consumption per mass of polymer produced depended heavily on the type and intensity of mixing used in the
reactor. Also, it was found that plotting this specic initiator
consumption versus reactor temperature produced a minimum
in the curve. This minimum was not represented using conventional kinetic modeling in a perfectly-mixed tank reactor, but it
was shown that the minimum could be could be explained by
imperfect mixing effects.10
This evidence of imperfect mixing effects in LDPE autoclave reactors led to further simulation and experimental research. Georgakis and Marini1113 used a three-compartment
model to represent imperfect mixing of the feed stream. This
model represented the feed zone by two small perfectly mixed
tanks connected in series. These feed compartments were connected by ows to and from a larger perfectly mixed compart3206

December 2005

ment. The major parameter of the model was the recycle rate
from the large zone to the feed compartments. This recycle rate
parameter was chosen to match common reactor measurements. Using this three-compartment model, Marini and Georgakis1213 computed conversion and polymer properties. They
also computed the largest real eigenvalue for the reactor model,
and found that the range of operation over which this eigenvalue was negative was expanded with imperfect mixing.
Villa et al.14 carried out a full stability analysis of the LDPE
autoclave including the effects of the decomp reactions. They
showed that the perfectly mixed reactor has a tiny (or sometimes no) range of operating conditions for stable operation
where polymer is produced. However, by using simple threecompartment mixing models, they predicted a broad range of
productive, stable operation when there is imperfect mixing.
Villa et al.14 provide a clear physical explanation for the
stabilizing effect of imperfect mixing. Their study demonstrated that small regions in the feed plume of the reactor
signicantly reduce initiator and radical concentrations that
reach larger downstream zones. This causes a net reduction in
initiator efciency when compared to perfect mixing, and consequently reduces heat release and reactor sensitivity to temperature. Villa et al. showed that this reduction in temperature
sensitivity is analogous to a stabilizing controller that reduces
feed initiator concentration to the downstream zones when
operating temperature increases. Subsequent studies of imperfect mixing in the LDPE autoclave have also assessed aspects
of reactor operation and stability.8,15
Other research has examined more complex, and presumably
more realistic, mixing models in the context of process efciency and product properties.16 20 These studies have utilized
compartment mixing models of varying levels of complexity
and backow characteristics. Such models have been used to
predict industrial plant data in certain operating regions, but
little guidance exists concerning the appropriate level of detail
required to accurately capture the effects of mixing on LDPE
reactor operation.
Recently, computational uid dynamics (CFD) has been
used to provide an increased level of mixing detail in models of
autoclave and tubular LDPE reactors.2126 Such CFD studies
showed that mixing in autoclaves can play a very signicant
role in regions near initiator feeds, where large gradients in
temperature and composition exist.2124 These gradients can
subsequently affect process productivity and efciency. However, due to limitations in computing power, such detailed CFD
models have not been used to examine mixing effects on LDPE
autoclave operation over wide ranges in feed conditions or
design parameters. Furthermore, stability and polymer property
issues have not been studied thus far with the level of mixing
detail provided by CFD. It is envisioned that CFD will aid in
understanding the underlying physics that govern mixing effects on the operation of the LDPE autoclave.
This article applies a modeling approach that combines the
computational simplicity of compartment models with the mixing detail of CFD simulation. The approach models detailed
mixing effects on stability and productivity in the adiabatic
LDPE autoclave over wide operating ranges. General effects
governing spatial variations and stability are rst established in
the relatively simple three-compartment model used by Villa et
al.14 and then these simple models are compared to the more
detailed approach involving CFD. The analysis provides guideVol. 51, No. 12

AIChE Journal

Table 1. Kinetic Mechanism for Free-Radical

Homopolymerization of Ethylene

Table 3. Constant Values Specied in this Work for

Modeling the LDPE Autoclave Reactor*

feff kd
2R decomposition of initiator)


initiation of growing chain

Chain Propagation:
Termination by combination:



Inlet ow velocity
Residence time
Impeller rotation rate
Reactor volume
Mixture density
Heat capacity
Species diffusivity
Heat conductivity
Mixture viscosity
Initiator efciency
Reactor pressure
Heat of polymerization
Heat of decomposition

50 m/s
32.8 s
200 RPM
500 L
499 g/L
2.768 J/g-K
2.88 10-5 m2/s
0.1998 J/m-s-K
1.6 103 kg/m-s
2000 atm
21386 cal/mol
30200 cal/mol

*Note: Values for physical constants are taken from Read et al.22

Pn Pm O
Ethylene decomposition:*
Cmon1.89kmd1 kpd1 0.0714kpd2
decomp products
30200 cal/mol
*Note: Ethylene decomposition (decomp) is modeled as a heat release only;
decomp products are not explicitly tracked. This simplied form of the decomp
model proposed by Zhang et al.27 relates the rate of heat generation from
monomer decomp to lumped kinetic parameters (k md, k pd1, k pd2) from the
decomposition mechanism as follows28:

of Zhang et al.27 is also applied to simulate the heat released as

ethylene decomposes at high temperature.28
This decomp modeling approach allows reactor behavior to
be accurately assessed at the upper temperature stability limit.
Table 3 shows the process parameters that are kept constant
for this work unless specically stated otherwise. The major
parameters studied in this work are the feed temperature, feed
initiator fraction, and reactor residence time. The initiator used
is tert-butyl peroxyacetate (TBPOA), a relatively active initiator with a half life of about 2 s at 200C. The half life of this
initiator decreases signicantly at higher temperatures.

Mixing Effects in Simple Compartment Models of

the LDPE Autoclave Reactor

R heat,decomp HdecompCmon
1.89kmd kpd1 0.0714kpd2Cmon

lines for accurately modeling the reactor over wide ranges of

operating conditions, and specic recommendations are made
for assessing mixing effects in industrial reactor congurations.

Reaction Kinetics
Table 1 presents the reactions comprising the kinetic mechanism used in this work for modeling free-radical ethylene
homopolymerization in the high-pressure autoclave reactor. As
can be seen, the common free radical initiation, propagation,
and termination steps are modeled. In this work, a single
chemical initiator is used, and termination is presumed to occur
primarily by coupling in accordance with experimental evidence.29
Table 2 presents the rate constant parameters used in this
study to dene reactive source terms. The rate constants are
assumed to be of the Arrhenius form with both temperature and
pressure dependencies. A simplied form of the decomp model

In the study of Villa et al.,14 a three-compartment model (see

Figure 2) was used to show that imperfect mixing expands the
stable operating region of LDPE autoclaves considerably due
to reduced initiator efciency. Figure 3 illustrates these results,
and also shows that at low-feed temperatures, imperfect mixing
can provide a desirable stable operating branch when none
exists for perfect mixing. In the gure, continuation diagrams
compare the stability characteristics of perfectly mixed and
imperfectly mixed models for an adiabatic reactor operating at
four different feed temperatures. A residence time of 32.8 s is
used, which falls in the range of industrial practice. In Figures
3a b (feed temperatures of 420 and 400 K), one sees that stable
operation is possible under both perfect mixing and imperfect
mixing conditions. Although the stability region for imperfect
mixing is considerably wider due to the lowered initiator efciency, stable ignition is even possible in the perfectly-mixed
case at the 32.8 s residence time. Further, at such a high-feed
temperature, stable ignition is attainable for very low initiator

Table 2. Kinetic Parameters Used to Study the LDPE Autoclave*




E A (cal/mol)

V A (cal/atm-mol)


Propagation (k p )
Initiator decomposition (k d )
Termination by coupling (k tc )
Initial monomer decomp. (k md )
Decomp. propagation 1 (k pd1 )
Decomp. propagation 2 (k pd2 )


1.14 107
1.06 1016
3.00 109
4.00 1019
1.59 1020
4.39 1020




*Rate constants (k) are computed using the formula: k k 0 exp[(E A V A P)/R gasT], where P is the absolute pressure in atm, T is the absolute temperature
in K and R gas is the ideal gas constant.

AIChE Journal

December 2005

Vol. 51, No. 12


Figure 2. Three-compartment imperfect mixing used by

Villa et al.14
Base parameters used in this work are taken from Zhang and
Ray32: V 1 /V tot 1/ 24, V 2 /V tot 2/ 24, f 1 1/3. The
recycle ratio R is varied.

inuence the internal ows, the recycle ratio (R) is varied.

Higher values of R indicate more vigorous internal ows and
reduce the local ow time scales in an approximately linear
fashion. Lower values of R indicate correspondingly less vigorous internal ow. Values of R between about 2 and 20 have
been used to t measured data for industrial LDPE autoclave
For adiabatic reactors models such as that considered in this
section, the initiator half life is dependent on recycle ratio (or
local compartment residence times) through its inuence on
local temperatures. Since initiator dispersion is the primary
factor that determines the degree of mixing in LDPE autoclave
reactors, such temperature variations have the potential to
provide much more complicated behavior than observed under
isothermal conditions. Indeed, it is shown in this section that
increasing recycle ratio does not always decrease spatial variation in radical concentration or polymerization reaction environment within the range of stable operation. This nonintuitive
behavior is explained using the simulation examples presented
Figures 4a through 4d show the effect of recycle ratio on
several important quantities for the standard adiabatic threecompartment model with a 300 K feedstream containing 50
ppm of the TBPOA initiator. Under these conditions, there are
both upper and lower limits for stability as recycle ratio (R) is
varied. At values of R 5, the reactor will extinguish because
insufcient heat is mixed back into the feed zone to achieve a
stable amount of free radical generation. At values of R greater
than about 210, the entire reactor will reach the conditions of
global monomer decomposition. In between these stability
limits, signicant changes in internal conditions and kinetic
rates occur. As recycle ratio is increased from the lower stability limit to the upper stability limit, the temperature and bulk

feed fractions because the feed stream itself has enough energy
to support ignition of the entire reactor volume.
At the lower feed temperatures of 360 K and 300 K (Figure
3c d), imperfect mixing provides a desirable stable operating
branch, but no such stable conditions exist for the perfectlymixed model using the TBPOA initiator at the current residence time. For these lower feed temperatures, a signicant
temperature rise in the partially isolated feed zone must occur
in tandem with backmixing of hot material to decompose the
initiator and ignite the reactor. These conditions are provided
by the three-compartment imperfect mixing model, but not by
the perfectly-mixed model.
It should be noted that low feed temperatures in the range of
about 293363 K (20 90C) are common industrially,33 because they allow for greater adiabatic temperature rise and
higher monomer conversion before the dangerous decompositions reactions become important. Thus, the results of this section
emphasize the necessity of including mixing effects in models of
industrial LDPE autoclave reactors, since perfectly-mixed models
may not predict any desirable stable branch for industrially relevant values of feed temperature and residence time.

Effect of internal ows in adiabatic reactor operation

In this section, the effects of internal ow rates on the spatial
uniformity and efciency of the LDPE autoclave reactor are
examined for the three-compartment model of Figure 2. To

December 2005

Figure 3. Stability diagrams comparing perfectly mixed

and imperfectly mixed models.
A residence time of 32.8 s is used. Plots: (a) T f 420 K, (b)
T f 400 K, (c) T f 360 K, (d) T f 300 K.

Vol. 51, No. 12

AIChE Journal

Figure 4(a d). Effect of recycle ratio on spatial variations of temperature and composition.
An adiabatic, 3-compartment model is used with feed conditions of 300 K and 50 ppm of the TBPOA initiator.

polymer fraction both become more uniform in space. Two

interrelated effects work to increase uniformity of bulk polymer fraction and temperature with increasing R: (1) increased
back-mixing of hot material raises the feed zone temperature,
such that it approaches the temperature of downstream zones,
and (2) the feed zone produces an increasing amount of polymer/heat that is passed on to the downstream zones.
In contrast to what is seen for polymer content and temperature, Figure 4c shows that radical concentrations vary in the
three zones by over an order of magnitude despite increased
internal ows. As the temperature of zone 1 increases due to
increased back-mixing, the initiator half-life decreases in an
exponential fashion. In contrast, the ow-time scale decreases
with R more slowly, in an approximately linear fashion. As a
result, the initiator is consumed to a greater extent in the feed
compartments as R is increased (see Figure 5a), and signicant
spatial variations in initiator and radical concentrations result.
Furthermore, as the recycle ratio is increased, the higher temperatures and radical concentrations in zone 1 cause it to
produce a much larger portion of the total conversion than it
would under perfectly mixed conditions (see Figure 5b).
Clearly, if very high recycle ratios were attainable on the stable
operating branch, the radical concentrations in the reactor
would eventually become spatially uniform. However, in the
stable range of operation, spatial uniformity of radicals is not
improved with stirring rate due to the inuence of temperature
on initiator half-life.
As shown in subsequent sections, operating conditions inuence reactor uniformity and efciency much more than
AIChE Journal

December 2005

internal ows. Therefore, design parameters such as feed initiator fraction, feed temperature, and residence time are the
focus of highly-detailed mixing model studies in the subse-

Figure 5(a b). Effect of recycle ratio on fractional conversion and initiator consumption in each
The feedstream is introduced at 300 K with 50 ppm of
the TBPOA initiator.

Vol. 51, No. 12


Figure 6. Procedure for selecting and using a compartment model of the LDPE autoclave reactor.

quent sections. Additional details on the effect of internal ows

on reactor performance may be found elsewhere.28

Detailed Modeling of Mixing and Reaction in the

LDPE Autoclave
Because spatial variations of the active radical species and
initiator are difcult to measure in the LDPE autoclave, experiments cannot easily provide a visual picture of how these
species are mixed. Simple compartment models have shown
that the spatial variations of these quantities are extremely
important in determining reactor operation.1214,22,34 Although
simple compartment models can be used to match industrial
measurements, there is no guarantee that simple models accurately capture mixing details under all conditions, especially
since the initiator efciency or kinetic rate constants are often
varied to t the data.18,35 If mixing details are not properly
represented, such simple models may not extrapolate well to
conditions outside of the experimental range. Furthermore, as
discussed by Smit,36 any model that captures some degree of
incomplete mixing can represent experimental measurements
such as the nonideal behavior of initiator consumption under
imperfect mixing. Therefore, in order to properly model the
physics of reactor mixing, it is desirable to know how much
mixing detail is necessary to accurately represent LDPE autoclave reactor behavior over a wide operating range.

CFD simulation and equivalent compartment model

This work assesses the appropriate level of mixing detail for
accurate modeling of the LDPE autoclave by reducing a CFD
model with a great deal of mixing detail to obtain an equivalent, simpler compartment model. The CFD/compartment
model approach involves a precise representation of spatially
varying temperature and compositions elds, and can be applied generally to many processes. Specic details of this
methodology are provided elsewhere,37 but the basic concept
and application in the case of LDPE autoclave reactor modeling is presented in Figure 6.
The combined CFD/compartment model approach starts
from a single, converged reactive CFD simulation. In the
current case, the LDPE autoclave reactor is modeled using the
Fluent CFD package,38 and a reactor design from Read et al.22

December 2005

that is chosen to match typical industrial autoclaves. Due to the

high level of agitation common in LDPE autoclaves, the ow
is modeled as fully turbulent. Constant density and viscosity
are assumed since the conversion is about 20% or less, and
these physical properties in the supercritical mixture are similar
to those of liquid water.16 The CFD mesh is shown in the upper
lefthand corner of Figure 6, and involves about 100,000 grid
cells to represent the 500 L reactor volume. Wherever possible,
model parameters such as vessel geometry, reaction kinetics,
and feed conditions are taken from the literature. CFD results
for initiator consumption, exit conversion, and adiabatic temperature rise have been conrmed to be in the operable range
for industrial reactors.4,6,10
Once a single reactive CFD simulation is converged at a base
set of operating conditions, an equivalent compartment model
is selected to represent the reactive ow eld in a simplied
way. The 100,000 grid cells in the base CFD simulation are
subdivided into numerous grid cell groupings in order to make
the zones represented by the groupings uniform in temperature,
composition, and the important reaction rates such as chain
propagation and termination. Each of these grid cell groupings
is represented as a perfectly-mixed reactor in a compartment
model of interconnected tanks. The zone volumes are taken to
be the sum of the grid cell volumes that comprise that zone, and
the interconnecting ows are also taken from the base CFD
simulation. The compartments that result from this subdivision
procedure can be of various shapes and sizes, and they generally follow the composition and temperature gradients that
exist in the reactor. The ows between compartments also
mimic the dominant circulation patterns determined by the
CFD simulation.
The procedure that constructs the compartment model from
CFD output les has been automated and can handle an arbitrary number of compartments. Although stability of a singlestage LDPE autoclave reactor has been considered here, the
methodology can be also be applied to more complex vessels
such as multi-stage autoclaves. The specic criteria used in the
compartment model selection may be found in Wells28 and
Wells and Ray.37 Typically, about 100 perfectly-mixed tanks
are needed to represent the temperature and composition gradients that exist in the reactor. A single set of compartment
model selection criteria can be applied to accurately represent
CFD results over a wide operating range, and the computational simplicity of the selected compartment models allows
them to be computed up to 1,000 times faster than the full CFD
simulations. As a result, the compartment models can be used
in stability analysis or other computational techniques where
full CFD simulations would be difcult or impossible to apply.
In subsequent sections, the compartment model/CFD approach
provides a detailed view of LDPE autoclave reactor mixing and
its effects on reactor performance and stability.

Quantication of effective volume usage and initiator

This section denes numerical quantities that will be used
throughout this article to compactly represent the degree of
vessel mixing and reactor efciency. These values are not
model parameters, but are instead quantities used to assess
mixing thoroughness based on converged simulation results.
Vol. 51, No. 12

AIChE Journal

Figure 7. Mixing effects on initiator efciency.

A feed temperature of 360 K and a residence time of 32.8 s are
used. Contours are shown on a plane that cuts through the 3-D
reactor volume by intersecting the inlet pipe, outlet pipe,
impeller shaft, and impeller blades. The effective reactor
volume fraction is shown for 0.05. The base CFD case for
the 100-compartment model is for a feed of 360 K and 120
ppm TBPOA.

One such measure for quantifying the effective reactor volume

can be dened as follows


Figure 8. Effect of initiator feed fraction on reactor exit

temperature and stability range.
T f 420 K, residence time 32.8 s. The effective reactor
volume fraction is shown for 0.05. The base CFD case for
generating the 100-compartment model is the circled point
with a feed stream at 420 K and 120 ppm of TBPOA.


f rV





n zones

Mixing effects on initiator consumption


Here, H is the unit step function, and f rV ( ) is the effective

reactor volume fraction, that is, the fraction of the reactor
volume actually used for reaction dependent on the choice of
use parameter . Values of in the range from 0.05 to 0.10
seem reasonable. The value V i represents the volume of a
particular zone i measured in liters, and R p,i is the polymerization rate (mol/L-s) in that zone. Zones can be either a
perfectly-mixed tank in a compartment model or an individual
grid cell in a CFD simulation. R p,max is the maximum polymerization rate that occurs throughout all reactor zones. Thus,
for 0.05, the value of f rV ( ) is the fraction of the reactor
volume that is utilized for producing polymer at a rate greater
than 5% of the maximum. Generally, values of f rV close to one
indicate good reactor mixing. Conversely, values of the measure near zero indicate very poor mixing. For perfectly-mixed
reactors, f rV will always be unity, although a value f rV 1 can
also occur in reactors with a small degree of imperfect mixing.
In addition to the reaction volume effectiveness measure
dened earlier, the classic measure of initiator consumption
reported by van der Molen et al.10 can be used to provide
signicant insight into imperfect mixing effects

10 6 w I,f
x p MW I

Figure 7 shows the specic initiator consumption curve of

the form used by van der Molen and coworkers to describe
initiator efciency.9 10 The points along the curve represent
different amounts of initiator in the feed, and radical concentration contours in the reactor are presented at two of these
points. The contour graphs are shown for a cutting plane in the


Here, w I,f is the initiator feed fraction, x p is the monomer

conversion, and MW I is the initiator molecular weight. The
value of I is the amount of initiator required to produce
polymer, and is typically dened as earlier in units of mmol of
AIChE Journal

initiator per kg of polymer produced. This quantity can be

considered as an inverse measure of initiator efciency.

December 2005

Figure 9. Effect of initiator feed fraction on reactor exit

temperature and stability range.
T f 300 K, residence time 32.8 s. The effective reactor
volume fraction is shown for 0.05. The base CFD case for
generating the 100-compartment model involves a feed
stream at 420 K with 120 ppm of TBPOA.

Vol. 51, No. 12


reactor that intersects that feed and exit pipes, as well as the
impeller and stirring blades. From continuation and stability
analysis of the compartment model, the region of stable operation is also determined.
As in the studies of van der Molen et al.,9 10 the current
analysis shows that the initiator consumption has a minimum at
an intermediate internal operating temperature. This behavior
can be understood by examining the two radical contours. At
the higher amount of initiator in the feed (120 ppm), the total
polymer production is also higher, and the adiabatic reactor
exit temperature is about 535 K. At this temperature, the
initiator half-life is short compared to ow timescales, and the
initiator is poorly dispersed. This poor initiator dispersion can
be seen from the radical concentration prole shown for the
high initiator feed fraction. Due to the poor initiator dispersion,
the radicals are concentrated near the feed region, and the
bimolecular termination reaction destroys many radicals before
they can add signicant monomer to form polymer. Also, only
about 1/4 of the reaction volume is effectively used in the
polymerization reaction. This causes the initiator consumption
per kilogram polymer produced to be quite high, increasing
production costs.
Conversely, at a much lower initiator feed fraction of 30
ppm, the operating temperature drops considerably to about
500 K. At this temperature, the initiator half-life is similar to
ow time scales, and initiator is dispersed relatively evenly. As
a consequence, radicals are present in most of the reactor, the
bimolecular termination reaction is less dominant, and the
reactor volume is used effectively in polymerization. This
behavior is consistent with the better initiator usage, meaning
that less initiator is used per kilogram of polymer produced.
Unfortunately, as can be seen in Figure 7, the conditions for the
best radical dispersion and lowest specic initiator consumption are also near the reactor extinction limit. As a result,
reactor operation requires a tradeoff between efciency of
initiator usage and robust process stability.

Effect of feed conditions on mixing and stability

The visual picture of mixing and its relationship to stability
is further explored in Figure 8. In this gure, continuation
diagrams show the effect of initiator feed fraction on exit
temperature for different feed temperatures. Several models
with varying levels of mixing detail are compared. The stability
limit for global reactor decomp is shown along with radical
concentration proles for selected points along the curves.
CFD simulation results (discrete points) are presented for all
operation conditions that would converge in Fluent to a stable
solution at the given feed temperature. Thus, the discrete CFD
results represent the stability range obtained using the fullydetailed model, and these results can be used for comparison to
the stability range predicted by other models.
As can be seen in Figure 8, there is only one stability limit
point at high temperatures due to monomer decomposition for
the feed temperature of 420 K. For this feed temperature and
the TBPOA initiator, the feed stream itself is hot enough to
maintain ignition without any signicant back-mixing. Thus, a
region of stable operation exists with reasonable conversion for
each model type; however for perfectly mixed reactors, the
stable region is too narrow to be practical, while stable reactor
operation is feasible for imperfect mixing. Note that, as ex3212

December 2005

pected, the 100-compartment model more closely predicts the

exact range of stable operating conditions (from CFD) than the
three-compartment model with untted parameters. Also the
reasonable t to the CFD results observed for the 100-compartment model in Figure 8 is related to initiator dispersion and
wastage. In the case of low initiator feed concentrations (10
ppm), the feed zone reaction rates and temperature are low,
resulting in long initiator half-lives. This allows time to disperse the initiator and radicals throughout the reactor. Thus, the
volume usage and mixing of the reactor is quite good at low
initiator feed rates and operating temperatures, and all models
yield similar results. As more initiator is fed, the total polymerization rate and operating temperature increase, initiator
half-life decreases, and radical dispersion worsens. At 20 ppm
of initiator feed, for instance, only about half of the reactor
contains radical concentrations signicant enough to produce
reasonable amounts of polymer. For 180 ppm of initiator in the
feedstream, higher feed zone reaction rates and temperatures
cause the initiator and radicals to be even more concentrated in
the feed zone, reducing effective volume usage to below 10%.
The increase in initiator wastage that is observed with increasing initiator feed is a signicant stabilizing effect, similar
to that seen previously. These results show that the exit temperatures of the more detailed models are less sensitive to
initiator feed fraction than the simpler models. For this reason,
the predicted stable region is much broader in the detailed
models. At the highest initiator feed fractions, even the 100compartment model cannot match the CFD simulations precisely. This disparity results from the base CFD case (at 120
ppm of feed initiator) from which the compartment model was
generated. This particular base case does not represent the
poorest mixing conditions well because compartment volumes
near the feed are not small enough to represent the very high
gradients in radical concentration present under those poorly
mixed conditions. Using additional base CFD simulations at
high feed initiator concentrations can bring the resulting 100compartment models much closer to the CFD results in this
range of initiator levels. Nevertheless, the compartment model
based on only one CFD case performs well in the range of
practical operation, since exit temperatures exceeding about
555 K are generally avoided industrially due to the threat of
Figure 9 shows the analogous stability diagram for a 300 K
feed stream. The lower feed temperature introduces an additional low temperature stability limit point at reactor extinction.
Therefore, at the lower feed temperature of 300 K, perfect
mixing models do not yield a desirable, stable operating branch
for the given residence time. For imperfect mixing, the untted
three-compartment model indicates a desirable, stable operating region, but the range is not as broad as that predicted by the
CFD model. The 100-compartment model, however, closely
matches the range of stable operating conditions observed from
the CFD simulations at exit temperatures below about 555 K
(282C), a reasonable upper limit for autoclave reactor operation that avoids the threat of monomer decomp. As before,
using additional CFD base cases at higher initiator feed concentrations will improve the resulting 100-compartment models in this range.
We have shown that feed initiator fraction has a very signicant effect on adiabatic reactor operation and degree of
mixing. In addition, Figure 10 shows that feed temperature also
Vol. 51, No. 12

AIChE Journal

Figure 11. Effect of residence time on reactor exit temperature and stability range.
T f 360 K, w I,f 60 ppm TBPOA. The effective reactor
volume fraction is shown for 0.05. The base CFD case
for generating the 100-compartment model involves a feed
stream at 420 K with 120 ppm of TBPOA.

Figure 10. Effect of feed temperature on radical concentration proles.

The initiator feed fraction is 80 ppm TBPOA and the residence time is 32.8 s. The effective reactor volume fraction is
shown for 0.05.

leading to efcient volume usage. As residence time increases

from 16.4 s to about 100 s, more reaction is allowed to take
place and the exit temperature and monomer conversion increase accordingly. However, at residence times greater than
about 100 s, the exit temperature becomes relatively insensitive
to further residence time increases. This occurs because mixing
becomes extremely poor at high residence times, and longer
residence times only expand the unused portion of the reactor.
This conclusion is supported by the radical concentration pro-

can signicantly affect radical dispersion at a xed initiator

feed fraction (80 ppm). Higher feed temperatures signicantly
increase the initiator decomposition rate in the feed zone. The
resulting decreased initiator half life leads to poorer initiator
and radical dispersion and less efcient volume usage. Thus,
from the standpoint of radical dispersion, the effect of increasing feed temperature at xed initiator feed is similar to that
seen when initiator feed is increased at a xed feed temperature.

Effect of reactor residence time

Figure 11 shows the effect of reactor residence time on
radical dispersion and reactor stability for a feed temperature of
360 K, and a feed initiator fraction of 60 ppm TBPOA. The
perfectly mixed reactor has no stable steady state except at
near-zero conversion or global reactor decomposition. The
untted three-compartment model predicts a relatively narrow
region of stability compared to the more exact CFD results;
however, the 100-compartment model technique represents the
stable region quite well.
The radical concentration proles illustrate that residence
time plays a large role in radical location. At the lowest
residence time (16.4 s), the exit temperature is such that the
initiator and radicals can disperse well throughout the reactor,
AIChE Journal

December 2005

Figure 12. Fraction of unused volume as a function of

initiator feed fraction.
The feed temperature is 300 K, and the reactor residence
time is 32.8 s. The base CFD case for generating the 100compartment model involves the circled point for a feed
stream at 300 K with 120 ppm of TBPOA.

Vol. 51, No. 12


les, which show only a very small volume being used for
reaction at the highest residence time (328 s).

Requirements for accurate LDPE autoclave reactor

models over wide operating ranges
To accurately capture the stability region for LDPE autoclave reactor operation, a model must be able to predict the
correct heat generation over the operating range of interest. As
was shown in the previous sections, heat generation is largely
dependent on the effective volume for polymerization. As
initiator feed rate, feed temperature, or residence time is increased, a decrease in effective volume is a signicant stabilizing force. A three-compartment model can vary the effective
volume over a fairly large range by allowing temperature to
vary spatially and by distributing radicals amongst the zones.
However, for a xed selection of the three-compartment model
parameters, the effective volume cannot drop below the smallest compartment volume. Furthermore, once a three-compartment model is parameterized, the volumes in which signicant
reaction occurs can only change very discretely. This situation
can be improved by selecting new three-compartment model
parameters for different conditions, but if prediction of unknown operating conditions is desired, it is difcult to accurately select these parameters.
In contrast, the 100-compartment/CFD modeling technique
can vary the volumes of reaction in a nearly continuous manner, and this volume variation occurs automatically as the
relative time scales for uid motion and reaction change. Thus,
tting the observed physical reaction volume requires no
change in model parameters, in contrast to the three-compartment model approach. Figure 12 examines this advantage of
the 100 compartment/CFD strategy over the simple 3-compartment model approach. The dead volume is plotted for the
various models as initiator feed fraction varies at a feed temperature of 300 K. In this case, the dead volume fraction
(1 f rV (0.05)) is considered to be the portion of the reactor
with polymerization rates less than 5% of the maximum polymerization rate. Using this measure, one sees that the dead
volume in the CFD simulations increases continuously as initiator feed fraction is varied in the stable range. This continuous variation in dead volume can be visualized by radical
concentration gradients shown in the contour graphs for selected points.
Although the 100-compartment model is clearly discrete in
nature, it matches the dead volume variation of the CFD
simulation quite well for all but the highest initiator feed
fractions, where exit temperatures approach the decomp limit.
In contrast, the three-compartment model does not match the
stability region or the dead volume fraction well. For this low
feed temperature of 300 K, the three-compartment model has a
large maximum in radical concentrations in the second zone,
with low polymerization rates in the rst and third zones that
fall below the 5% threshold. Thus, for the three-compartment
model parameters chosen, the dead volume fraction of the
three-compartment model is about 0.92 for the entire stable
From the results presented in Figure 12, it appears that
success in predicting the proper stability region requires the
model to vary the dead reactor volume. For this reason, it is
interesting to consider the development of simple alternatives

December 2005

Figure 13. Simplied model for representing a geometrically-expanding plume region.

The compartment volumes are dened as: V i V 1 i1 ,
where the growth factor is xed for this work at 1.2.
Note that V tot i1
V i.

to the three-compartment model that retain the volume variation of highly detailed models but do not require CFD simulation. A proposed modeling strategy to achieve this goal is
illustrated in Figure 13.
The simplied model presented in Figure 13 consists of a
sequence of perfectly-mixed reactors with recycle much like
the three-compartment model technique. However, many more
volumes are included to represent a plume that can shrink or
grow as the initiator half-life changes. Furthermore, the tanks
in the sequence increase geometrically in size as they move
away from the main feed, allowing for a wide range of effective reaction volumes. This modeling strategy is designed to
physically represent a feed plume that engulfs additional uid
as it progresses down the reactor.
The model illustrated in Figure 13 is similar in its physical
representation of the ow to the plume engulfment model
proposed by Zwietering.40 The plume engulfment model has
been applied to general reactive mixing problems, and considers incoming uid volumes to have varying ages. Fluid
volumes that have just entered the reactor are young, and
become closer to the bulk conditions as their age increases.
Unfortunately, the method of Zwietering requires the estimation of age time constants that may be difcult to obtain.
In contrast to the Zwietering mixing model,40 the plume
model presented in Figure 13 involves a more convenient
stirred-tank representation and is relatively simple to formulate. The plume model has only three parameters: the number
of compartments N compt. the recycle ratio R, and the factor
which describes the geometric rate of increase of compartment
Vol. 51, No. 12

AIChE Journal

previous research, results show that perfectly-mixed reactors

have much narrower stable operating regions than imperfectlymixed reactors. At lower feed temperatures and residence
times, perfectly-mixed reactors can have no desirable stable
operating branch, in contrast to imperfectly-mixed reactors
with the same feed conditions. In adiabatic operation, the
extent of stirring or recycle ow can affect the internal reactor
temperature elds, and can, therefore, have a very large effect
on both the ow and the reaction time scales throughout the
reactor. Simple three-compartment models show that a range of
stable recycle ratios can exist in the LDPE autoclave reactor,
and for many feed conditions, no amount of stirring can make
Table 4. Comparison of Models for LDPE
Autoclave Reactors
Figure 14. Comparison of the plume model to other
model formulations.
The effect of initiator feed fraction on exit temperature is
shown for feed temperature of 360 K. The base CFD case for
generating the 100-compartment model involves a feed
stream at 420 K with 120 ppm of TBPOA.

volumes along the reactor train. The rate of side feed to the
reactors from the recycle stream is proportional to the volume
of the receiving compartment. This new simplied model could
be parameterized by knowledge of plant data, similar to the
three-compartment model approach. However, the plume
model also has the ability to expand and contract its physical
region of reaction nearly continuously and automatically as
conditions change, similar to the 100 compartment/CFD methodology.
The plume model presented here represents a straightforward approach that can be applied in practice, and it is expected to capture the stability region more completely than the
simpler three-compartment model. This hypothesis is conrmed in Figure 14, which compares the performance of the
plume model (for 1.2, R 15, and 20 zones) to that of the
three-compartment model, the 100 compartment model, and
CFD data. As can be seen from this gure, the plume model
captures a span in the stable region that is similar to the
100-compartment model. In general, the plume model performs
quite well given its simplicity, but it does not provide the same
level of spatial detail as the 100-compartment model. Other
advantages and disadvantages of the various modeling approaches employed here are summarized in Table 4, and should
be considered when modeling the operation of LDPE autoclave
In general, because of its relative simplicity, good accuracy,
and ease of parameterization, the new plume model may represent a good compromise between effort and accuracy when
simulating the LDPE autoclave reactor. The plume model may
be especially attractive if a CFD model has not yet been
developed for the reactor, since model setup is often the most
time-consuming part of the overall CFD simulation process.41

Summary and Conclusions

In this work, analysis using both simplied and detailed
models emphasizes the importance of including mixing effects
in models of adiabatic LDPE reactors. In agreement with
AIChE Journal

December 2005

Model IPerfect mixing:

Very simple and fast computation, including stability
No adjustable ow/volume parameters
Does not require knowledge of the ow eld
Does not offer stabilization from poor initiator usage or
variable effective reaction volume
Very limited stability region
Model IIThree-compartment model technique:
Simple and fast to compute, including stability analysis
Includes the stabilizing effect of poor initiator mixing,
greatly expands stability region compared to perfect
Only requires rough knowledge of ow eld
Only designed to represent poor mixing in feed zone
Requires choice of free parameters for ows/volumes
Lacks ability to have variable downstream volume
wastage; can signicantly underestimate initiator wastage
and range of stable operating conditions
Most general method for representing reactive mixing
Provides great spatial detail
Includes the stabilizing effect of poor initiator mixing as
well as variable effective reaction volume; more
accurately represents feasible operability range
No free parameters for ows and volumes
Time consuming to set up and compute; difcult to use in
stability analysis
Requires detailed knowledge of reactor geometry
Model IVCombined CFD/Compartment model:
Retains spatial detail of CFD solution
Accurately represents stability region
Much faster than CFD, allowing stability analysis
No free parameters for ows and volumes
Requires at least one CFD simulation or detailed
knowledge of ow/reactive eld
Model VNew plume mixing model:
Can represent stability region fairly well
Simple to compute, allowing stability analysis
Does not require CFD or detailed knowledge of ow/
reactive eld
Loses some knowledge of spatial details
Requires specication of free parameters for

Vol. 51, No. 12


radical concentration uniform in the desirable stable operating

Models of the LDPE autoclave reactor using a combined
compartment model/CFD approach provide a great degree of
spatial detail, and also reproduce the stable operation region
with a high degree of accuracy. Such detailed models allow the
effective reaction volume for polymerization to vary continuously as feed conditions are changed. Lower feed temperatures
and initiator feed fractions provide the highest degree of uniformity in initiator and radical concentration. Such uniform
conditions result in the most efcient volume usage for reaction
and the highest initiator efciency in producing polymer. However, obtaining such efcient conditions requires operation
near the stability limit for reactor extinction.
Analysis of several different models shows that the variation
in effective volume usage is key to capturing the broad stability
range observed in CFD simulations. On the basis of this knowledge, this article proposes a new simplied imperfect mixing
model of the LDPE autoclave that represents the feed plume as
a series of interconnected compartments with geometrically
increasing volumes. The plume model captures the wide range
of stable operation observed in CFD simulations. Furthermore,
no tting to CFD simulations is required, and yet the predictions are quite close to the CFD results. Thus the plume model
may represent a good tradeoff between computational effort
and accuracy.

The authors are indebted to the industrial sponsors of the University of
Wisconsin Polymer Reaction Engineering Laboratory, and to the U.S.
Department of Energy for nancial assistance. We also thank Iasson
Mustakis and Carlos Villa for their insight and support.




computational uid dynamics

monomer concentration, mol/L
heat capacity of mixture, cal/(g-K)
dead polymer chain of length n monomer units
activation energy used in the Arrhenius expression for rate
constants, cal/mol
effectiveness factor for initiator
fraction of exit stream from compartment i that enters
compartment j
fraction of volume considered to be active or used in
producing polymer
volume fraction of a given zone i
unit step function
pre-exponential factor for rate constants, variable units
rate constant for initiator decomposition, s1
rate constant for initial monomer decomposition, L/(mol-s)
rate constant for propagation, L/(mol-s)
rate constant for decomp propagation (path 1), L/(mol-s)
rate constant for decomp propagation (path 2), s1
rate constant for termination by combination, L/(mol-s)
low-density polyethylene
monomer species
molecular weight of initiator, g/mol
monomer molecular weight, g/mol
number of compartments
number of spatial zones used in averaging
absolute pressure, atm
live polymer chain of length n monomer units
mass ow rate exiting compartment n, kg/s
ideal gas constant, 1.987 cal/(mol-K)
recycle ratio, dimensionless
free radical

December 2005



rotations per minute

rate of heat generation by monomer decomposition, cal/L-s
rate of polymerization, mol/L-s
internal reactor temperature, K
temperature in compartment k, Kelvins
feed temperature, K
tert-butyl peroxyacetate (peroxide initiator, 132.2 g/mol)
activation volume, cal/(atm mol)
total reactor volume, L
volume of a spatial zone or compartment n, L
initiator feed fraction, ppm by weight
initiator weight fraction in compartment k, dimensionless
polymer weight fraction in compartment k, dimensionless
monomer conversion, dimensionless
Use factor for quantifying effective reactor volume based
on simulation results
specic initiator consumption, mmol initiator consumed/kg
polymer produced
heat of polymerization, cal/mol
heat of decomposition, cal/mol
total growing radical concentration, mol/L
total growing radical concentration in compartment k,
multiplicative expansion factor used in plume model, dimensionless
mixture density, g/L

Literature Cited
1. McCoy M, Reisch MS, Tullo AH, Short PL, Thayer AM, Tremblay JF,
Storck WJ. Facts and gures for the chemical industry. Chem and Eng
News. 2003;81:25 66.
2. Roedel MJ. The molecular structure of polyethylene. I. Chain branching in polyethylene during polymerization. J A Chem Soc. 1953;75:
6110 6112.
3. Kiparissides C, Verros CG, MacGregor JF. Mathematical modeling,
optimization, and quality control of high-pressure ethylene polymerization reactors. J Macromolecular Sci. Reviews in Macromolecular
Chemistry and Physics. 1993;C33:437527.
4. Doak KW. Low density polyethylene. In: Mark HF. Encyclopedia of
Polymer Science and Technology. New York: Wiley Interscience
Publishing, 1985:383 429.
5. Beasley JK. Polymerization at High Pressure. In: Allen SG, Bevington
JC. Comprehensive Polymer Science. New York: Pergammon Press,
6. Gemassmer AM. Autoclave Process for the High Pressure Polymerization of Ethylene. Erdol und Kohle-ErdgasPetrochemie. 1978;31:
7. Christl RJ, Roedel MJ. USA Patent No. 2897183, 1959.
8. Pladis P, Kiparissides C. Steady-State Multiplicity in High Pressure
Ethylene Polymerization Reactors. 3rd Annual AIChE Polymer Producers Conference, Houston, TX: 1999.
9. van der Molen TJ, van Heerden C. The Effect of Imperfect Mixing on
the Initiator Productivity in the High Pressure Radical Polymerization
of Ethylene. 1st International Symposium on Chemical Reaction Engineering, Washington, D.C. American Chemical Society; 1970.
10. van der Molen TJ, Koenen A, Oosterwijk H, van der Bend H. Lightoff temperature and consumption of 16 initiators in LDPE production.
Quaderni dellIngegnere Chimico Italiano. 1982;18:715.
11. Georgakis C, Marini L. The Effect of Mixing on Steady-State and
Stability Characteristics of Low Density Polyethylene Vessel Reactors. Seventh International Symposium on Chemical Engineering,
Boston, Massachusetts, American Chemical Society; 1982.
12. Marini L, Georgakis C. Low-Density Polyethylene Vessel Reactors.
Part I: Steady State and Dynamic Modeling. AIChE J. 1984;30:401
13. Marini L, Georgakis C. Low-Density Polyethylene Vessel Reactors.
Part II: A Novel Controller. AIChE J. 1984;30:409 414.
14. Villa CM, Dihora JO, Ray WH. Effects of Imperfect Mixing on
Low-Density Polyethylene Reactor Dynamics. AIChE J. 1998;44:
1646 1656.
15. Lee HJ, Yeo YK. Modeling and Simulation of High Pressure Autoclave Polyethylene Reactor Including Decomposition Phenomena. J
Chem Eng of Japan. 2000;33:323329.

Vol. 51, No. 12

AIChE Journal

16. Donati G, Gramondo M, Langianni E, Marini L. Low density polyethylene in vessel reactors. Quaderni dellIngegnere Chimico Italiano.
1981;17:88 96.
17. Singstad P. Modelling and Multivariable Control of High Pressure
Autoclave Reactors for Polymerization of Ethene, The Norwegian
Institute of Technology; 1992. PhD Diss.
18. Chan WM, Gloor PE, Hamielec AE. A Kinetic Model for Olen
Polymerization in High-Pressure Autoclave Reactors. AIChE J. 1993;
19. Nordhus H, Moen , Singstad P. Prediction of Molecular Weight
Distribution and Long-Chain Branching Distribution of Low-Density
Polyethylene from a Kinetic Model. J Macromol Sci Pure Appl Chem.
20. Ghiass M, Hutchinson RA. Simulation of Free Radical High-Pressure
Copolymerization in a Multizone Autoclave Reactor: Compartment
Model Investigation. Macromol Symposia. 2004;206:443 456.
21. Torvik R, Gravdahl AR, Fredriksen GR, Moen , Laurell J. Design of
HPPE stirred autoclaves using 3D computational uid dynamics. 5th
International Workshop on Polymer Reaction Engineering: Berlin;
22. Read NK, Zhang SX, Ray WH. Simulations of a LDPE Reactor Using
Computational Fluid Dynamics. AIChE J. 1997;43:104 117.
23. Tosun G, Bakker A. A study of macrosegregation in low-density
polyethylene autoclave reactors by computational uid dynamic modeling. Ind Eng Chem Res. 1997;36:296 305.
24. Zhou W, Marshall EM, Oshinowo L. Modeling LDPE tubular and
autoclave reactors. Ind Eng Chem Res. 2001;40:55335542.
25. Fox RO. Computational methods for turbulent reacting ows in the
chemical process industry. Revue de LInstitut Francais du Petrole.
26. Kolhapure NH, Fox RO. CFD analysis of micromixing effects on
polymerization in tubular low-density polyethylene reactors. Chem
Eng Sci. 1999;54:32333242.
27. Zhang SX, Read NK, Ray WH. Runaway phenomena in low-density
polyethylene autoclave reactors. AIChE J. 1996;42:29112925.
28. Wells GJ. Modeling methods for predicting polymer process dynamics
and product properties: the Effects of Mixing and Catalyst Characteristics, University of Wisconsin-Madison; 2003. PhD Diss.
29. Odian G. Principles of Polymerization. New York: John Wiley and
Sons; 1991.
30. Chen CH, Vermeychuk JG, Howell JA, Ehrlich P. Computer Model
for Tubular High-Pressure Polyethylene Reactors. AIChE J. 1976;22:
463 471.
31. Zhang SX. Modelling and Experimental Studies of Free Radical
Polymerization Reactors. University of Wisconsin-Madison; 1996.
PhD Diss.
32. Zhang SX, Ray WH. Modeling of Imperfect Mixing and Its Effects on
Polymer Properties. AIChE J. 1997;43:12651277.
33. Sittig M. Low Density Polyethylene Processes. Polyolen Production
Processes: Latest Developments. Noyes Data Corporation; 1976.
34. Kwag BG, Choi KY. Effect of Initiator Characteristics on HighPressure Ethylene Polymerization in Autoclave Reactors. Ind Eng
Chem Res. 1994;33:211217.
35. Goto S, Yamamoto K, Furui S, Sugimoto M. Computer Model for
Commercial High-Pressure Polyethylene Reactor Based on Elementary Reaction Rates Obtained Experimentally. J Appl Poly Sci: Applied
Polymer Symposium. 1981;36:21 40.
36. Smit L. The Use of Micromixing Calculations in LDPE-Reactor Modelling. 4th International Workshop on Polymer Reaction Engineering;
37. Wells GJ, Ray WH. Methodology for modeling detailed imperfect
mixing effects in complex reactors. AIChE J. 2005;51:1508.
38. Fluent Incorporated. Fluent Users Guide: Version 6.0; 2001.
39. Britton LG, Taylor DA, Wobser DC. Thermal Stability of Ethylene at
Elevated Pressures. Plant/Operations Progress. 1986;5:238 251.
40. Zwietering TN. A backmixing model describing micromixing in single-phase continuous-ow systems. Chem Eng Sci. 1984;39:1765
41. Versteeg HK, Malalasekera W. An Introduction to Computational
Fluid Dynamics: The Finite Volume Approach. London: Longman
Group; 1995.

AIChE Journal

December 2005

State equations
The state equations for zone k of a compartment model with
any number of compartments are given in the equations below.
If not specied, all rate constants, concentrations, mass fractions, and temperatures are at the conditions of the compartment k. Compartment volumes are assumed to be time invariant, and constant physical properties are also assumed.
dw Ik

d 0k






f Q 2k f C k



d eff




f Q w k C




0 MWmon



f Q T H




dw pk

dT k


f Q w k C MW


k Cmon0

poly p



(1.89kmd kpd1 ) 0.0714kpd2 Cmon]


Denition of parameters for the compartment models

To dene the generalized compartment models used in this
work,37 the set of tank volumes {V n ; n 1 . . . N compt} and the
interconnecting ow fractions { f ji ; i, j 0 . . . N compt} must
be dened. The interconnecting ow fraction f ji is the amount
of the exit stream from compartment i that enters compartment
j. The overall inlet ow to the reactor is denoted stream 0.

100-Compartment models
The 100-compartment model used in this work is selected
from a CFD base case using criteria that minimize variations in
the temperature, propagation rate, and termination rate in each
selected zone. Details of this selection process are given in
Wells and Ray.37

3-Compartment model
For the 3-compartment model (see Figure 2), the volumes
are taken to be the values used by Zhang and Ray,32 namely
V 1 V tot/24
V 2 2V tot/24
V 3 21V tot/24


The ow fractions for the three compartment model depend on

the recycle ratio (R), and are dened as follows
f 01 1
Vol. 51, No. 12


f 12 1


f 23 1

V tot

N compt1


31 R

f 32

V i V i1; 1 i N compt

Once the compartment volumes are known, the ow fractions

f ji can be determined from the recycle ratio since the recycle
ows from the last compartment are proportional to the volume
of the receiving compartment. These ow fractions are dened
as follows

31 R

f ii 1, i 1 . . . 3
f ij 0, all other cases


f 01 1
f ii1 1; 1 i N compt

Plume model
The plume model (see Figure 13) denes the ow fractions
and volume fractions based on the number of compartments in
the model (N compt), the recycle ratio (R), and the volume
growth factor (). Based on the denitions given in Figure 13,
the compartment volumes can be related to the total reactor
volume and growth factor using the following recursive relationship



December 2005

f ii 1, i 1 . . . Ncompt
f Ncompt

; 1 i Ncompt
R 1Vtot VNcompt

f ij 0, all other cases


Manuscript received Nov. 24, 2004, and revision received Mar. 10, 2005.

Vol. 51, No. 12

AIChE Journal

Low-Order Models for Catalyst Particles:

A Dynamic Effectiveness Factor Approach
lvarez-Ramrez, Francisco J. Valdes-Parada, and J. Alberto Ochoa-Tapia
Jose A
Division de Ciencias Basicas e Ingeniera, Universidad Autonoma MetropolitanaIztapalapa, Mexico, D.F., Mexico
DOI 10.1002/aic.10593
Published online August 31, 2005 in Wiley InterScience (www.interscience.wiley.com).

Models for heterogeneous reactors are composed of distributed parameter ordinary

and partial differential equations to describe balances both for a uid and a catalyst
particle. For steady-state conditions, the effectiveness factor (EF) approach has been
primarily used to reduce the model dimensionality, which yields an important computational effort during reactor design stages. The idea behind EF is to express the reaction
rate in a catalyst particle by its rate under surface/bulk conditions multiplied by the EF.
As a consequence, only few arithmetic operations to obtain the catalyst particle total
reaction rate have to be made, avoiding in this form repeated solutions of the more
sophisticated distributed parameter model. The aim of this paper is to extend the EF
approach to obtain a simple model for the dynamics of the catalyst particle total reaction
rate. As a preliminary step, a Laplace domain approach is used to introduce a dynamic
EF (DEF) concept as a linear operator that transforms reaction rate signals from
surface/bulk to catalyst particle conditions. It is shown that the DEF becomes the transfer
function of a model governing the dynamics of the catalyst particle reaction rate.
However, the resulting transfer function is innite-dimensional. This implies that, contrary to the steady-state case, an exact simple modeling of the reaction rate dynamics is
not possible. An approach to obtain low-order models, based on an approximate model
matching, is proposed. For a rst-order model, the time constant is computed for common
particle geometries showing its dependency with the Thiele modulus. In fact, the results
show that the (dominant) time constant is a decreasing function of the Thiele modulus. In
this form, the faster the chemical reaction dynamics, the faster the reaction diffusion
catalyst particle dynamics. 2005 American Institute of Chemical Engineers AIChE J, 51:
3219 3230, 2005

Keywords: catalyst particle, reaction diffusion, low-order model, approximate model


Dynamic models for heterogeneous reactors are composed
of distributed parameter, ordinary, and partial differential equa-

lvarez-Ramrez is also afliated with Programa de Investigacion en MatemJ. A

aticas Aplicadas y Computacion, Instituto Mexicano del Petroleo.
Correspondence concerning this article should be addressed to J. A
at jjar@xanum.uam.mx.

2005 American Institute of Chemical Engineers

AIChE Journal

December 2005

tions to describe balances both for a uid and a disperse phase

composed by catalyst particles. Simple models for the uid can
correspond to continuous stirred tank assumptions, which yield
coupled ordinary differential and algebraic equations. Given that
intraparticle resistance to mass transfer becomes signicant for
catalyst particles, mass and energy balances can lead to distributed
parameter differential equations to describe the concentration distribution into the catalyst particle. The uid and the catalyst
particle models are coupled through a mass-transfer ux on the
particle surface, which depends on uid conditions.
Vol. 51, No. 12


The design and analysis of heterogeneous reactors may

involve repeated solutions of the model. For instance, within an
optimization-based design methodology, the reactor model solution is required at each optimization trial. On the other hand,
on-line reactor dynamic, or even steady-state, simulation would
involve solving the reactor model for each simulation step.
Accurate numerical solution of distributed parameter models,
as in the catalyst particle case, can be made with standard
nite-element and nite-difference methods. However, such
methods lead, in general, to a large set of algebraic and differential equations. Under this situation, the computational burden
involved in repeated reactor model solutions can be excessive
or even prohibitive. Given its parameter distributed nature, in
general, the largest numerical burden is associated with the
catalyst particle model. This practical problem constitutes an
incentive to nd approximate dynamic models for heterogeneous reactors, specically for catalyst particles. Indeed, this is
the aim of the paper.
For steady-state conditions, the effectiveness factor (EF)
approach has been largely used to reduce the model dimensionality.1 The idea is to express the reaction rate in a catalyst
particle by the reaction rate under surface/bulk conditions
multiplied by the EF. As a consequence, only few arithmetic
operations are required to obtain the catalyst particle reaction
rate, avoiding in this form repeated solutions of the highdimensional distributed parameter model. Few studies have
focused on the issue of approximate methods for catalyst
particles under dynamic conditions.2-4 The proposal is to approximate the dynamic distributed parameter by a single initialvalue differential equation describing the dynamics of the
average particle concentration. The low-dimensional model has
been derived as a long-time approximation and its accuracy
increases for longer times.4 A criticism to this approximation
approach is that low-order models are oriented to describe
average concentration dynamics, which are rarely of interest in
most practical reactor design and simulation situations. Instead,
variables such as the reaction rate provided by the catalyst
particle are more useful to evaluate reactor performance. The
attractiveness of the EF approach relies on the fact that it leads
directly to the catalyst particle reaction rate as a function of
bulk/surface conditions. In this way, reactor models are readily
simplied by incorporating the particle reaction rate as a pseudo-reaction rate in the uid mass balance. Unfortunately, an
extension of the EF idea for dynamic conditions is not yet
The aim of this paper is to twofold:
(1) To extend the EF approach to dynamic conditions. To
the best of our knowledge, a dynamic EF concept has not been
reported previously in the open literature.
(2) To obtain a simple model for the dynamics of the reaction rate for a catalyst particle. To this end, we constrain
ourselves to the linear case as a prerequisite to address the more
interesting nonlinear case. In principle, this latter case can be
addressed following similar ideas as those reported by Szukiewicz.4
In a rst step, a Laplace domain approach is used to introduce
a dynamic EF (DEF) concept as a linear operator that transforms reaction rate signals from surface/bulk to catalyst particle conditions. Similar to feedback control theory for hyperbolic and parabolic partial differential equations (PDEs),5,6
what is done to compute the DEF is to formulate an output for

December 2005

linear parabolic reaction diffusion equations, which is the spatial integral averaging of the dynamic reaction rate. In this
form, it is shown that the DEF becomes the transfer function of
a model governing the dynamics of the catalyst particle reaction rate. However, the resulting transfer function is innitedimensional. This implies that, contrary to the steady-state
case, an exact simple (such as nite-dimensional) modeling of
the reaction rate dynamics is not possible. To overcome this
problem, an approach to obtain low-order models based on an
approximate model matching is proposed. For a rst-order
model, the time constant is computed showing its dependency
with the Thiele modulus. In fact, the results show that the
(dominant) time constant is a decreasing function of the Thiele
modulus. In this form, the faster the chemical reaction dynamics, the faster the reaction diffusion catalyst particle dynamics.
Numerical simulations are used to illustrate our ndings.

Dynamic Effectiveness Factor

In this section, a DEF concept that extends the classical one
based on nondynamic conditions is introduced. To this end,
similar to the case of the dynamic permeability concept,7 the
catalyst particle dynamics will be represented in the Laplace/
Fourier domain. Because we are concerned with fundamental
concepts rather than with EF computations for general cases,
analytic solutions for the reaction diffusion process will be
used. In this way, the concepts will be based on simple isothermic reaction diffusion processes leading to linear differential equations.

Classical EF concept
As mentioned earlier, the classical EF concept pertains to
nondynamic operating conditions. To recall the classical EF
concept, consider an isothermic catalyst particle where a single
rst-order chemical reaction is taking place. In a general geometry, by assuming constant diffusivity D and a Fickian
diffusion mechanism J Dc, the nondynamic process can
be modeled as follows (throughout, overbars denote nondynamic conditions):
Dc kc 0


where x 3 is the system coordinates vector, c is the

nondynamic reactive concentration, k is the reaction constant,


x 12 x 22 x 32

The domain D 3 of the catalyst particle is assumed to be

simply connected. The boundary conditions for Eq. 1 are given
c x c b
c x is finite

for all x D
for all x D


where c b is a constant boundary concentration and D denotes

the boundary of the domain D. The linear Eq. 1 has a unique
solution c (x) satisfying the boundary conditions of Eq. 2 [that
Vol. 51, No. 12

AIChE Journal

is, c (x) c b for all x D]. The EF expresses the reaction rate
in a catalyst particle by the rate under surface/bulk conditions.1
That is, the classical nondynamic EF (denoted by ) is dened
as follows:

D kc dV
D kc bdV


where dV dx1dx2dx3 is a generalized volume differential.

Analytic solutions of Eqs. 1 and 2 can be obtained for simple
geometries, and the EF expression can be computed by integration of the solution c (x) according to Eq. 3. Let




be the Thiele modulus, where L is a characteristic thickness

depending on the geometry. In fact, L is the slab length for
rectangular geometry and L is the cylinder and sphere radius
for cylindrical and spherical geometries, respectively. For common particle geometries analytical expressions are known in
advance. In fact,




for rectangular geometry

for cylindrical geometry

for spherical geometry


where I1 and I0 are the rst- and zeroth-order modied Bessel

functions of the rst kind, respectively.
The importance of the EF to obtain a reduced-order model
relies on the fact that it depends only on reaction and diffusion
parameters by the Thiele modulus , and not on the bulk/
surface conditions. This means that the EF is invariant under
perturbations of the bulk/surface uid conditions. This property
is exploited to obtain a useful reduced catalyst particle model.
In principle, to compute the total reaction rate D kc dV provided by the catalyst particle, one should solve the boundaryvalue problem of Eqs. 1 and 2. Standard efcient nite-element
or nite-differences numerical methods can be used to this end.
However, if such a computation has to be made iteratively, as
in optimization-based designs, an excessive and sometimes
prohibitive computational burden could be introduced. A reduced-order model is obtained from the EF concept as follows.



R b

kc bdV

be the reaction rate for particle and surface/bulk conditions,

respectively. From Eq. 3, because R b Vkc b , one obtains the
relationship R Vkc b . In this way, the effects of changes in
the uid conditions, reected in changes of c b , can be easily
estimated, leading to a signicant reduction of the computational effort during extensive reactor simulations.

Extending the effectiveness factor concept for dynamic

An extension of the classical EF concept to dynamic conditions should retain the property of relating reaction rates from
bulk/surface to particle conditions. Indeed, this property makes
possible the reduction of the complexity of the catalyst particle
model. To this end, we will interpret the EF as a scaling factor
between two reaction functions. To do this we proceed as
follows. It has been shown that R R b . The EF can be seen
as a steady-state operation that transforms reaction rate values
at boundary conditions R b into reaction rate values at catalyst
particle reaction diffusion conditions R ; that is,

: R b 3 R

An extension of the EF concept to dynamic conditions

should retain the following salient characteristics:
(1) It should be a linear transformation between dynamic
reaction rate functions at boundary conditions and dynamic
reaction rate functions at catalyst particle (reaction diffusion)
(2) When a dynamic EF is seen as an operator between such
reaction rate functions, it should converge uniformly to the
steady-state operator .
To introduce the DEF concept as a linear transformation
between dynamic reaction rate functions, we resort to Laplace
domain methods. Recall that if f (x, t) is a time function, its
Laplace transform F(x, s) L[ f (x, t)] is given by F(x, s)
0 estf (x, t)dt, where s denotes the Laplace variable. If cb(t) is
a dynamic boundary condition, its Laplace transform is Cb(s)
L[cb(t)] and the corresponding dynamic reaction rate at boundary condition is Rb(s) VkCb(s). The dynamic reaction rate at
particle conditions R(s) is computed as


kC x, sdV


where C(x, s) is the solution of the dynamic boundary-value

problem (notice that sc is the Laplace transform of c/t):
DC kC sC

kc xdV



with boundary conditions

C x, s C bs

for all x D

Cx, s is finite

for all x D

AIChE Journal

December 2005

Vol. 51, No. 12


Similar to the steady-state EF concept, we dene the DEF as


The DEF can be expressed in terms of a normalized frequency as follows. Let

The DEF is a linear operator (s) that transforms dynamic

total reaction rates at boundary conditions Rb(s) into dynamic
reaction rates at particle conditions R(s).

d D/L 2
be the diffusion characteristic frequency. Introduce the normalized Laplace variable as (recall that s has frequency units)

That is, the DEF is dened as the transformation

s : R bs 3 Rs


In other words, the DEF is a linear operator satisfying the

Rs s R bs
It is asked that (s) 3 as s 3 0 (that is, steady-state
conditions). By serving as the functional factor, the DEF
should give some information on the effectiveness of a single
catalyst particle when it is subjected to external perturbations in
the surface/bulk conditions.
It should be remarked that the denition of the DEF as a
linear dynamic operator between input [Rb(s)] and output [R(s)]
signals is motivated by control theory concepts. Specically,
similar to the approach by Christodes and Daoutidis,5,6 the
signal R(s) is an output of the linear parabolic PDE given by
Eqs. 8 and 9. The computation of the DEF is equivalent to the
computation of the transfer function between the input Rb(s)
and the output R(s). In this way, as (s) 3 when s 3 0 is
required that would correspond to the steady-state gain.



In the time domain, the Eq. 14 normalization is equivalent to

normalizing the time variable as td. In fact, td 1
can be
interpreted as a characteristic diffusion timescale. For (s) k
s the dynamic Thiele modulus can be written as
s n 2 s n


In this way, in terms of normalized variables, the DEF

expressions for common geometries are the following:

s n,

2I1 sn

sn I0 sn

sn tanhsn sn 2

for rectangular geometry

for cylindrical geometry

for spherical geometry


Computation of the DEF

In principle, to compute the DEF (), one should nd the
solution C(x, s) of the boundary-value problem (Eqs. 8 and 9)
and use it to calculate R(s) by the integral in Eq. 7. An
expression for the DEF can be easily obtained from the nondynamic EF expressions (see, for instance, Eq. 5) by noticing
that Eq. 8 can be written as
DC sC 0


where (s) k s can be interpreted as a dynamic reaction

rate constant. By considering the boundary condition Eq. 9, Eq.
11 can be solved in a similar way as the nondynamic case (Eqs.
1 and 2). To this end, let us introduce the dynamic Thiele

s L 2


Thus, if () is an expression for the nondynamic EF,

an expression for the DEF is obtained by a formal substitution
x (). That is, the DEF is obtained as

s s


Because (0) 3 , one obtains the desired convergence

(0) to nondynamic conditions.


December 2005

For convenience, the dependency of the DEF has been made

explicit. In this way, (sn, ) can be seen as a family of linear
operators parameterized by the classical Thiele modulus .

Frequency response of DEF

For convenience the discussion throughout will be made in
terms of the normalized variable sn. For each Thiele modulus
, the DEF (sn, ) is a linear operator acting on reaction rate
signals Rb(s). Some salient characteristics of (sn, ) can be
obtained by studying its frequency response in the Fourier
domain. This is done by making the formal substitution s x i
(sn x in), where is the frequency and i 1. For
simplicity in presentation, computations are made only for
rectangular geometry. Similar results are obtained for cylindrical and spherical geometries. Figures 1 and 2 show, respectively, the Nyquist plot Re[(n)] vs. Im[(n)] and the magnitude Bode plot (n, ) , for three different values of the
Thiele modulus . For comparison, the limiting case 0
corresponding to a reaction-free process is also included. The
following salient features are observed:
For each Thiele modulus , the Bode plot displayed in
Figure 2 shows that the DEF magnitude (n, ) is a
monotonous decreasing function of the frequency n. This
means that the best reaction rate performance of a catalyst
particle is found at nondynamic conditions. That is, the largest
reaction rate RT(n) is obtained when the reaction diffusion
process is operated at steady-state conditions. Dynamic perturVol. 51, No. 12

AIChE Journal

Figure 1. Nyquist plot for four different values of the

Thiele modulus .
Observe the high frequency convergence of the Fickian diffusion behavior (n, ) 1/in.

bations will deteriorate the performance of the catalyst particle.

This is an incentive to introduce control actions aimed at
reducing dynamic perturbation effects on the system performance.
For nondynamic conditions one has that
(1) (2)
for 1 2. That is, the nondynamic EF is a decreasing
function of the classical Thiele modulus . For each frequency
n, this property is maintained for the magnitude of the dynamic effectiveness factor (n, ). In fact, Figure 2 shows
that, as expected, (n, 1) (n, 2) for 1 2 and
all n. As a consequence, the DEF for the reaction-free case
0 becomes an upper bound for the DEF (n, ). That is,
the largest DEF is obtained for reaction-free conditions. The
Nyquist plot in Figure 1 shows that the curve 0 surrounds
the curves for 0.
Recall that the diffusion characteristic frequency d has
been used to normalize the frequency variable. In this way, n
1 corresponds to d. Figure 2 shows that the DEF
magnitude (n, ) () for n 1. That is, dynamic
perturbations with frequencies lower than the diffusion characteristic frequency d do not have a signicant effect on the
performance of a catalyst particle. As a consequence, the
Nyquist plot of (n, ) approaches the unit circle behavior for
n 1. This is illustrated in Figure 1 where the point corresponding to n 1 has been marked. From the physical
perspective, the behavior (n, ) () for n 1 can be
explained by the fact that, because surface concentration uctuations with d are relatively slow compared with diffusion dynamics, the reactant supply into the catalyst particle
bulk is not decreased. This allows the reaction diffusion process to perform similarly to steady-state conditions. As previously observed by Christodes and Daoutidis,5,6 when working
with Galerkin approximation techniques, this suggests that the
AIChE Journal

December 2005

dynamics of the reaction diffusion system are dominated by a

few slow modes that can capture, to some degree, the dominant
asymptotic spatiotemporal behavior of the overall reaction rate
For d 1, the magnitude (n, ) goes through a fast
roll-down behavior to achieve values close to zero at high
frequencies. This behavior is reected in the Nyquist plot
(Figure 1) as a fast convergence of the curve (n, ) to the
origin. The drastic deterioration of the catalyst particle performance when the perturbation frequency is increased beyond the
diffusion characteristic frequency d is attributed to a reduction
in the effective usage of the particle space. In fact, when the
frequency of the perturbation is large, (N, ) 1 because reactants cannot penetrate the particles during one perturbation period. At high frequencies, the catalyst particle
appears as a semi-innite particle because of the limited penetration of the diffusion wave. In this form, for high-frequency
perturbations the DEF is close to zero because a signicant
space of the catalyst particle is unused.
For high frequencies, Figure 1 shows that the Nyquist plot
(n, ) approaches a 45 slope in the fourth quadrant, which
corresponds to a plot of the form (n, ) (in)1/2. This
behavior can be easily derived by analyzing the reaction-free
case 0. For 0, one has that (n, 0) tanh[(in)1/2]/
(in)1/2. One can show that tanh[(iN)1/2] 1 holds at highfrequency conditions, which leads to the approximation (N,
) (iN)1/2. Thus the high-frequency behavior (N, )
(iN)1/2 corresponds to a regime controlled by the diffusion
Summing up, the better performance for the reaction diffusion
process (Eqs. 8 and 9) is found for steady-state (that is, zero
frequency) operating conditions. Besides, two regimes separated by the characteristic diffusion frequency d are observed
in the behavior of the DEF (n, ):
(1) A low-frequency regime where the performance of the

Figure 2. Magnitude Bode plot corresponding to the

conditions of Figure 1.
Notice that (n, ) is a decreasing function of the frequency n.

Vol. 51, No. 12


reaction diffusion process is not seriously deteriorated. Disturbances with frequency components in the range d do
not have a signicant effect on the supply of molecules to the
reaction mechanism.
(2) A high-frequency regime where the performance is seriously depleted. Because of high-frequency perturbations, the
molecules are unable to penetrate the particle, such that the
reaction rate dynamic R(sn) is seriously affected.
It is noticed that for the low-frequency regime (n, )
(). That is, the DEF is almost equal to the nondynamic EF.
This shows that for long times (corresponding to n 3 0) or for
low-frequency perturbations one can take the quasi-steadystate model R(t) ()Vkcb(t) to compute the reaction rate
behavior R(t). Indeed, this is an assumption commonly made in

Low-Order Dynamic Modeling Based on DEF

From the interpretation of the DEF as a linear operator
between reaction rates, by considering that Rb(sn) kVCb(sn),
the exact dynamics of the particle reaction rate R(sn), in terms
of boundary concentration dynamics Cb(sn), are described in
the Laplace domain as follows:
Rs n gs n, C bs n



gs n, kV s n,
Equation 17 describes the dynamics R(sn) affected by the
boundary concentration dynamics Cb(sn). Because the operator
g(sn, ) is causal, there exist matrices (A, B, C) such that the
dynamics of Eq. 17 can be described in state-space form. That
is, if z is a state vector, it is possible to write10
dzt n
Azt n Bc bt n
dt n
Rt n Czt n


where, given that sn sd, tn td tD/L2 is a dimensionless

time variable. One has that the equality g(sn, ) C(snI
A)1B must be satised. Unfortunately, the order of the DEF
operator is innite, implying that the state vector z is innitedimensional. To see that the order of (sn, ) is innite, let us
look at the simplest rectangular geometry case: (sn, )
tanh[(sn)]/(sn). In terms of fundamental exponentials, the
operator (sn, ) can be written as
e sn e sn
2 s e 2sn e 2sn

s n,


One observes that a single root is located at 2. However,

because the equation is transcendental, the part e sn

contains innitely many (complex) roots. This property

recovers, in some cases, the fact that the reaction rate dynamics
R(t) depends on a distributed reaction diffusion mechanism.
The above result shows that, although for the steady-state
case the EF concept leads to a simple model for the reaction
rate R , for the dynamic case such a concept leads to an
innite-dimensional model that is as sophisticated as the original one given by the partial differential equation (Eq. 8).
However, an advantage of the DEF-based model R(sn) g(sn,
)Cb(sn), compared to the original one given by Eq. 8, is that
the dependency on the spatial variable x has been removed. In
fact, the transfer function g(sn, ) kV(sn, ) depends only
on the temporal operator snc c/t. We will exploit this
characteristic to approximate the innite-dimensional model
R(sn) g(sn, )Cb(sn) by a simpler nite-dimensional one,
R(sn) h(sn, )Cb(sn), where h(sn, ) denotes a nite-dimensional transfer-function approximating the innite-dimensional
one, g(sn, ).
The magnitude Bode plot of the DEF (see Figure 2) shows
that the transfer function g(sn, ) behaves as a low-pass lter.
In this form, the type of parabolic PDE systems given by Eqs.
8 and 9 is characterized by a few dominant modes that can
capture the dominant long-term spatiotemporal behavior of the
concentration equation. This suggests that the innite-dimensional dynamics of the reaction rate R(s) can be adequately
approximated with a nite-dimensional low-order model. As
mentioned earlier, this is equivalent to nding a rational transfer function h(sn, ) approximating the innite-dimensional
one, g(sn, ). In the following, we will explore the use of
simple rst-order and second-order models to approximate the
reaction rate dynamics.
The idea of exploiting the existence of dominant modes for
computing (nite-dimensional) models for reaction diffusion
equations is not new. Methods based on a singular perturbation
formulation by Galerkins method have led to an accurate
description of reaction-rate dynamics. In turn, such low-order
approximations have allowed successful application to the control of distributed processes.6,8,9 The distinctive feature of our
approach is that it relies on an approximation of a transfer
function that, as done in this work, can be made with classical
frequency response tools.

First-Order Approximate Model

A rst-order transfer function is surely the simplest approximate model for the reaction dynamics. Similar approximate
modeling for the average concentration was previously proposed by Szukiewicz.4 The Bode plot of the dynamic effectiveness factor (see Figure 2) shows that the magnitude (n,
) is a monotonous decreasing function with limits (0, )
3 () as 3 0 and (n, ) 3 0 as 3 . This
suggests the following stable rst-order model structure:

hs n,

The order of the operator (sn, ) is equal to the number of

roots of the characteristic equation
Es n, 2 s n e sn e sn


December 2005

ns n 1

where () is a steady-state gain and n() is a (normalized)

time constant. Given that R(sn) h(sn, )Cb(sn) and h(0, )
Vol. 51, No. 12

AIChE Journal

(), one has R ()c b for steady-state conditions. To

retain the steady-state characteristic of the exact dynamics
R(sn) g(sn, )Cb(sn), the equality


Imn ,

ns n 1


Interestingly, by recalling that g(sn, ) kV(sn, ), the

approximation problem h(sn, ) g(sn, ) is equivalent to
nding a rst-order approximation of the form ()/[n()sn
1] to the exact DEF (n, ). That is,


ns n 1


Given the linearity of the approximate model, without losing

generality one can assume that R(tn 0) 0. In fact, these
initial condition is the steady-state value R 0 corresponding
to the input cb 0. In this way, under the aforementioned
initial condition assumption, the output R(tn) can be seen as a
deviation variable from the steady-state value R 0. In this
way, a time-domain realization of the approximate rst-order
model for the particle reaction rate dynamics is
dRt n
n 1Vk c bt n Rt n
dt n

n, Ren , i Imn ,


(2) For convenience in notation, let


f n,

ni n 1

This function can be written in terms of its real and imaginary parts as follows:
f n,

n n
n 2 n2 1
n 2 n2 1


Im0, n n n
where Im[(0, )] is the derivative of Im[(0, )] with
respect to the frequency n. By assuming the equality Im[(0,
)]n ()n()n, an estimate of the time constant is
given by



Figure 3 shows that the derivative Im[(0, )] 0 for ,

which implies that the time constant given by Eq. 28 is well
dened [that is, n() 0].
(6) Because (n, ) is a sophisticated function, the attempt
to analytically obtain Re[(n, )] and Im[(n, )] is not
possible in general. Instead, numerical methods can be used to
compute the time-constant estimate as specied by Eq. 28.
Specically, for a given Thiele modulus, the imaginary part as
a function of frequency is computed by a numerical routine.
This can be made with standard FORTRAN or MATLAB
packages. Next, the derivative Im[(0, )] is approximated as
a forward nite difference:


(3) Pose an approximate model-matching problem to compute the time constant n(). By matching the real and imaginary parts of f (n, ) and (n, ) (see Eqs. 24 and 25), the
following two approximation problems should be solved:
AIChE Journal

n n
n 2 n2 1

It has been established that the complex function (n, ) is

innite-dimensional, whereas f (n, ) is only one-dimensional. In general, an exact solution to the model-matching
problem f (n, ) (n, ) does not exist. Thus an approximate model-matching problem should be posed instead. Because the model-matching problems expressed in Eqs. 26 and
27 cannot be solved for the whole frequency range, model
matching can be satised only for a limited frequency range. It
is suggested to pose the approximate model-matching problems
in a frequency range around the steady-state condition n 0.
In fact, given that the steady-state equality f (0, ) (0, )
() was forced to hold, extending the matching to frequencies sufciently close to n 0 will guarantee the asymptotic
convergence R(t) 3 Vk ()cb.
(4) Compute a zeroth-order Taylor expansion of both sides
of Eq. 26 at steady-state conditions. Such expansion gives
Re[(0, )] (), which is an outcome of Eq. 20.
(5) Compute a rst-order Taylor expansion of both sides of
Eq. 27 at steady-state conditions. Taking note that Im[(0, )]
0, one then obtains the following relationship:


Notice that, for cb(tn) c b , R(tn) 3 Vk ()c b asymptotically.

The frequency domain (that is, sn 3 in) is used to compute
the time constant in Eq. 22 according to the following procedure:
(1) Decompose the complex function (n, ) into its real
(Re) and imaginary (Im) parts as follows:



should hold, which leads to the approximate model

hs n,

n 2 n2 1

Ren ,

December 2005


Im, Im0,

for a sufciently small 0. Our experience has shown that

102 sufces to obtain a stable nite-difference approxiVol. 51, No. 12


where () 0 is a damping factor. Similar to the rst-order

case, a time-domain realization of the approximate secondorder model is

n 2

dRt n
d 2Rt n
2 n
Rt n Vk
dt n
dt n2

Because n()2s 2n 2n()()sn 1 is a stable polynomial,

one has that, for cb(t) c b , R(tn) 3 Vk ()c b asymptotically.
Similar steps that led to an expression for the time constant
for the rst-order model can be followed to estimate the model
parameters n() and (). The corresponding steps are the
(1) Same step as in the rst-order model case.
(2) In this case,
f n,

1 n 2 n2 i2 n n


which, when decomposed into its real and imaginary parts,

f n,
Figure 3. Derivatives Re[(0, )] and Im[(0, )] of the
DEF (n, ) around steady-state condition n

mation. Finally, an estimate of the time constant is computed

with Eq. 28.
Figure 4 describes the function n() given by Eq. 28.
Observe that n() is a monotonous decreasing function. That
is, the faster the chemical reaction, the faster the reaction rate
dynamics of the catalyst particle. The slower dynamics corresponds to a purely diffusional process with limiting (normalized) time constant n(0) 1/3. Figure 5 shows the magnitudes
of the exact DEF (n, ) and the approximate rst-order
()/[n()n 1] complex functions. Acceptable approximation is observed for normalized frequencies n, which
corresponds to frequencies below the diffusion characteristic
frequency d. Some deviations are observed in the high-frequency range. It is also observed that the smaller the Thiele
modulus, the better the rst-order approximation. This is because for a small Thiele modulus, the whole catalyst space is
penetrated by diffusion molecules, yielding a uniform particle
reaction rate dynamics.

1 n 2 n2
2n n
n ,


where (n, ) [1 n()2 2n ] [2n()()n]2.

(3) The model-matching problems are obtained as follows:
Ren ,

1 n 2 n2
n ,



Second-Order Approximate Model

Second-order models can also be used, which are still simple
approximate models that do not introduce a serious increment
in the computational complexity. Assume a stable second-order
model parameterized as follows:
hs n,

2 2
n s n 2 n s n 1


December 2005

Figure 4. Estimated time constant for a rst-order

model from a model matching problem around
steady-state condition.
Vol. 51, No. 12

AIChE Journal

matched second-order model around steady-state conditions for

the reaction diffusion process does not exist.

Extension to the nonlinear case

Figure 5. Actual (sn, ) and approximate rst-order

{[ ()]/[n()sn 1]} dynamic EF.

Imn ,

2n n
n ,


(4) A second-order Taylor expansion of Eq. 33 around

steady-state condition n 0 is computed. In Figure 3 one can
observe that Re[(0, )] 0, which yields the approximation

Re0, n2 2n 2 n2


In this way, by matching second-order frequency power, the

following time-constant estimate is obtained:



Figure 3 shows that the second derivative Re[(0, )] 0.

This implies that n() as given by Eq. 36 is not well dened.
The above analysis is interesting as it shows the impossibility of computing a time constant for a stable second-order
model around steady state conditions. Although the parameter
of the second-order model Eq. 31 can be estimated by means of
other (e.g., optimization-based) methods, it was shown that a
AIChE Journal

In this work, we have considered reaction diffusion processes with linear reaction rates. This allowed a model-reduction methodology based on a transfer function; that is, the
effectiveness factor in dynamic conditions. Although by itself
the DEF concept introduced in the preceding sections can be of
interest to researchers working in reaction engineering, the
model-reduction problem for linear reaction rates is rarely of
practical importance for such applications as process control.
In fact, it is very likely that the reaction rates are nonlinear in
real processes. As described earlier, the model-reduction technique based on frequency-response tools cannot be used for
reaction diffusion processes with nonlinear reaction rates because a transfer function cannot be computed (only local transfer functions can be computed). To overcome this problem,
Szukiewicz11 proposed to linearly approximate the reaction
rate and compute a rst-order model reduction based on such
approximation. By using heuristic arguments, the nonlinear
nature of the reaction diffusion processes is subsequently recovered. A drawback of Szukiewiczs proposal is its lack of a
systematic procedure to evaluate the model-reduction error.
Moreover, high-order extensions to improve the approximation
are not reported.
Well-established model-reduction methods (for example,
Galerkin) for reaction diffusion equations are based on the
following observation. The concentration dynamics is driven
by diffusion mechanisms modeled as Dc, and the reaction
activity is described, in general, by a nonlinear function f (c).
The latter is a pointwise function, and so it does not introduce
spatial memory. The former, given its distributed nature,
involves an innite-dimensional operator, which introduces an
innite number of dynamic modes. Thus, a procedure aimed at
obtaining a nite-dimensional model should focus on the approximation of the diffusion operator. This idea has found
sound applications in the computation of nite-dimensional
controllers for the reaction diffusion processes by Galerkin
methods.6,8,9 Given the linear nature of the diffusion operator,
the same idea can be exploited to obtain reduced-order models
by classical response-frequency tools. The rationale behind this
proposal is that, as in Galerkin methods, a reduced-order model
retains the dominant modes of the diffusion operator.
In this way, a rst-order model contains an estimate of the
dominant time constant of the process dynamics that, for many
process applications, sufces to compute an industrial [such as
proportional integral (PI)] feedback controller.12
To illustrate these ideas in a preliminary way, consider a
reaction diffusion process with a nonlinear reaction rate

December 2005

Dc fc
with Dirichlet-type boundary conditions (see Eq. 9). In the
spirit of the Galerkin method, let us compute an approximation
to the distributed part of the above equation. To this end, given
the pointwise nature of f (c), consider the reactionless equation
c/t Dc and compute the corresponding transfer function
Vol. 51, No. 12


as done above. This is equivalent to taking (sn, 0). Subsequently, compute a rst-order approximation to give (0) 1
and n(0) 1/3 (see Figure 4). This implies that the dominant
time constant for the diffusion operator (in rectangular coordinates) is equal to 1/3. By reincorporating the nonlinear reaction
rate evaluated at average concentration conditions, the approximate rst-order model for the average concentration dynamics
c can be written as
dc 3D

c t c fc
Rp b
where cb(t) is the boundary concentration and Rp is the particle
radius. It is interesting to remark that for spherical coordinates
the dominant time constant is n(0) 1/15, which recovers the
result reported by Szukiewicz.11 One advantage of the approach described above is that high-order approximations can
be obtained by increasing the approximation order of the transfer function (sn,0). In such a case, the models can be expressed as in the following:
t fc
where (t) is the output of a linear system and (s) h(s)cb(s),
and h(s) is a rational approximation to the transfer function
h(s) : cb(s) 3 c (s) corresponding to the reactionless diffusion
It is interesting to note that when the method described
above is used in the linear case, the approximate dynamics are
given by

Figure 6. Comparison of the computed time constant

from three different methods.

to obtain reduced-order models, the issue of computing reduced-order models for nonlinear kinetics should be explored
in a subsequent work.

This section is used to illustrate the performance of the
model Eq. 23. For rectangular geometry, the dynamic reaction
diffusion problem can be written as

3c bt c 2c

c 2c

t n x 2

In this way, the estimated time constant is n () (3 2)1.

As 3 0, one obtains n (0) 1/3. Figure 6 compares this
simple approximation with that obtained with the more formal
method of the preceding section (see Figure 4). For all Thiele
moduli, the simple expression n () yields larger time-constant values. Because the time constant shown in Figure 4 was
computed from a matching problem around steady-state conditions, such a result could be expected, given that a smaller
time-constant value leads to faster dynamics convergence to
steady-state conditions. In fact, the approximation n () does
not take in account the effects of the reaction dynamics on the
diffusion dynamics. For completeness, Figure 6 also shows the
estimated time constant from Szukiewiczs methodology. Observe that the result is quite similar to the result obtained in this
work in that the former was also derived from a steady-state
approximation problem. Perhaps the main advantage of the
approach in this paper with respect to that of Szukiewicz is that
a formal framework to compute reduced-order models is used,
which allows an easier extension to more general situations
(such as computation of high-order approximations).
Summing up, some possible extensions to the nonlinear case,
of the model-reduction method described herein can be designed by following ideas from Galerkins methods. Given that
the main focus of this paper is on dening a DEF and its usage

December 2005


where x [0, 1]. The usual boundary conditions are

c x, t n c bt n
cx, tn

for x 1

for x 0

and the reaction rate provided by the catalyst particle is (see

Eq. 7)

Rt n Q 2

c x, t ndx


where Q is a constant ow rate. On the other hand, the

corresponding rst-order model (Eq. 23) can be written as (use
the identication k 3 2)
dRt n
n 1Q 2c bt n Rt n
dt n


Given the dimensionless description of Eqs. 37 40, only the

Thiele modulus has to be specied. Because both the exact and
Vol. 51, No. 12

AIChE Journal

Figure 7. Step response for unit step of the exact and

the approximate solutions for two different
values of the Thiele modulus.

the approximate models are linear, analytical solutions can be

obtained, although such solutions are input dependent. That is,
the analytical solution depends on the type of input (such as
steps, ramps, etc.) considered. For the sake of simplicity in
computations, we have used numerical methods to evaluate the
system response in the face of different inputs cb(tn). Specically, a nite-differences method with a mesh of size 200 was
used to reduce the partial differential equation problem of Eqs.
37 and 38 to a set of initial-value problems, which was integrated with a RungeKutta 4/5 integration method. Simpson
quadrature was used to evaluate the integral in Eq. 39. Finally,
a numerical solution to Eq. 40 was obtained with a Runge
Kutta 4/5 integration method. For all simulations, the value
A 1 was used.
Figure 7 shows the step response of the exact and approximate solutions for a step change from c b 0 to c b 1.0 at t
0.0. It is observed that the response of the exact solution is
faster than the approximate one. This is because n() is an
upper bound of the estimate of the dominant time constant of
the process. In fact, the dynamics of the exact reaction diffusion process contains innitely many dynamical modes whose
effects are not recovered by the simple rst-order model.
Figure 5 shows good agreement between the exact and approximate dynamics in the low-frequency range. This is reected in
the fact that, as time increases, the approximate solution is
closer to the exact one, as shown in Figure 6.
For 1.0, Figure 8 shows simulation results for an
oscillatory surface concentration of the form cb(t) 1.0 0.25
sin(tn). As expected from the results shown in Figure 5, good
agreement between the exact and the approximate solutions is
found for frequencies below the characteristic diffusion frequency d,n 1.0. In fact, for n 0.1 and n 1.0 small
deviations are observed during the transient dynamics. However, larger derivations (about 5%) are displayed for n 10.0,
where the important discrepancies in the frequency response
can be observed. In this form, the larger the oscillation freAIChE Journal

December 2005

Figure 9. Exact and quasi-steady-state solutions for

1.0 and n 0.75.

quency, the larger the approximation error. Similar results were

found for other Thiele modulus values.
Finally, let us illustrate the fact that (n, ) () for n
d,n 1.0. As noted earlier, this property yields the quasisteady-state consideration R(t) A () 2cb(tn). Figure 9
shows the good agreement between the exact and the approximate reaction rates for n 1.0. The plateau in the frequency
response of the DEF as shown in Figure 5 can be considered as
a support to the quasi-steady-state assumption commonly made
in practice for heterogeneous reactor models.
It should be mentioned that, given the closeness of the
different estimated time constants in Figure 6, minor differ-

Figure 8. Frequency response of the exact and the approximate solution for 1.0 and three different values of the oscillation frequency.
Vol. 51, No. 12


ences between the numerical results discussed above and numerical results from Szukiewiczs estimation were obtained.
Such differences are found in the high-frequency (such as
initial-conditions effect) regime.

methods, more sophisticated second- and higher-order model

structures should be explored in a future work. By following
some approximation ideas from Szukiewicz,11 extensions of
the low-order model methodology to nonlinear reactions rate
kinetics should be studied.

In this paper, the problem of representing the distributed
parameter dynamics of a catalyst particle by means of loworder (lumped parameter) models was addressed. As a preliminary step, an extension of the classical effectiveness factor
(EF) concept to dynamic conditions was introduced. To do this,
the EF was interpreted as an operator that transforms total
reaction rates from surface to particle conditions. Based on this
interpretation, a dynamic EF (DEF) was dened as a linear
operator that transforms total reaction signals from surface to
particle conditions. In this way, the DEF retains some information on the performance of a particle under dynamic perturbations. For common geometries, the computation of the DEF
can be easily obtained from the nondynamic expression by a
direct substitution of the classical Thiele modulus by a sort of
complex Thiele modulus. The DEF results showed that the
better catalyst performance is found at steady-state conditions;
that is, dynamic perturbations deteriorate the catalyst particle
By departing from the DEF concept, a systematic method to
obtain low-order models for the reaction dynamics was proposed. An algorithm to estimate the time constant for rstorder models was proposed by using model-matching methods.
The results showed that the larger the Thiele modulus, the
smaller the time constant. Although our analysis showed that a
stable second-order model cannot be computed from matching


December 2005

Literature Cited
1. Aris R. The Mathematical Theory of Diffusion and Reaction in Permeable Catalysts. Vol. 1. Oxford, UK: Clarendon Press; 1975.
2. Kim DH. Linear driving force formulas for diffusion and reaction in
porous catalyst. AIChE J. 1989;35:343-346.
3. Goto M, Hirose T. Approximate rate equation for intraparticle diffusion with and without reaction. Chem Eng Sci. 1993;48:1912-1915.
4. Szukiewicz MK. New approximate model for diffusion and reaction in
a porous catalyst. AIChE J. 2000;46:661-665.
5. Christodes PD, Daoutidis P. Feedback control of hyperbolic PDE
systems. AIChE J. 1996;42:3063-3086.
6. Christodes PD, Daoutidis P. Finite-dimensional control of parabolic
PDE systems using approximate inertial manifolds. J Math Anal Appl.
7. Sheng P, Zhou MY. Dynamic permeability in porous media. Phys Rev
Lett. 1988;61:1591-1594.
8. Baker J, Christodes PD. Finite-dimensional approximation and control of nonlinear parabolic PDE systems. Int J Contr. 2000;73:439456.
9. Christodes PD. Nonlinear and Robust Control of Partial Differential
Equation Systems: Methods and Applications to Transport-Reaction
Processes. Boston, MA: Birkhauser; 2001.
10. Chen CT. Introduction to Linear System Theory. New York, NY: Holt,
Rinehart & Winston; 1970.
11. Szukiewicz MK. An approximate model for diffusion and reaction in
a porous pellet. Chem Eng Sci. 2002;57:1451-1457.
12. Morari M, Zariou E. Robust Process Control. New York, NY:
Prentice-Hall; 1989.
Manuscript received Jan. 11, 2005, and revision received Apr. 4, 2005.

Vol. 51, No. 12

AIChE Journal

Two-Scale Continuum Model for Simulation of

Wormholes in Carbonate Acidization
Mohan K. R. Panga and Murtaza Ziauddin
Schlumberger, Sugar Land, TX 77478

Vemuri Balakotaiah
Dept. of Chemical Engineering, University of Houston, Houston, TX 77204
DOI 10.1002/aic.10574
Published online September 6, 2005 in Wiley InterScience (www.interscience.wiley.com).

A two-scale continuum model is developed to describe transport and reaction mechanisms in reactive dissolution of a porous medium, and used to study wormhole formation
during acid stimulation of carbonate cores. The model accounts for pore level physics by
coupling local pore-scale phenomena to macroscopic variables (Darcy velocity, pressure
and reactant cup-mixing concentration) through structure-property relationships (permeability-porosity, average pore size-porosity, and so on), and the dependence of mass
transfer and dispersion coefcients on evolving pore scale variables (average pore size
and local Reynolds and Schmidt numbers). The gradients in concentration at the pore
level caused by ow, species diffusion and chemical reaction are described using two
concentration variables and a local mass-transfer coefcient. Numerical simulations of
the model on a two-dimensional (2-D) domain show that the model captures the different
types of dissolution patterns observed in the experiments. A qualitative criterion for
wormhole formation is developed and it is given by O(1), where keffDeT/uo.
Here, keff is the effective volumetric dissolution rate constant, DeT is the transverse
dispersion coefcient, and uo is the injection velocity. The model is used to examine the
inuence of the level of dispersion, the heterogeneities present in the core, reaction
kinetics and mass transfer on wormhole formation. The model predictions are favorably
compared to laboratory data. 2005 American Institute of Chemical Engineers AIChE J, 51:
32313248, 2005

Keywords: wormholes, carbonate acidizing, reactive dissolution, porous media

Acid treatment of carbonate reservoirs is a widely practiced
oil and gas well stimulation technique. The primary objective
of this process is to increase the production rate by increasing
permeability of the damaged zone near the wellbore region.
The injected acid dissolves the material near the wellbore and
creates ow channels that establish a good connectivity beCorrespondence concerning this article should be addressed to V. Balakotaiah at

2005 American Institute of Chemical Engineers

AIChE Journal

December 2005

tween the reservoir and the well. While dissolution increases

permeability, the relative increase in permeability for a given
amount of acid is observed to be a strong function of the
injection conditions. At very low injection rates, acid is spent
soon after it contacts the medium resulting in face dissolution.
The penetration depth of the acid is restricted to a region very
close to the wellbore. On the other hand, at very high injection
rates, acid penetrates deep into the formation but the increase in
permeability is not large because the acid reacts over a large
region leading to uniform dissolution. At intermediate ow
rates, long conductive channels known as wormholes are
formed. These channels penetrate deep into the formation and
Vol. 51, No. 12


tion process, and the coupling between reaction and transport

mechanisms that plays a key role in reactive dissolution is not
completely accounted for in these models. In this work, we
present a two-scale continuum model that describes reaction
and transport in a porous medium retaining the essential physics of dissolution. The model is used to investigate the inuence of different factors on wormhole formation.

Literature Review

Figure 1. Typical dissolution patterns observed in carbonate acidizing: (a) face dissolution, Q 0.04
cc/min, PVinj 43.1, (b) conical Q 0.11 cc/
min, PVinj 10, (c) wormhole Q 1.05 cc/min,
PVinj 0.8, (d) ramied Q 10 cc/min, PVinj
2.1, and (e) uniform Q 60 cc/min, PVinj 6.7.
Hydrochloric acid is used in these experiments and the acid
injection rate is increased from (a) to (e) (Fredd and Fogler15).
The cores are approximately 3.8 cm in dia. and 10.2 cm in

facilitate the ow of oil. Thus, for successful stimulation of a

well it is required to produce wormholes with optimum density
and penetrating deep into the formation. A detailed description
of eld practices for carbonate acidizing can be found in the
literature.1 4
Several experimental studies have been conducted in the past
to understand wormhole formation and to predict the conditions required for creating wormholes.514 In those experiments, acid was injected into a core at different injection rates
and the volume of acid required to break through the core, also
known as breakthrough volume, was measured for each injection rate. A common observation in experimental studies is that
dissolution creates patterns that are dependent on the injection
rate. These dissolution patterns were broadly classied into
three types: uniform, wormholing and face dissolution patterns
corresponding to high, intermediate and low injection rates,
respectively. Figure 1 shows typical dissolution patterns observed in experiments15 on carbonate cores treated with HCl at
different injection rates. It is also observed that wormholes
form at an optimum injection rate, and because only a selective
portion of the core is dissolved the volume required to stimulate the core is minimized (see Figure 2). Furthermore, the
optimal conditions for wormhole formation were observed to
depend on various factors such as acid/mineral reaction kinetics, diffusion coefcients of the acid species, concentration of
acid, temperature, geometry of the system (linear/radial ow),
and so on. Many theoretical studies have been conducted in the
past to understand the phenomena of ow channeling associated with reactive dissolution in porous media, and to obtain an
estimate of the optimum injection rate. However, the models
developed thus far describe only a few aspects of the acidiza3232

December 2005

Wormhole formation during reactive dissolution of carbonates is a subject that has been actively studied in the last thirty
years. To explain wormhole formation, numerous models,
ranging from detailed pore-scale models (for example, network
models) that account for reaction, transport and dissolution at
the pore scale to single wormhole (tube) models that consider
only the mechanisms occurring inside the wormholes, have
been developed in the literature. In this section, a brief review
of different models developed to study wormhole formation in
carbonates is presented.
Relating the important dimensionless groups of the system to
experimental observations is one of the approaches followed to
model acidization process. For example, Fredd and Fogler12,15
and Fredd16 have reported the dependence of wormhole formation on the Damkohler number and predicted an optimum
Damkohler number of 0.29 for different uid/mineral systems.
Daccord et al.7,8,17 also used a dimensionless parameter based
approach and coupled it with the concept of fractals to obtain
the propagation rate of wormholes. The use of a few dimensionless groups to explain experimental observations is a difcult exercise because the actual number of parameters in the
system is large and it is difcult to study wormholing phenomena systematically using this approach.

Figure 2. Typical breakthrough curve

acidization experiments.



The pore volume of acid (HCl, 22C) required for breakthrough is high at very low and very high injection rates, and
is minimum at optimum injection rate, Q opt. The length and
diameter of the cores are 10.2 cm and 3.8 cm, respectively.
The initial porosities and permeabilities are in the range of
0.15 0.2 and 0.8 2 md, respectively.

Vol. 51, No. 12

AIChE Journal

In single wormhole models, a cylindrical tube represents the

wormhole and the effects of uid leakage, reaction kinetics,
and so on, on wormhole propagation rate are investigated using
these models.10,14,18,19,20,21 One of the key results from the
studies on single wormhole models is on the interaction of
wormholes and competition between them. It has been reported14,20 that the growth rates of multiple wormholes in a domain
is dependent on the separation distance between them. Panga et
al.22 used a continuum model and made a similar observation in
their numerical simulation of wormhole density. The simple
structure of single wormhole models offers the advantage of
studying reaction, diffusion and convection mechanisms inside
the wormhole in detail; however, these models cannot be used
to study wormhole initiation, the other dissolution patterns
(face, ramied, and so on) observed in the experiments, and the
effect of heterogeneities on wormhole formation.
Hoefner and Fogler,9 Fredd and Fogler,12 and Daccord et
al. have developed network models to describe reactive dissolution. Network models represent the porous medium as a
network of tubes interconnected to each other at the nodes. The
ow of acid inside these tubes is described using the HagenPoiseuille relationship for laminar ow. The acid reacts at the
wall of the tube and dissolution is accounted for in terms of an
increase in the tube radius. Network models predict the dissolution patterns and qualitative features of dissolution like the
optimum ow rate observed in experiments. However, a core
scale simulation of the network model is computationally very
An intermediate approach to describing reactive dissolution involves the use of averaged or continuum models.
Unlike the network models that describe dissolution from
the pore scale, and the models based on the assumption of
existing wormholes, the averaged models describe dissolution at the Darcy scale. The Darcy scale model requires
information on the pore scale processes, which are obtained
from a pore scale model. The predictions of the pore scale
model depend on the pore structure that changes with time
due to dissolution. Obtaining detailed pore structure of a
core and approximating its change during dissolution is very
difcult and is one of the disadvantages of using a Darcy
scale model. However, by using different pore scale models
that are representative of the core, the sensitivity of the
results obtained from Darcy scale models can be studied.
Averaged models for carbonate acidization have been developed by Liu et al.23 Chen et al.24,25 and Goler et al.26
These models were shown to capture qualitative and some
quantitative features of dissolution. The model developed by
Liu et al. and Chen et al. does not consider the effect of mass
transfer on the reaction rate and is valid only in the kinetic
regime, while the model developed by Goler et al. is valid
only in the mass-transfer controlled regime (to be dened
later). This work presents a continuum model that captures
both the extremes of reaction (kinetic and mass transfer
controlled) simultaneously by using two concentration variables and a mass-transfer coefcient. This allows the description of a wide range of acids as demonstrated later. The
model proposed here is similar to the widely used two-phase
models of catalytic reactors,27 the main difference being
coupling of the ow, dissolution/reaction and pore scale
mass-transfer processes.
AIChE Journal

December 2005

Figure 3. Different length scales used in the model.

Model for Carbonate Dissolution

Reaction between a carbonate porous medium and acid
leads to complete dissolution of the medium, thereby increasing the permeability to a large value. At very low
injection rates in a homogeneous medium, this would give
rise to a planar reaction/dissolution front where the medium
behind the front is completely dissolved and the medium
ahead of the front is not dissolved. However, the presence of
natural heterogeneities in the medium leads to an uneven
increase in permeability along the front leading to regions of
high and low permeabilities. The high permeability regions
attract more acid, which further dissolves the medium creating channels that travel ahead of the front. Thus, adverse
mobility (K/, where K is the permeability and is the
viscosity of the uid) arising because of differences in the
permeabilities of the dissolved and undissolved medium and
heterogeneity are required for channel formation. This reaction-driven instability has been studied using linear and
weakly nonlinear stability analyses by some authors.28,29,30,31 This instability is similar to the viscous ngering instability where adverse mobility arises due to a
difference in viscosities of the displacing and displaced
The shape (wormhole, conical, and so on) of the channels is,
however, dependent on the relative magnitudes of convection
and dispersion in the medium. For example, when transverse
dispersion is more dominant than convective transport, reaction
leads to conical and face dissolution patterns. On the other
hand, when convective transport is more dominant, the concentration of acid is more uniform in the domain leading to a
uniform dissolution pattern. The model presented here describes the phenomena of reactive dissolution as a coupling
between processes occurring at two scales, namely the Darcy
scale and the pore scale. Different length scales are shown in
Figure 3. In the following subsections, these two parts of the
model are discussed.
Vol. 51, No. 12


Darcy scale model

The Darcy scale model equations are given by




C f
UCf De Cf kc a Cf Cs (3)
k cC f C s RC s
RC sa





Here U (U, V, W) is the Darcy velocity vector, K is the

permeability tensor, P is the pressure, is the porosity, C f is
the cup-mixing concentration of the acid in the uid phase, C s
is the concentration of the acid at the uid-solid interface, De is
the effective dispersion tensor, k c is the local mass-transfer
coefcient, a is the interfacial area available for reaction per
unit volume of the medium, s is the density of the solid phase
and is the dissolving power of the acid, dened as grams of
solid dissolved per mole of acid reacted. The reaction kinetics
is represented by R(C s ). For a rst-order reaction R(C s ) reduces to k s C s where k s is the surface reaction rate constant
(having the units of velocity).
Equation 3 gives a Darcy scale description of the transport of
the acid species. The rst three terms in the equation represent
accumulation, convection and dispersion of the acid, respectively. The fourth term describes the transfer of acid species
from the uid phase to the uid-solid interface and its role is
discussed later in this section. The velocity eld U in the
convection term is obtained from Darcys law (Eq. 1) relating
velocity to the permeability eld K and the gradient of pressure. The rst term in the continuity Eq. 2 accounts for the
effect of local volume change during dissolution on the ow
eld. While deriving the continuity equation, it is assumed that
the dissolution process does not change the density of the uid
phase signicantly.
The transfer term in the species balance Eq. 3 describes the
depletion of the reactant at the Darcy scale due to reaction. An
accurate estimation of this term depends on the description of
transport and reaction mechanisms inside the pores. In the
absence of reaction, concentration of the acid species is uniform inside the pores. Reaction at the solid-uid interface gives
rise to concentration gradients in the uid phase inside the
pores. The magnitude of these gradients depends on the relative
rate of mass transfer from the uid phase to the uid-solid
interface and reaction at the interface. If the reaction rate is
very slow compared to the mass-transfer rate, the concentration
gradients are negligible. In this case, the reaction is considered
to be in the kinetically controlled regime, and a single concentration variable is sufcient to describe this situation. However,
if the reaction is very fast compared to the mass transfer, steep
gradients develop inside the pores. This regime of reaction is

known as the mass-transfer controlled regime. To account for

the gradients developed due to mass transfer control requires
the solution of a differential equation describing diffusion and
reaction mechanisms inside each of the pores. Since this is
impractical, we use two concentration variables C s and C f , for
the concentration of the acid at the uid-solid interface and in
the uid phase respectively, and capture the information contained in the local concentration gradients as a difference
between the two variables using the concept of a local masstransfer coefcient (Eq. 4).
Equation 4 balances the amount of reactant transferred to the
surface to the amount reacted. For the case of rst order
kinetics, R(C s ) k s C s , Eq. 4 can be simplied to

December 2005




In the kinetically controlled regime (k s k c ), the concentration at the uid-solid interface is approximately equal to the
concentration of the uid phase (C s C f ). In the mass transfer
controlled regime (k s k c ), the value of concentration at the
uid-solid interface is very small (C s 0). Since the rate
constant is xed for a given acid, the magnitude of the ratio
k s /k c is determined by the local mass-transfer coefcient k c ,
which is a function of the pore geometry, the reaction rate and
the local hydrodynamics. Due to dissolution and heterogeneity
in the medium, the ratio k s /k c is not a constant in the medium
but varies with space and time which can lead to a situation
where different locations in the medium experience different
regimes of reaction. To describe such a situation it is essential
to account for both kinetic and mass transfer controlled regimes
in the model. Equation 5 describes the evolution of porosity in
the domain due to reaction.
To complete the model (Eqs. 15), information on the permeability tensor K, dispersion tensor De , mass-transfer coefcient k c , and interfacial area a is required. These quantities
depend on the pore structure and are inputs to the Darcy scale
model from the pore scale model. Instead of calculating these
quantities from a detailed pore scale model taking into consideration the actual pore structure, we use structure-property
relations that relate permeability, interfacial area and average
pore radius of the pore scale model to its porosity. However, if
a detailed calculation including the pore structure can be made,
then the quantities K, De , k c , and a obtained from such a
calculation can be used as inputs from the pore scale model to
the Darcy scale model.

Pore scale model

(a) Structure-Property Relations. Dissolution changes the
structure of the porous medium continuously, thus, making it
difcult to correlate the changes in local permeability to porosity during acidization. The results obtained from averaged
models, which use these correlations, are subject to quantitative
errors arising from the use of a bad correlation between the
structure and property of the medium, although the qualitative
trends predicted may be correct. Since a denitive way of
relating the change in the properties of the medium to the
change in structure during dissolution does not exist, we use
Vol. 51, No. 12

AIChE Journal

gime of reaction for a given acid (Eq. 6). The mass-transfer

coefcient depends on the local pore structure, reaction rate
and local velocity of the uid. Gupta and Balakotaiah33 and
Balakotaiah and West34 investigate the contribution of each of
these factors to the local mass-transfer coefcient in detail. For
developing ow inside a straight pore of arbitrary cross section,
it is shown34 that a good approximation to the Sherwood
number, the dimensionless mass-transfer coefcient, is given

Figure 4. Variation of permeability with porosity for different values of .

semiempirical relations that relate the properties to local porosity. The relative increase in permeability, pore radius and
interfacial area with respect to their initial values are related to
porosity in the following manner

1 o

K o o o1


K o
K o



a r o

a o or p


Here K o , r o and a o are the initial values of permeability,

average pore radius and interfacial area, respectively. Notice
that, for 1 the structure-property relation for permeability
evolution reduces to the well-known Carman-Kozeny correlation K 3 /(1 ) 2 relating permeability to the porosity of
the medium. The parameter is introduced into Eq. 7 to extend
the relation to a dissolving medium. Figure 4 shows a typical
plot of permeability versus porosity for different values of the
parameter . In addition, the effect of structure-property relations on breakthrough time has also been tested by using
different correlations (in a later section). The model yields
better results if structure-property correlations that are developed for a particular system of interest are used. Note that, in
the earlier relations permeability, which is a tensor, is reduced
to a scalar for the pore scale model. In the case of anisotropic
permeability, extra relations for the permeability of the pore
scale model are needed to complete the model.
(b) Mass-Transfer Coefcient. Transport of the acid species from the uid phase to the uid-solid interface inside the
pores is quantied by the mass-transfer coefcient (k c ). It
plays an important role in characterizing dissolution phenomena because the mass-transfer coefcient determines the reAIChE Journal

December 2005

2k cr p
Sh 0.35


Rep1/ 2 Sc1/3


where k c is the mass-transfer coefcient, r p is the pore radius

and D m is molecular diffusivity, Sh is the asymptotic Sherwood number for the pore, Rep is the pore Reynolds number,
d h is the pore hydraulic diameter, x is the distance from the
pore inlet and Sc is the Schmidt number (Sc /D m ; where
is the kinematic viscosity of the uid). Assuming that the
length of a pore is typically a few pore diameters, the average
mass-transfer coefcient can be obtained by integrating the
above expression over a pore length, which gives
Sh Sh b Rep1/ 2 Sc1/3


where the constants Sh and b (0.7/m1/2, m pore length to

diameter ratio) depend on the structure of the porous medium
(pore cross sectional shape and pore length to hydraulic diameter ratio). Equation 11 is of the same general form as the
Frossling correlation used extensively in correlating masstransfer coefcients in packed-beds. (For a packed bed of
spheres, Sh 2 and b 0.6. This value of b is close to the
theoretical value of 0.7 predicted by Eq. 11 for m 1.)
The two terms on the righthand side of Eq. 11 are contributions to the Sherwood number due to diffusion and convection
of the acid species, respectively. While the diffusive part, Sh
depends on the pore geometry, the convective part is a function
of the local velocity. The asymptotic Sherwood number for
pores with cross-sectional shape of square, triangle and circle
are 2.98, 2.50 and 3.66, respectively.34 Since the value of
asymptotic Sherwood number is a weak function of the pore
cross-sectional geometry, we use a typical value of 3.0 in our
calculations. The convective part depends on the pore Reynolds number and the Schmidt number. For liquids, the typical
value of Schmidt number is around one thousand and assuming
a value of 0.7 for b, the approximate magnitude of the con2
vective part of the Sherwood number from Eq. 11 is 7 Re1/
p .
The pore Reynolds numbers are very small due to the small
pore radius, and the low injection velocities of the acid, making
the contribution of the convective part negligible during the
initial stages of dissolution. As dissolution proceeds, the pore
radius and the local velocity increase, making the convective
contribution signicant. (Remark: For typical values cited
above, the two contributions are equal when 7 Re1/
3 or
Rep 0.2.) However, the effect of this convective mass
transfer on acid consumption may not be signicant because of
the extremely low interfacial area in the high porosity regions
where convection is dominant. The effect of reaction rate on
the mass-transfer coefcient is observed to be weak.34 For
example, the asymptotic Sherwood number varies from 48/11
Vol. 51, No. 12


(4.36) to 3.66 in the case of a very slow reaction to a very fast

(c) Fluid Phase Dispersion Coefcients. For homogeneous, isotropic porous media, the dispersion tensor is characterized by two independent components, namely, the longitudinal, D eX and transverse, D eT , dispersion coefcients. In the
absence of ow, dispersion of a solute occurs only due to
molecular diffusion and D eX D eT os D m , where D m is
the molecular diffusion coefcient and os is a constant that
depends on the structure of the porous medium (for example,
tortuosity or connectivity between the pores). With ow, the
dispersion tensor depends on the morphology of the porous
medium, as well as the pore level ow and uid properties. The
relative importance of convective to diffusive transport at the
pore level is characterized by the Peclet number in the pore,
dened by
Pe p



where U is the magnitude of the Darcy velocity, and d h is the

pore hydraulic diameter. For a well-connected pore network,
random walk models and analogy with packed beds may be
used to show that

D eX
os XPe p



D eT
os TPe p



where X and T are numerical coefcients that depend on the

structure of the medium ( X 0.5, T 0.1 for packed-bed
of spheres). Other correlations used for D eX , are of the form

D eX
os Pe pln


December 2005

7 104
7 103

1.4 10
1.4 103

Dimensionless Model Equations and Limiting

The model equations for rst-order irreversible kinetics are
made dimensionless for the case of constant injection rate at the
inlet boundary by dening the following dimensionless variables







2k sr o






Equation 16 is based on Taylor-Aris theory, and it is normally

used when the connectivity between the pores is very low.
These as well as the other correlations in the literature predict
that both the longitudinal and transverse dispersion coefcients
increase with the Peclet number.
Table 1 shows typical values of pore Peclet numbers calculated, based on the core experiments (permeability of the cores
is approximately 1 mD) of Fredd and Fogler.12 The injection
velocities of the acid (0.5 M HCl) are varied between 0.14 cm/s
and 1.4 104 cm/s, where 0.14 cm/s corresponds to the
uniform dissolution regime and 1.4 104 cm/s corresponds
to the face dissolution regime. The values of pore diameter,
molecular diffusion and porosity used in the calculations are

Pe p

0.1 m, 2 105 cm2/s and 0.2, respectively. It appears from

the low values of pore level Peclet number in the face dissolution regime that dispersion in this regime is primarily due to
molecular diffusion. The Peclet number is close to order unity
in the uniform dissolution regime showing that both molecular
and convective contributions are of equal order. In the numerical simulations presented later, it is observed that the dispersion term in Eq. 3 does not play a signicant role at high
injection rates (uniform dissolution regime) where convection
is the dominant mechanism. As a result, the form of the
convective part of the dispersion ( X Pe p , Pe p ln(3Pe/ 2), and
so on), which becomes important in the uniform dissolution
regime, may not affect the breakthrough times at low permeabilities. In this work, we use the dispersion relations given by
Eqs. 13 and 14 to complete the averaged model. (Remark: As
a rst approximation, we have assumed that the mass transfer
coefcient to be the same in the axial and transverse directions.
However, as in the case of the dispersion coefcient, the
convective contribution to mass transfer could be different in
the ow and transverse directions. This can be accounted for by
replacing the scalar k c by the transfer matrix (a tensor). We do
not pursue this here.)

D eX
os Pe p2.

Injection Velocity (cm/s)





Table 1. Pore Level Peclet Numbers at Different

Injection Rates

k sa oL


N ac
Pe L



P Pe
u oL


u oL

2r o


where L is the characteristic length scale in the (ow) x

direction, H is the height of the domain, u o is the inlet velocity,
C o is the inlet concentration of the acid and P e is the pressure
at the exit boundary of the domain. The initial values of
permeability, interfacial area and average pore radius are repVol. 51, No. 12

AIChE Journal

resented by K o , a o and r o , respectively. The parameters obtained after making the equations dimensionless are the Thiele
modulus 2, the Damkohler number Da, the acid capacity
number N ac , the axial Peclet number Pe L , aspect ratio o , and
. The Thiele modulus (2) is dened as the ratio of diffusion
time to reaction time based on the initial pore size and the
Damkohler number (Da) is dened as the ratio of convective
time to reaction time based on the length scale of the core. The
acid capacity number (N ac ) is dened as the volume of solid
dissolved per unit volume of the acid, and the axial Peclet
number Pe L is the ratio of axial diffusion time to convection
time. Notice that in the above parameters, inlet velocity u o
appears in two parameters Da and Pe L . To eliminate inlet
velocity from one of the parameters, so that the variable of
interest, that is, injection velocity, appears in only one dimensionless parameter (Da) we introduce a macroscopic Thiele
modulus 2 which is dened as 2 k s a o L 2 /D m DaPe L .
The macroscopic Thiele modulus is a core scale equivalent of
the pore scale Thiele modulus (2), and is independent of
injection velocity. The dimensionless equations in the twodimensional (2-D) case are given by




t x y

DaA cf
c f uc f c f

2 r

X ur



os Da

os Da
T ur

DaN acA c f







0 @ y 0, 0


cf 1 @ x 0


c f
AIChE Journal


December 2005

and y o


cf 0 @ t 0


x, y, a, l 0 f a,l @ t 0


A constant injection rate boundary condition given by Eq. 21

is imposed at the inlet of the domain and the uid is contained
in the domain by imposing zero ux boundary conditions (Eq.
23) on the lateral sides of the domain. The boundary conditions
for the transport of acid species are given by Eqs. 24 26. It is
assumed that there is no acid present in the domain at time t
0. To simulate wormhole formation numerically, it is necessary
to have heterogeneity in the domain which is introduced by
assigning different porosity values to different grid cells in the
domain according to Eq. 28. The porosity values are generated
by adding a random number (f) uniformly distributed in the
interval [ o , o ] to the mean value of porosity o . The
quantity a dened as a o / o is the magnitude of heterogeneity, and the parameter l is the dimensionless length scale
of heterogeneity which is scaled using the pore diameter that is,
l L h /(2r o ) L h /( L), where L h is equal to the length scale
of the heterogeneity. Unless stated otherwise, L h is taken as the
size of the grid in numerical simulations.
The earlier system of equations can be reduced to a simple
form at very high or very low injection rates to obtain analytical relations for pore volumes required to breakthrough. Face
dissolution occurs at very low injection rates where the acid is
consumed as soon as it comes in contact with the medium. As
a result, the acid has to dissolve the entire medium before it
reaches the exit for breakthrough. The stoichiometric pore
volume of acid required to dissolve the whole medium is given


The boundary and initial conditions used to solve the system of

equations are given below

c f

PV FaceD

s1 o 1 o

C o o
N ac o


where C o is the inlet concentration of the acid, and o is the

initial porosity of the medium. At very high injection rates, the
residence time of the acid is very small compared to the
reaction time and most of the acid escapes the medium without
reacting (Remark: It is difcult to reach this asymptote experimentally). Because the conversion of the acid is low, the
concentration in the medium could be approximated as the inlet
concentration. Under these assumptions the model could be
reduced to

k sC oa
s 1


Denoting the nal porosity required to achieve a xed increase

in the permeability by f (this could be calculated from Eq. 7),
the earlier equation can be integrated for the breakthrough time
Vol. 51, No. 12


t bth

k sC o



Thus, the pore volume of acid required for breakthrough at

high injection rates is given by

PV UniformD

t bthu o
su o

k sC o a o oL


DaN ac o


The breakthrough volume increases with increasing velocity.

To achieve a xed increase in the permeability, a large
volume of acid is required in the uniform dissolution regime
where the acid escapes the medium after partial reaction.
Similarly, in the face dissolution regime a large volume of acid
is required to dissolve the entire medium. In the wormholing
regime only a part of the medium is dissolved to increase the
permeability by a given factor, thus, decreasing the volume of
acid required than that in the face and uniform dissolution
regimes. Since spatial gradients do not appear in the asymptotic
limits (Eq. 29 and Eq. 30) the results obtained from 1-D, 2-D
and 3-D models for pore volume of acid required to achieve
breakthrough should be independent of the dimension of the
model at very low and very high injection rates for a given
acid. However, optimum injection rate and minimum volume
of acid that arise due to channeling are dependent on the
dimension of the model. The pore volume required for breakthrough vs the injection rate is shown in Figure 5 for 1-D, 2-D
and 3-D models.

2-D Dissolution Patterns

In this section, numerical simulations illustrating the effect
of heterogeneity, different transport mechanisms and reaction
kinetics on dissolution patterns are presented. The model is
simulated on a rectangular 2-D porous medium of dimensions
2 cm 5 cm ( o 0.4). Acid is injected at a constant rate at
the inlet boundary of the domain, and it is contained in the
domain by imposing a zero ux boundary condition on the
lateral sides of the domain. The simulation is stopped once the
acid breaks through the exit boundary of the domain. Here,
breakthrough is dened as decrease in the pressure drop by a
factor of 100 (or increase in the overall permeability of the
medium by 100) from the initial pressure drop. The numerical
scheme used to solve the governing equations is described in
Appendix-A. The value of initial porosity in the domain is 0.2.
The effect of injection rate on the dissolution patterns is studied
by varying the Damkohler number (Da), which is inversely
proportional to the velocity. In addition to the dimensionless
injection rate (Da), the other important dimensionless parameters in the model are 2 , N ac , 2 , a and l. The effect of these
parameters on wormhole formation is investigated.
(a) Magnitude of Heterogeneity. As discussed earlier, het3238

December 2005

Figure 5. Qualitative trends in breakthrough curves for

1-D, 2-D and 3-D models.
The optimum injection rate and the minimum pore volume
decrease from 1-D to 3-D due to channeling.

erogeneity is an important factor that promotes pattern formation during reactive dissolution. Without heterogeneity, the
reaction/dissolution fronts would be uniform despite an adverse
mobility ratio between the dissolved and undissolved media. In
a real porous medium, the presence of natural heterogeneities
triggers instability leading to different dissolution patterns. To
simulate these patterns numerically, it is necessary to introduce
heterogeneity into the model. Heterogeneity could be introduced in the model as a perturbation in concentration at the
inlet boundary of the domain or as a perturbation in the initial
porosity or permeability eld in the domain. In the present
model, heterogeneity is introduced into the domain as a random
uctuation of initial porosity values about the mean value of
porosity as given by Eq. 28. The two important parameters
dening heterogeneity are the magnitude of heterogeneity a,
and the dimensionless length scale l. The effect of these
parameters on wormhole formation is investigated in this, and
the following subsections.
The inuence of the magnitude of heterogeneity (a) is
studied by maintaining the length scale of heterogeneity constant (which is the grid size) and varying the magnitude from
a small to a large value. Figures 6a 6e show the porosity
proles of numerically simulated dissolution patterns for different Damkohler numbers on a domain with a large magnitude
of heterogeneity in initial porosity distribution. The uctuations (f) in porosity ( 0.2 f ) are uniformly distributed in
the interval [0.15, 0.15] (a 0.75). Figures 6f 6j show
the porosity proles for the same Damkohler numbers used in
Figures 6a 6e, but with a small magnitude of heterogeneity in
the initial porosity distribution. The uctuations (f) in porosity
( 0.2 f ) for this case are distributed in the interval
[0.05, 0.05] (a 0.25). It could be observed from the
gures that wormholes do not exhibit branching when the
magnitude of heterogeneity is decreased. This observation suggests that branching of wormholes observed in carbonate cores
Vol. 51, No. 12

AIChE Journal

Figure 6. Porosity proles at different Damkohler numbers with uctuations in initial porosity distribution in the interval [0.15, 0.15] are shown in
Figures a e.
Figures fj show porosity proles for the same Damkohler
numbers as used in Figures a e, but for uctuations in the
interval [0.05, 0.05]. The values of Damkohler numbers for
different patterns are: (a) Da 3 10 4 ( o 30), (b)
Da 10 4 ( o 10), (c) Da 500 ( o 0.5), (d) Da
40 ( o 0.04), and (e) Da 1 ( o 0.001). The values
of parameter o are reported for later use. The values of other
parameters xed in the model are 2 10 6 , 2 0.07,
N ac 0.1, o 0.4.

could be a result of a wide variation in magnitude of heterogeneities present in the core. Figures 6a 6j show that at very
large Damkohler numbers (low injection rates), the acid reacts
soon after it contacts the medium resulting in face dissolution,
and at low values of Damkohler number (high injection rates),
acid produces a uniform dissolution pattern. Wormholing patterns are created near intermediate/optimum values of the
Damkohler number. While changing the magnitude of heterogeneity changes the structure of the wormholes, an important
observation to be made here is that the type of dissolution
pattern (wormhole, conical, and so on) remains the same at a
given Damkohler number for different magnitudes of heterogeneity. Thus, heterogeneity is required to trigger the instability and its magnitude determines wormhole structure, but the
type of dissolution pattern formed is governed by the transport
and reaction mechanisms. Figure 7 shows the pore volume of
acid required to breakthrough the core at different injection
rates with different levels of heterogeneity for the porosity
proles shown in Figure 6. The curves show a minimum at
intermediate injection rates because of wormhole formation. It
could be observed from the breakthrough curves that the minimum pore volume/breakthrough time and optimum injection
rate (Damkohler number) are approximately the same for both
levels of heterogeneity.
The second parameter related to heterogeneity that is introduced in the model is the length scale of heterogeneity l. The
effect of this parameter on wormhole structure is dependent on
the relative magnitudes of convection, reaction and dispersion
levels in the system. The role of this parameter on wormhole
formation is thus discussed after investigating the effects of
convection, reaction and transverse dispersion in the system.
AIChE Journal

December 2005

Figure 7. Breakthrough curves for different magnitudes

of heterogeneity used in Figure 6.
The volume of acid required to breakthrough the core at
different injection rates is not a strong function of the magnitude of heterogeneity.

(b) Convection and Transverse Dispersion. In the previous subsection, it was shown that the magnitude of heterogeneity affects wormhole structure but its inuence on optimum
Damkohler number is not signicant. The dissolution pattern
produced is observed to depend on the relative magnitudes of
convection, reaction and dispersion in the system. Because of
the large variation in injection velocities (over three orders of
magnitude) in core experiments, different transport mechanisms become important at different injection velocities, each
leading to a different dissolution pattern. For example, at high
injection velocities convection is more dominant than dispersion and it leads to uniform dissolution, whereas at low injection velocities dispersion is more dominant than convection
leading to face dissolution. A balance between convection,
reaction and dispersion levels in the system produces worm-

Figure 8. Reaction front thickness in the longitudinal

and transverse directions to the mean ow.
The front thickness is represented by l X in the ow direction
and l T in the transverse direction.

Vol. 51, No. 12


holes. In this subsection, a qualitative analysis is rst presented

to identify the important parameters that determine the optimum velocity for wormhole formation and the minimum pore
volume of acid. Numerical simulations are performed to show
the relevance of these parameters.
Consider a channel in a porous medium (see Figure 8)
created because of reactive dissolution of the medium. The
injected acid reacts in the medium ahead of the tip and adjacent
to the walls of the channel and increases the length, as well as
the width of the channel. If the growth of the channel in the
direction of ow is faster than its growth in the transverse
direction then the resulting shape of the channel is thin and is
called a wormhole. On the other hand, if the growth is much
faster in the transverse direction compared to the ow direction
then the channel assumes a conical shape. To nd the relative
growth in each direction, it is necessary to identify the dominant mechanisms by which acid is transported in the direction
of ow and transverse to the ow. Because of a relatively large
pressure gradient in the ow direction, the main mode of
transport in this direction is convection. In the transverse
direction, convective velocities are small and the main mode of
transport is through dispersion. If we denote the length of the
front in the medium ahead of the tip where the acid is consumed by l X , and the front length in the transverse direction by
l T , a qualitative criterion for different dissolution patterns can
be given by

In a similar fashion, the length scale l T in the transverse

direction is given by the dispersion-reaction equation
D eT

where D eT is the effective transverse dispersion coefcient.

The length scale l T in the transverse direction is given by


0.1 to 1 f Wormhole


O1 f uniform dissolution



C f
keff Cf


where u tip is the velocity of the uid at the tip of the wormhole
and k eff is an effective rate constant dened as

D eT
k eff


l T k effDeT



The qualitative criteria for different channel shapes in Eqs.

3234 in terms of parameter are given by O(1) for
face dissolution, 0.1 to 1 for wormhole formation and
O(1) for uniform dissolution. The parameter

k ck s
k s k c eT
u tip


used for determining the conditions for wormhole formation

includes the effect of transverse dispersion through D eT , reaction rate constant k s , pore-scale mass-transfer coefcient k c ,
structure-property relations through a , effect of convection
through velocity u tip , and is independent of domain length L. It
should be noted that the above quantities change with time and,
thus, provides only an approximate measure for wormhole
formation but it is an important parameter to study wormholing. For the case of mass transfer controlled reactions, the
parameter reduces to k c a D eT /u tip , while for kinetically controlled reactions it reduces to k s a D eT /u tip.
The optimum injection velocity

An approximate magnitude of l X can be obtained from the

convection-reaction equation

u tip

The ratio of transverse to axial length scales is given by

O1 f Face dissolution

2C f
k effCf
y 2

u opt

k ck s
a D k effD eT
k s k c eT


scales as square root of effective rate constant and transverse

dispersion coefcient. The parameter in Eq. 39 can be
written in terms of dimensionless parameters Damkohler number Da and Peclet number Pe L as


k eff
k sa k ca

Pe L

Thus, the length scale over which the acid is consumed in the
ow direction is given by



A D T 1/ 2M

o A DT 1/ 2 M



u tip

December 2005

M u o/u tip
Vol. 51, No. 12

AIChE Journal

Again, for kinetically controlled reactions, the above equation

reduces to

lT 1 DT


Figure 9. Plot shows pore volume of acid required to

breakthrough versus the parameter 1
different values of macroscopic Thiele modulus 2.
The reaction is in the kinetic regime (2 0.07). Minimum
acid volume is required for injection rates where o O(1).
This minimum volume of acid required decreases with increasing values of 2.

For clarity, only kinetically controlled reactions ( 2 r/Sh

1) are analyzed in the remainder of this subsection. The analysis of mass transfer controlled reactions ( 2 r/Sh 1) is
presented in the next section. For kinetically controlled reactions, o can be reduced to

Da Da

Pe L

k sa oD m
u o2


It is observed in the numerical simulations that o 0.1 to 1

gives a good rst approximation to wormhole formation criterion in Eq. 41. Figure 6 shows the values of o for different
dissolution patterns in the kinetic regime. From the gure it can
be seen that wormholing patterns occur at o 0.5 as
indicated by the scaling o O(1). For small values of o
(0.001), uniform dissolution is observed and for large values
of o (30), face dissolution is observed. The value of the
parameter o gives an estimate of the optimum injection velocity. The minimum pore volume required for breakthrough,
however, depends on the diameter of the wormhole because the
volume of acid required to dissolve the material in the wormhole decreases as the wormhole diameter decreases. Since the
diameter of the wormhole depends on the thickness of the front
l T in the transverse direction, it is necessary to identify the
parameter that controls the transverse front thickness. The
parameter that determines the front thickness can be obtained
from Eq. 37


AIChE Journal

D eT

k effL2



1/ 2

December 2005

1/ 2


From Eq. 45, it can be seen that the front thickness or the
wormhole diameter is inversely proportional to the square root
of macroscopic Thiele modulus 2. Thus, for increasing values
of macroscopic Thiele modulus (or decreasing levels of dispersion), the diameter of the wormhole decreases, thereby
decreasing the minimum pore volume required to breakthrough. Figure 9 shows pore volume of acid required for
breakthrough versus reciprocal of the parameter o for three
different values of 2 for a kinetically controlled reaction
(2 0.07). The minimum pore volume required to breakthrough decreases with increasing values of macroscopic
Thiele modulus 2. Figure 10 shows the nal porosity proles
at the optimum injection rate in Figure 9 for different values of
macroscopic Thiele modulus. It can be seen from the gure that
the wormhole diameter decreases with increasing values of 2.
The earlier analysis shows that optimum injection rate and
minimum pore volume required for breakthrough are determined by o and macroscopic Thiele modulus 2. The breakthrough curves in Figure 9 are plotted again with respect to
Damkohler number Da in Figure 11 for different values of
macroscopic Thiele modulus 2. It can be seen from the gure
that the optimum Damkohler number is dependent on the value
of 2. Thus, changing the value of 2 changes the optimum
Damkohler number whereas the parameter is always of order
unity for different values of 2 (Figure 9). We believe that
gives a better criterion than the optimum Damkohler number
for predicting wormhole formation. As shown in Figure 11, 2
does not affect the number of pore volumes required to breakthrough in the high injection rate regime. This is because
dispersion effects are negligible at high injection rates, where
convection and reaction are the dominant mechanisms. The
slope of the breakthrough curve at low injection rates and the
minimum pore volume are dependent on the value of 2
showing that dispersion becomes an important mechanism at
lower injection rates where wormholing, conical and face dissolution occur. The breakthrough curve for 2 104 shows a
minimum pore volume that is higher than that required for

Figure 10. Porosity proles at the optimum injection

rate for the breakthrough curves shown in
Figure 9 for different values of 2.
The values of 2 in the gures are: (a) 2 104 (b) 2
105 (c) 2 106. The gures show the effect of macroscopic Thiele modulus 2 on the wormhole diameter. [Color
gure can be viewed in the online issue, which is available
at http://www.interscience.wiley.com.]

Vol. 51, No. 12


sion and reaction are the dominant mechanisms leading to

face dissolution.
(c) Reaction Regime. The magnitude of 2 r/Sh or k s /k c in
the denominator of the local equation


Figure 11. Breakthrough curves in Figure 9 are plotted

as function of the reciprocal of Damkohler
The gures show that optimum Damkohler number depends
on 2. The breakthrough curves are independent of 2 in
the high injection rate asymptote because dispersion effects
are negligible compared to convection and reaction.

larger values of 2, and it also reaches the low injection rate

asymptote at injection rates higher than that required for larger
values of 2. This is due to high dispersion level in the system
for 2 104, because of which acid is spread over a larger
region at low injection rates, thus, reacting with more material
and leading to consumption of more acid. Eventually, all the
breakthrough curves for different values of 2 will reach the
low injection rate asymptote but the value of injection rate at
which they reach the asymptote will depend on the value of 2,
or the level of dispersion in the system (Remark: It is observed
in the simulations that the effect of axial dispersion on the
dissolution patterns is negligible when compared to transverse
dispersion. This was veried by suppressing axial and transverse dispersion terms alternatively and comparing it with
simulations performed by retaining both axial and transverse
dispersion in the model. Transverse dispersion is a growth
arresting mechanism in wormhole propagation because it
transfers the acid away from the wormhole and therefore
prevents fresh acid from reaching the tip of the wormhole.)
Figure 11 shows that convection and reaction are dominant
mechanisms at high injection rates leading to uniform dissolution, and at very low injection rates transverse disper-




determines whether a reaction is in the kinetic or mass transfer

controlled regime. In practice, a reaction is considered to be in
the kinetic regime if 2 r/Sh 0.1, and in the mass transfer
controlled regime if 2 r/Sh 10. For values of 2 r/Sh
between 0.1 and 10, a reaction is considered to be in the
intermediate regime. The Thiele modulus 2 is dened with
respect to the initial conditions, but the dimensionless pore
radius r and Sh change with position and time making the term
2 r/Sh a function of both position and time. At any given time,
it is difcult to ascertain whether the reaction in the entire
medium is mass transfer or kinetically controlled because these
regimes of reaction are dened for a local scale and may not
hold true for the entire system. In Table 2, the values of Thiele
modulus, the initial values of 2 r/Sh (r 1), and the ratio of
interface concentration C s to uid phase concentration C f for
different acids used in the experiments of Fredd and Fogler12
are tabulated for initial pore radii in the range of 1 m20 m.
A typical value of 3 is assumed for Sherwood number in the
calculations. The ratios of C s /C f in the table show that all the
acids except HCl are in the kinetic regime during the initial
stages of dissolution. The reaction between HCl and calcite is
in the intermediate regime. As the reaction proceeds, the pore
size increases, thereby increasing the value of 2 r/Sh, leading
to transitions between different regimes of reaction. To describe these transitions and to capture both the reaction regimes
simultaneously, we use two concentration variables in the
It has been shown earlier that o Da/Pe L O(1)
gives an approximate estimate of the optimal injection conditions for kinetically controlled reactions and the diameter of the
wormhole or the pore volume of acid required to breakthrough
was observed to depend on the macroscopic Thiele modulus
2. The extensions of these parameters to mass transfer controlled reactions are discussed here. For the case of a masstransfer controlled reaction ( 2 r/Sh 1), the species balance
Eq. 19 can be reduced to

Table 2. Ratio of Interface to Cup-Mixing Concentration for Different Acids

0.25-M EDTA
pH 13
0.25-M DTPA
pH 4.3
0.25-M EDTA
pH 4
0.25-M CDTA
pH 4.4
0.5-M HCl


2 [r o 1 m20 m]

2 /Sh

C s /C f

5.3 105




4 106

4.8 105










D m (cm2/s)

k s (cm/s)

6 106

6 10

4.5 106
3.6 105

1.4 10

2.3 104
2 101

December 2005

Vol. 51, No. 12

AIChE Journal

om P T/Pe L

a oD m2
2u o2r o


From Eq. 44, it can be shown that the minimum pore volume
depends on the parameter m . Equation 48 shows that structure property relations have a stronger inuence on the optimal
criterion for mass transfer controlled reactions when compared
to kinetically controlled reactions where

Figure 12. Breakthrough curve of a mass transfer controlled reaction (2 10).

The optimum value of om is 0.13. The values of other
parameters are N ac 0.1, o 0.2, f [0.15, 0.15],
m 3779.


c f uc f c f


x ur

os PT

os PT
y ur



A D 1/ 2M
Pe L T

This result is expected because the mass transfer coefcient is

a function of the structure of the porous medium. Figure 12
shows the pore volume of acid required to breakthrough for a
mass transfer controlled reaction (2 10) as a function of the
reciprocal of om . The breakthrough curve shows a minimum
at om 0.13. However, because of a strong dependence on
the structure property relations for mass transfer controlled
reactions, the value of om for wormhole formation is expected
to be a function of the structure-property relations. The effect
of structure-property relations on om is investigated in the
following subsection.
Figure 13 shows a comparison of breakthrough curves for
kinetic and mass transfer controlled reactions as a function of
dimensionless injection rate 2 /Da. Note that this denition of
dimensionless injection rate is independent of the reaction rate
constant. In the breakthrough curves shown in Figure 13, the
effect of reaction regime on breakthrough curves is investigated by changing the reaction rate or pore scale Thiele modulus from a very low (2 0.001) to a very large value (2


a oLD m
2u or o


is an equivalent to the Damkohler number for mass transfer

controlled reactions dened as the ratio of convection time to
diffusion time and
m2 P TPe L

a oL 2
2r o


is an equivalent to the macroscopic Thiele modulus 2. Note

that molecular diffusion or mass transfer coefcients do not
appear in the earlier denition because the Peclet number is
dened based on molecular diffusion assuming that the main
contribution to dispersion is from molecular diffusion. The
parameter that determines the optimal injection rate can be
derived from Eq. 41, and is given by

Pe L

1/ 2

M om


1/ 2


The x-coordinate is independent of reaction rate (parameters: N ac 0.1, o 0.2, f [0.15, 0.15]).

AIChE Journal

Figure 13. In the plot the reaction rate constant or 2 is

varied to simulate breakthrough curves of kinetic (2 0.001, 0.07) and mass transfer
(2 10, 100) controlled reactions.

December 2005

Vol. 51, No. 12


mination of this effect requires us to consider extremely ne

grids, which is beyond the scope of this work.
(d) Acid Capacity Number. The acid capacity number N ac
( C o / s ) depends on the inlet concentration of the acid.
Figure 14 shows the breakthrough curves for acid capacity
numbers of 0.05 and 0.1. From the breakthrough curves it can
be seen that the minimum shifts proportionally with the acid
capacity number. For low values of acid capacity number (N ac
1), the time scale over which porosity changes signicantly
is much larger than the time scale associated with changes in
concentration. In such a situation, a pseudo-steady-state approximation can be made and Eqs. 19 and 20 can be reduced to

c f
DaA cf
c f


os Da
x ur

os Da
y ur


Figure 14. Pore volume required for breakthrough is inversely proportional to the acid capacity
number (parameters: 2 0.07, o 0.2, f
[0.15, 0.15], 103).


100), thereby changing the reaction regime from kinetic to

mass-transfer control. It could be observed that the optimum
injection rate increases with increasing Thiele modulus. Thus,
acids like HCl which have a Thiele modulus larger than EDTA
should be injected at a higher rate to create wormholes. The
minimum volume required to breakthrough the core is observed to be higher for lower values of Thiele modulus. This
observation is consistent with experimental data reported in
Fredd and Fogler,12 where the minimum volume required for
EDTA is higher than the minimum volume required for HCl to
breakthrough the core. It can also be observed from Figure 13
that the injection rate is independent of reaction rate constant
for large values of 2 (see breakthrough curves of 2 10 and
100), because the system is mass transfer controlled. Figure 13
demonstrates the effect of competition between mass transport
and reaction at the pore scale on optimal conditions for injection. Because increasing temperature increases the rate constant a similar behavior as observed in Figure 13 for increasing
rate constants is expected when the temperature is increased.
Before closing this section, we consider briey the role of
heterogeneity length scale (l L h / 2r o ) on wormhole formation. In all the simulations presented in this work, we have
taken L h to be the grid size (which in physical units is 1 mm).
In practice, this length scale in carbonates can vary from the
pore size to the core size. It may be seen intuitively that when
L h l T and l X the structure of the wormholes is not inuenced by L h , as transverse dispersion dominates over the small
length scales. Similarly, when L h l T and l X , wormhole
formation is again not inuenced by L h , as it is a local phenomenon now dictated by dispersion and reaction at smaller
scales. Thus, we expect the heterogeneity length scale to play
a role in determining the wormhole structure only when L h is
of the same order of magnitude as the dispersion-reaction (l T ),
and convection-reaction (l X ) length scales. Quantitative deter-

where N ac t. Since the earlier equations are independent of

N ac , the breakthrough time BT is independent of N ac . The
breakthrough volume, dened as t/ o BT /N ac o , is, therefore, inversely proportional to acid capacity number at low
values of N ac as demonstrated in Figure 14.
(e) Effect of Structure-Property Relations. In the previous
subsections, the effect of heterogeneity, injection conditions,
reaction regime and acid concentration on wormhole formation
were investigated using the structure-property relations given
by Eqs. 79. It has been observed that the optimum injection
rate and breakthrough volume are governed by parameters o
and 2 for kinetically controlled reactions and om and m
mass transfer controlled reactions for a given set of structureproperty relations. In this subsection, the effect of structureproperty relations on the optimal conditions is investigated
using a different correlation given by


December 2005

DaA c f


exp b



The relations for average pore radius and interfacial area are
given by Eqs. 8 and 9. By changing the value of b in Eq. 52,
the increase in local permeability with porosity can be made
gradual or steep. Figures 15 and 16 show the effect of b on
evolution of permeability and interfacial area with porosity. It
can be seen from the gures that for low values of b, the
changes in permeability and interfacial area with porosity are
gradual until the value of local porosity is close to unity, and
the change is very steep for large values of b. Figure 17 shows
the effect of structure property relations on the breakthrough
curve for very low and large values of b. A mass transfer
controlled reaction is considered in these simulations because
the effect of structure property relations on optimal conditions
Vol. 51, No. 12

AIChE Journal

Figure 15. Evolution of permeability with porosity for different values of b.

The initial value of porosity o is equal to 0.36.

is signicant for mass-transfer controlled reactions as discussed

earlier. It can be seen that the optimum om , although different
for different structure property relations is approximately order
unity for large changes in the qualitative behavior of structureproperty relations. The value of minimum pore volume to
breakthrough is also observed to depend on the structureproperty relations. The lower value of minimum pore volume
for a large value of b or a steep change in evolution of
permeability is because of a rapid increase in adverse mobility
ratio between the dissolved and undissolved medium at the

Figure 16. The change in interfacial area is very gradual

for low values of b and steep for large values
of b.
AIChE Journal

December 2005

Figure 17. The effect of structure-property relations on

breakthrough volume is shown in the gure
by varying the value of b.
For low values of b the evolution of permeability and
interfacial area are gradual and the evolution is steep for
large values of b. The parameters used in the simulation are
o 0.36, f [0.03, 0.03], 2 50, m 534,
o 0.2.

reaction front. This leads to faster development and propagation of wormholes leading to shorter breakthrough times or
lower pore volumes to breakthrough.

Experimental Comparison
Wormholing in carbonates is a 3-D phenomenon. Comparison of model predictions with experimental data requires a
3-D model. Since the current model is 2-D, and due to a lack
of 2-D experimental data in carbonates, the model predictions
are compared to 2-D experiments on salt-packs reported in
Goler et al.26 In those experiments, undersaturated salt solution was injected into solid salt packed in a Hele-Shaw cell of
dimensions 25 cm in length, 5 cm in width, and 1 mm in
thickness. Because the thickness of the cell is very small
compared to the width and the length of the cell, this conguration may be considered 2-D (provided the wormhole diameter is bigger than the thickness, which is the case in these
experiments). The average values of permeability and porosity
of the salt-packs used in the experiments are reported to be
1.5 1011 m2 and 0.36, respectively. Solid salt dissolves in
the undersaturated salt solution, and creates dissolution patterns that are very similar to patterns observed in carbonates.
The dissolution of salt is assumed to be a mass transfer controlled process (Remark: This assumption breaks down as the
salt concentration in the injected solution approaches saturation
value). Figure 18 shows the experimental data on pore volumes
of salt solution required to breakthrough at different injection
rates for two different inlet concentrations (150 g/L and 230
g/L) of salt solution. The saturation concentration (C sat ) of salt
is 360 g/L and the density of salt ( salt ) is 2.16 g/cm3. The
dissolution of salt in an undersaturated salt solution is a process
very similar to dissolution of carbonate due to reaction with
Vol. 51, No. 12


Figure 18. Experimental data on salt dissolution reported in Goler et al.26

Under-saturated salt solution was injected into a salt-pack to
create dissolution patterns. The breakthrough curves in the
plot correspond to two different inlet concentrations, 150
g/L and 230 g/L, of salt solution.

acid, and the model developed here can be used for salt
dissolution by dening the acid concentration to be C f
C sat C salt . Thus, the acid capacity number for a salt solution
of concentration C o g/L is given by
N ac

C sat C o

Using the above equation, the acid capacity numbers for salt
concentrations of 230 g/L and 150 g/L are calculated to be 0.06
and 0.097 respectively.
To compare model predictions with experimental data, we
need information on initial average pore radius, interfacial area,
and structure-property relations. This data is difcult to obtain
directly and, hence, the model is calibrated with experimental
data to obtain these parameters. Using these parameters, the
model is simulated for a different set of experimental data for
comparison. From the earlier results, we know that for masstransfer controlled reactions, the pore volumes of salt solution
required to breakthrough is a function of the parameters om ,
, and structure property relations for a given inlet concentration. The model is rst calibrated to the breakthrough curve
corresponding to the inlet salt solution concentration of 150
g/L. For calibration, the largest uncertainty arises from lack of
information on structure-property relations, so the relation in
Eq. 52 is used with the value of b 1. The minimum pore
volume to breakthrough depends on m
, and its value is used
to calibrate to the experimental minimum after the structure
property relations are xed. Then, the pore volume to breakthrough curve is generated for different values of om . Figure
19 shows the calibration curve of the model with the experimental data. The value of m used for calibration is 534. This
value of m is used to simulate the model for inlet salt
concentration of 230 g/L (N ac 0.06). The comparison of

December 2005

Figure 19. Plot shows the calibration of the model with

experimental data for different structureproperty relations.

model predictions with experimental data is shown in Figure

20. The value of a o /r o can be calculated using Eq. 47, and is
found to be 912.49 cm2. Using this value of a o /r o , and the
optimum value om 0.33, the value of injection velocity is
calculated from Eq. 49 to be 1.29 103 cm/s (D m 2
10 5 cm2/s). This value is much lower when compared to the
experimental optimum injection velocity u o 0.045 cm/s. To
get a better estimate of the injection velocity a different value
of b 0.01 is used for the structure-property relations, and the
model is calibrated with the data for inlet salt concentration of
150 g/L (see Figure 19). The value of m used for calibration
is 1,195. The model predictions for this value of b for inlet salt
solution concentration of 230 g/L is shown in Figure 20. The
injection velocity is calculated using the procedure described
before and is found to be 3 103 cm/s. (One possible reason
for this calculated value to be one order of magnitude smaller
than the experimental value is our assumption of a constant

Figure 20. Model predictions with experimental data for

different structure-property relations.
Vol. 51, No. 12

AIChE Journal

os . In general, dissolution increases the connectivity between

the pores and, hence, the effective transverse dispersion coefcient. Thus, if we include the variation of the transverse
dispersion coefcient also in the structure-property relations, it
follows from Eq. 48 that the calculated u o will be higher.)
The earlier comparisons show that the model predictions in
terms of pore volume are in reasonable agreement with experimental data. However, the predictions of injection velocity are
not in good agreement with experimental values due to lack of
information on structure-property relations. Further investigation on obtaining structure-property relations and calibration
methods are required for experimental comparison.

Conclusions and Discussion

The main contribution of this work is the development of a
two-scale continuum model that retains the qualitative features
of reactive dissolution of porous media. We have also analyzed
in detail the 2-D version of the model and determined the
inuence of various parameters such as the level of dispersion,
the magnitude of heterogeneities, strength of acid, and pore
scale mass transfer on wormhole formation. It is observed that
wormhole formation depends on the magnitude of convection,
dispersion and reaction effects in the system. A qualitative
criterion is developed to predict conditions required for creating wormholes and it is given by O(1), where
k eff D eT /u o . The branched wormholes observed in experiments are found to be a result of the heterogeneities in the core.
The optimum injection rate and minimum volume required to
breakthrough were found to be weak functions of initial heterogeneity in the core. Pore-scale mass transfer and reaction
were observed to have a strong inuence on the optimal conditions. The model predictions are in qualitative agreement
with laboratory data on carbonate cores and salt-packs presented in the literature.
The model presented here as well as the numerical calculations can be extended in several ways. Our preliminary calculations indicated that the fractal dimension of the wormhole
formed depends both on the magnitude of heterogeneity and
the rate constant (2). Strong acids such as HCl and higher
levels of heterogeneities produce thinner wormholes but have a
higher fractal dimension. In contrast, weak acids and lower
levels of heterogeneities lead to fatter wormholes having lower
fractal dimension. It is not clear intuitively which of these is
preferred because in the former case, the interfacial area between the wormhole, and the reservoir rock is higher while in
the latter case, the permeability of the wormhole is larger. The
model presented here can be used to quantify the effect of
wormholes of different fractal dimension and size on the overall permeability.
The calculations presented in this work were limited to the
case of linear kinetics and constant physical properties of the
uid. The model can be extended to include multistep chemistry at the pore scale as well as changing physical properties
(for example, viscosity varying with local pH) on wormhole
structure. Likewise, all the calculations presented here are for
a xed aspect ratio. The model can be used to determine the
density of wormholes by changing the aspect ratio corresponding to that near a wellbore (for example, height of domain
much larger than width). Preliminary calculations of this type
were presented by Panga et al.22 These calculations, as well as
AIChE Journal

December 2005

additional simulations using a 3-D version of the model will be

pursued in future work.

This work was partially supported by an unrestricted grant by Schlumberger to the University of Houston. We thank Drs. Eduard Siebrits and
Philippe Tardy for many helpful comments.

Literature Cited
1. William B, Gidley J, Schechter RS. Acidizing Fundamentals. SPE
Monograph Series; 1992.
2. Schechter RS. Oil Well Stimulation. New Jersey: Prentice Hall Inc.;
3. Economides M, Hill A, Ehlig-Economides C. Petroleum Production
Systems. Prentice Hall Inc.; 1993.
4. Economides M., Nolte K (Eds.). Reservoir Stimulation. Wiley; 2000.
5. Daccord G, Lenormand R. Fractal patterns from chemical dissolution.
Nature. 1987; 325: 41 43.
6. Daccord G. Chemical dissolution of a porous medium by a reactive
uid. Phys Rev Lett. 1987; 58: 479 482.
7. Daccord G, Lenormand R, Lietard O. Chemical dissolution of a porous
medium by a reactive uid-1: model for the wormholing. Chem Eng
Sci. 1993a; 48: 169 178.
8. Daccord G, Lenormand R, Lietard O. Chemical dissolution of a porous
medium by a reactive uid-2: convection vs reaction, behavior diagram.
Chem Eng Sci. 1993b; 48: 179 188.
9. Hoefner ML, Fogler HS. Pore evolution and channel formation during
ow and reaction in porous media. AIChE J. 1988; 34: 4554.
10. Wang Y, Hill AD, Schechter RS. The optimum injection rate for
matrix acidizing of carbonate formations. SPE 26578. SPE Ann. Tech.
Conf. and Exhib., Houston, TX; 1993; 675 687.
11. Frick TP, Mostozadeh B, Economides MJ. Analysis of radial core
experiments of hydrochloric acid interaction with limestones. SPE
27402. SPE Intl. Symposium on Formation Damage Control, Lafayette,
LA; 1994;577592.
12. Fredd CN, Fogler HS. Inuence of transport and reaction on wormhole
formation in carbonate porous media. AIChE J. 1998; 44: 19331949.
13. Bazin B. From matrix acidizing to acid fracturing: A laboratory
evaluation of acid/rock interactions. SPE Prod. & Facilities. Feb. 2001;
14. Buijse MA. Understanding wormholing mechanisms can improve acid
treatments in carbonate formations. SPE Prod. & Facilities. Aug. 2000;
15. Fredd CN, Fogler HS. Optimum conditions for wormhole formation in
carbonate porous media: Inuence of transport and reaction. SPE J.
1999; 4: 196.
16. Fredd CN. Dynamic model of wormhole formation demonstrates conditions for effective skin reduction during carbonate matrix acidizing.
SPE 59537. SPE Permian Basin Oil and Gas Recovery Conference.
Midland, Texas; 2000: 1.
17. Daccord G, Touboul E, Lenormand R. Carbonate Acidizing: Toward a
quantitative model of the wormholing phenomenon. SPE Prod Eng.
1989; 63.
18. Hung KM, Hill AD, Sepehrnoori K. A mechanistic model of wormhole growth in carbonate matrix acidizing and acid fracturing. J Pet
Tech. 1989; 41: 59.
19. Huang T, Hill AD, Schechter RS. Reaction rate and uid loss: The
keys to wormhole initiation and propagation in carbonate acidizing. SPE
37312. SPE Intl. Symposium on Oileld Chemistry, Houston, Texas
1997; 775.
20. Huang T, Zhu D, Hill AD. Prediction of wormhole population density
in carbonate matrix acidizing. SPE 54723. SPE European Formation
Damage Conference; 1999;1 The Hague, Netherlands.
21. Gdanski R. A fundamentally new model of acid wormholing in carbonates. SPE 54719. SPE European Formation Damage Conference;
1999: 1 The Hague, Netherlands.
22. Panga MKR, Ziauddin M, Gandikota R, Balakotaiah V. A new model
for predicting wormhole structure and formation in acid stimulation of
carbonates. SPE 86517. SPE International Symposium and Exhibition
on Formation Damage Control; 2004;1 Lafayette, Louisiana.
23. Liu, X, Ormond A, Bartko K, Li Y, Ortoleva P. A geochemical

Vol. 51, No. 12


reaction-transport simulator for matrix acidizing analysis and design. J

Pet Sci Eng. 1996; 181.
24. Chen Y, Mu J, Chen W, Davis D, Ortoleva P. Optimizing matrix
acidizing for near-borehole remediation using the CIRF reaction-transport simulator. SPE 26184. SPE Gas Tech. Symposium. Calgary, Alberta; 1993; 471.
25. Chen Y, Fambrough J, Bartko K, Li Y, Montgomery C, Ortoleva P.
Reaction-Transport simulation of matrix acidizing and optimal acidizing
strategies. SPE 37282. SPE International Symposium on Oileld Chemistry. Houston, Texas 1997; 679.
26. Goler F, Zarcone C, Bazin B, Lenormand R, Lasseux D, Quintard M.
On the ability of a Darcy-scale model to capture wormhole formation
during the dissolution of a porous medium. J Fluid Mech. 2002; 457:
27. Froment GF, Bischoff KB. Chemical Reactor Analysis and Design.
John Wiley & Sons, 1990.
28. Chadam J, Hoff D, Merino E, Ortoleva P, Sen A. Reactive inltration
instabilities. IMA J Appl Math. 1986; 36: 207.
29. Sherwood JD. Stability of a plane reaction front in a porous medium.
Chem Eng Sci. 1987; 42: 1823.
30. Hinch EJ, Bhatt BS. Stability of an acid front moving through porous
rock. J Fluid Mech. 1990; 212: 279.
31. Crompton S, Grindrod P. A geometric approach to ngering instabilities for reaction fronts in fully coupled geochemical systems. IMA
J Appl. Math. 1996; 57: 29.
32. Homsy GM. Viscous ngering in porous media. Ann Rev Fluid Mech.
1987; 19: 271.
33. Gupta N, Balakotaiah V. Heat and mass transfer coefcients in catalytic monoliths. Chem Eng Sci. 2001; 56: 4771.
34. Balakotaiah V, West DH. Shape normalization and analysis of the
mass transfer controlled regime in catalytic monoliths. Chem Eng Sci.
2002; 57: 1269.


December 2005

Appendix A
Numerical algorithm
The equations are discretized on a 2-D domain using a
control volume approach. While discretizing the species balance equation, an upwind scheme is used for the convective terms
in the equation. The following algorithm is used to simulate ow
and reaction in the medium. The pressure, concentration and
porosity proles in the domain at time t are denoted by pt, ct and
t. Porosity and concentration proles in the domain are obtained
for time t t (ctt, tt), by integrating the species balance
and porosity evolution equations simultaneously using the ow
eld calculated from the pressure prole ( pt) by applying Darcys
law. Integration of concentration and porosity proles is performed using Gears method for initial value problems. The calculation for concentration and porosity proles is then repeated for
time thalf t t/2 using the velocity prole at time t. The ow
eld at t t/2 is then calculated using the concentration prole
chalf and porosity prole half. Using the ow prole at thalf the
values of concentration and porosity are again calculated for time
t t, and are denoted by cnew and new. To ensure convergence,
the norms ctt cnew and tt new are maintained below
a set tolerance. If the tolerance criterion is not satised the calculations are repeated for a smaller time step. The earlier procedure
is repeated until the breakthrough of the acid which is dened as
the decrease in the initial pressure by a factor 100.
Manuscript received Feb. 16, 2005, and revision received Apr. 25, 2005.

Vol. 51, No. 12

AIChE Journal


Measurement of Moisture Diffusivity through

Layered-Silicate Nanocomposites
Hiteshkumar T. Rana and Rakesh K. Gupta
Dept. of Chemical Engineering and Constructed Facilities Center, West Virginia University, Morgantown, WV 26506

Hota V. S. GangaRao
Dept. of Civil and Environmental Engineering and Constructed Facilities Center, West Virginia University,
Morgantown, WV 26506

L. N. Sridhar
Dept. of Chemical Engineering, University of Puerto Rico, Mayaguez, PR 00681
DOI 10.1002/aic.10605
Published online August 22, 2005 in Wiley InterScience (www.interscience.wiley.com).

Layered-silicate nanoparticles, when distributed within the matrix of a ber-reinforced

polymer (FRP), can retard the diffusion of environmental moisture and other chemicals
to the bermatrix interface where their presence can result in delamination and ber
weakening. Thus, the use of nanoparticles helps to preserve the integrity of FRPs and to
prolong the service life of composites when these are used in outdoor applications such
as bridges and utility poles. In the present work, diffusion coefcients of water, brine, and
alkali were determined at several different temperatures by soaking clay/vinyl ester
nanocomposites in the different liquids and measuring the increase in weight with
increasing time of immersion. It was found that the measured diffusivities decreased with
increasing clay content. However, the reduction in water diffusion coefcient was found
to be less severe when steady permeation experiments were carried out. This result is
theoretically explained based on the hydrophilic nature of the clay surface that tends to
immobilize some of the moisture. 2005 American Institute of Chemical Engineers AIChE J,
51: 3249 3256, 2005

Keywords: diffusivity measurement, moisture permeation, vinyl ester, polymer nanocomposites, montmorillonite, barrier properties

Glass-berreinforced polymer (GFRP) composites incorporating thermosetting polymer matrices are now being used in
the construction and repair of bridges and other civil structures.1 However, widespread use of GFRPs in construction,
where a service life in excess of 50 years is desired, has been
hindered by the lack of long-term durability and performance
Correspondence concerning this article should be addressed to R. K. Gupta at

2005 American Institute of Chemical Engineers

AIChE Journal

December 2005

data on which to base design calculations. A particular difculty has been that atmospheric moisture can diffuse to the
bermatrix interface and cause both delamination and ber
weakening.2 One way to reduce moisture ingress is to reduce
the effective moisture diffusion coefcient, and this can be
achieved by dispersing clay platelets in the matrix polymer,
forming a nanocomposite; the diffusing molecules have to
travel around the akelike barriers, and the additional path
length, by itself, is predicted to substantially decrease the
diffusivity (see, for example, Cussler et al.3 and Yano et al.4).
Note that the most common variety of clay for nanocomposite
synthesis is montmorillonite.5,6 It is a swelling clay consisting
Vol. 51, No. 12


of layers formed by a condensation of two silicon oxygen

tetrahedral sheets and one aluminum or magnesium oxygen
hydroxyl octahedral sheet; the layer thickness is 0.96 nm and
the lateral dimensions are of the order of a few hundred
nanometers each. These layers generally exist stacked together,
and they have to be separated or exfoliated to be effective
mass-transfer barriers. However, the clay is inherently hydrophilic, and a surface treatment is needed to make it organophilic for ease of dispersion in a polymer even when mechanical shear is used.
In our previous work,5 we studied moisture diffusion
through vinyl ester samples containing up to 5 wt % of organically treated montmorillonite; sorption experiments were done
on rectangular cross-sectional strips within which the clay
platelets were partially exfoliated. It was found that water
diffusivity decreased to half its value in the neat resin when the
clay content was only 1 wt %. The present research is a
continuation of the prior work, and here we have examined
moisture diffusion through the nanocomposite using both sorption and permeation methods; there is a difference in the
results, and this is partially explained on a theoretical basis. We
also present data on the diffusion of brine and alkali through
both the neat polymer and through the nanocomposite, and
these data provide a clue to other reasons why the diffusivity
obtained from sorption experiments differs from the corresponding value obtained from permeation experiments.
We note that the addition of passive mass-transfer barriers in
a polymer can only retard the rate of diffusion and not eliminate diffusion altogether. However, the timescale of diffusion
through a structure such as a bridge deck is so long that, unless
atmospheric humidity is always high, the natural reversal in
concentration gradient, resulting from a change in seasons or
the cessation in waterlogging, should prevent moisture from
reaching very far into the composite material. When the nanoller is clay, as in the present case, it is hydrophilic in nature,
and this has an additional benecial effect. Any moisture that
does diffuse into the GFRP tends to become sequestered on the
clay surface and does not reach the surface of the reinforcing
bers. Indeed, we have found7 that the presence of montmorillonite in the GFRP increased its durability, and this is demonstrated by the scanning electron micrographs shown in Figure 1. Figure 1a shows the fracture surface of a freshly
prepared GFRP sample without clay; the bers are circular in
cross section, and there is good bonding between the bers and
the matrix. Figure 1b shows that after 2 months of being soaked
in distilled water, the GFRP is severely damaged to the extent
that the ber cross section has become noncircular. Note that,
as expected, the bers that were closer to the sample surface
were damaged to a greater degree than those that were situated
away from the surface. Finally, Figure 1c shows that, with the
addition of 5 wt % clay, the damage that is visible in Figure 1b
can be avoided. Thus, adding clay appears to help in making
polymer nanocomposites more durable, although these simple
experiments cannot be expected to help quantify the increase in
service life.

Materials Used and Sample Preparation

The polymer used in this work was DerakaneTM 411-350
Momentum epoxy vinyl ester resin obtained from the Dow
Chemical Co. (Midland, MI), and it contained 45 wt % dis3250

December 2005

solved styrene; this is a slightly different chemical form of the

polymer that we used in previous investigations.5 It was cured
at room temperature, as recommended by the manufacturer.
One wt % of methyl ethyl ketone peroxide (9% active oxygen)
was used as the initiator, whereas 0.03 wt % of 6% cobalt
naphthenate was used as a catalyst. Both of these chemicals
were purchased from Sigma Aldrich Co. (St. Louis, MO). To
formulate the nanocomposites, Cloisite 10A, a montmorillonite treated with benzyl (hydrogenated tallow alkyl) dimethyl
quaternary ammonium chloride, was used as received from
Southern Clay Products (Gonzales, TX); the organic content of
this treated clay is approximately 39 wt %. Sodium hydroxide
used in this study was obtained from Aldrich Chemical (Milwaukee, WI), whereas sodium chloride was obtained from
Fisher Scientic (Pittsburgh, PA).
Transparent samples of neat vinyl ester were prepared by
pouring the mixture of resin, initiator, and catalyst into a
Teon mold. Typical dimensions of the samples used for
transient diffusion experiments were 50 12.5 mm, with the
thickness ranging from 0.2 to 0.7 mm. The polymer was
allowed to cure at room temperature for 24 h, and it was
subsequently postcured in an oven for 3 h at 95C; differential
scanning calorimetry showed that these conditions led to complete curing and resulted in a glass-transition temperature (Tg)
of 117C. In the case of vinyl ester resin clay nanocomposites,
the organically treated clay was added to the liquid resin and
manually stirred. The mixture was then degassed in a vacuum
oven to remove air bubbles. The catalyst and initiator were then
added, and the resin was poured into the molds and subjected
to the same cure cycle as the neat resin. Similar procedures
were used to make samples for the steady permeation experiments, although now all the samples were square in shape
(60 60 mm), but the region for mass transfer was circular,
with a diameter of 50.7 mm. Additional details are available in
Ranas thesis.7

Material Characterization
As reported previously,5 X-ray diffraction scans on the nanocomposite samples were indicative of either exfoliation or
breakdown of the ordered structure of clay. Low-magnication
transmission electron micrographs of the same samples showed
that the clay was reasonably uniformly distributed, and the clay
platelets were randomly oriented. High-magnication micrographs, however, revealed the presence of aggregates made up
of two to ten platelets; the distance between adjacent platelets
in these aggregates ranged from 4.4 to 5 nm, whereas the
platelet length ranged from 80 to 180 nm. Thus, the clay is only
partially exfoliated, and this has consequences for the resulting
barrier properties.

Diffusion Tests and Data Analysis

To determine the improvement in barrier properties of a
nanocomposite compared to the neat polymer, one may
carry out water absorption tests. Here one immerses a sample in water and measures the amount of water absorbed in
a xed amount of time; barrier properties are considered to
have been improved if the amount of moisture absorbed
decreases with added nanoller.8,9 However, more quantitative information can be obtained by computing the diffusivVol. 51, No. 12

AIChE Journal

Figure 1. (a) SEM of as-prepared GFRP sample; (b) SEM of GFRP sample after immersion in distilled water at room
temperature for 2 months; (c) SEM of GFRP sample containing 5 wt % Cloisite 10A after immersion in
distilled water at room temperature for 2 months.

ity from either sorption experiments or permeation experiments. In the present work, transient diffusion tests were
conducted by immersing the various samples in distilled
water or alkali of pH 13 or 0.1 M salt solution at a constant
temperature ranging from 4 to 62C. Beakers containing
these solutions were closed at the top with aluminum foil to
prevent evaporation. A refrigerator was used to conduct
experiments at temperatures below room temperature,
whereas an oven was used at temperatures above room
temperature. The samples were periodically removed, blotAIChE Journal

December 2005

ted dry with lint-free tissue, weighed, and reimmersed in the

beaker. Readings were initially taken as frequently as every
30 min, and the experiment was continued until there was no
weight gain recorded over a period of 10 days. The balance
used had an accuracy of 1 g, and at least three replicate
runs were carried out for a given set of conditions. To
compute the diffusion coefcient from data on mass gain as
a function of time, one may use Fickian theory. For unlled
samples, the process of one-dimensional, unsteady diffusion
is governed by10,11
Vol. 51, No. 12


D 2


in which c is the concentration of the diffusing species, t is

time, x is the position in the diffusing direction, and D is the
diffusion coefcient or diffusivity. Upon solving Eq. 1 with
constant boundary conditions, we obtain moisture uptake (that
is, increase in sample mass) Mt to be


D(2n 1)2 2 t
(2n 1) 2 2


where M is the equilibrium increase in sample mass and 2l is

the sample thickness. At the initial stages of diffusion, the
solution for Ficks law at small times reduces to

(2l) 2

1/ 2


so that the diffusion coefcient can be computed from the

initial slope of Mt/M vs. t1/2/2l using short-time water-uptake
Another method of determining the moisture diffusion coefcient is by steady permeation experiments. Red sealing wax
was used to attach a polymer lm to the mouth of an aluminum
dish containing 4 g of anhydrous CaCl2 desiccant; the diameter
of the polymer lm that was exposed to moisture was 50.7 mm.
The sealed assembly was weighed and put inside a chamber
that was maintained at a constant relative humidity (RH) of
77% at a temperature of 25C. As moisture diffused into the
dish, the mass of the assembly progressively increased, and the
rate of increase of mass ultimately became constant. This was
determined by periodic weighing of the assembly using an
electronic balance. To attain a steady state in a short amount of
time, the polymer lms were initially saturated with water
vapor at the relative humidity level of the chamber. Knowing
the steady mass ux, the diffusion coefcient can be calculated
as the ratio of the ux to the concentration gradient. Given that
the absolute humidity inside the assembly can be taken to be
negligible, the concentration gradient is simply the ratio of the
moisture content in the polymer in equilibrium with 77% RH at
25C to the thickness of the lm.
When diffusion takes place through a polymer lm containing dispersed platelets, the diffusing molecules have to go
around these barriers. The morphology that is likely to be most
effective in reducing diffusivity is one where the akelike
barriers are positioned parallel to the lm surface and perpendicular to the direction of mass transport. This idealized geometry is shown in Figure 2; although this is only rarely
achieved (but see Figure 5 of Yano et al.4), it is useful from the
perspective of theoretically calculating the maximum improvement in barrier properties.4 As seen from Figure 2, the path
length for steady diffusion increases, say by a factor of , as
compared to the unlled case. In addition, the area for mass
transfer decreases by a factor (1 ). Thus, in a steady
permeation experiment, the measured diffusion coefcient Ds
should equal (1 )D/, if D is the diffusion coefcient for
mass transfer through the neat polymer. Note that both and

December 2005

Figure 2. Morphology of a model layered-silicate nanocomposite.

should increase with increasing concentration. Note also that

the measured diffusivity should not be inuenced by any moisture that becomes bound to the (hydrophilic) surface of the
ller particles. In other words, if an increase in the diffusion
path is the main mechanism by which the clay platelets reduce
diffusivity, then the percentage reduction in the diffusion coefcient of moisture on adding a certain amount of clay to
polymer should be identical to the percentage reduction in the
diffusion coefcient of any other molecule, say oxygen, under
similar conditions. Furthermore, this percentage reduction
should not depend on the temperature at which the measurements are made. This expectation is not fullled in (transient)
sorption experiments (see, for example, published data in Tortora et al.12).
Upon considering transient diffusion, one nds that the mass
balance is altered by the presence of clay, and one has



where is the volume fraction of clay in the polymer. If we

now let the ux nx equal (D/)(c/x), Eq. 4 becomes
c D1 2c

1 x 2


and the effective transient diffusivity Dt is given by

Vol. 51, No. 12

AIChE Journal

Figure 3. Water sorption curve of postcured neat resin at 25C.



D s/D t 1


Finally, then

Because 1 it is obvious from Eq. 7 that Dt would be

greater than Ds and both Dt and Ds would be less than D,
clearly demonstrating that (1) both the effective steady-state
and transient diffusivities would decrease as a result of the
introduction of the nanoparticles and (2) the effective transient
diffusivity would reduce less than the steady-state diffusivity.
For these conclusions to be quantitatively valid, however, the
nanocomposite morphology must be the one described in Figure 2, and the ller particles must not interact with the diffusing

Results and Discussion

Water sorption through neat resin
Figure 3 shows the fractional increase in mass of vinyl ester
samples plotted vs. t1/2/2l upon immersing neat resin coupons
in distilled water at 25C. As expected, based on Eq. 3, water
uptake increases linearly at short times, but it levels off at
longer times as equilibrium is attained; the equilibrium moisture content is about 0.66 wt %, and it takes a little over 11 days
for equilibrium to be reached. Figure 3 demonstrates that data
from different runs superpose, but moisture diffusion does not
faithfully follow the predictions of Eq. 2 (solid line). In other
words, the process of moisture diffusion through the neat resin
is slightly non-Fickian because there is a mismatch between
theory and data at intermediate times. This nding, which
differs from our previous results,5 could be explained by the
minor change that was made in the chemical composition of the
resin by the manufacturer. Although better agreement can be
obtained between data and theoretical predictions if the theory
is modied to, say, make the diffusivity depend on moisture
content, we have chosen not to do this here because the
conclusions we seek to draw are not inuenced by whether the
Fickian theory holds exactly.
The moisture diffusion coefcient computed from the initial
AIChE Journal

December 2005

slope of the solid line in Figure 3 is 0.74 106 mm2/s, and

this is somewhat lower than the value measured by Shah et al.5
because of the slightly higher value of the equilibrium moisture
content in the present case. When experiments were done at
4C, results were qualitatively similar to those displayed in
Figure 3. In quantitative terms, although the diffusion coefcient was signicantly lower at 0.14 106 mm2/s, the equilibrium moisture content was only slightly lower at 0.6 wt %.
Consequently, it took twice as long to reach equilibrium at 4C
compared to the time taken at 25C. At a temperature of
42.5C, sorption experiments did not seem to reach equilibrium
even after 20 days, and instead there was a reduction in sample
mass beyond this time. It seemed that the vinyl ester was
hydrolyzing, and this speculation was conrmed when the
temperature of the experiment was raised to 62C. Now the
samples degraded visibly after only 4 days of immersion, and
the experiments were discontinued. Clearly, in the case of this
particular resin, an increase in temperature cannot be used to
shorten the timescale of the diffusion experiments. We hasten
to add, though, that the vinyl ester used for bridge applications
is different from the one used here, and it cures at a high
temperature. The resin used here was chosen for convenience
because it can be cured at room temperature.

Inuence of clay on moisture sorption

The water-sorption behavior of nanocomposite samples containing different amounts of montmorillonite was qualitatively
similar to that of neat resin samples at each temperature examined. All the samples showed a linear initial uptake, and the
calculated diffusion coefcients are listed as a function of clay
content in Table 1 at both 4 and 25C; at a given level of added
clay, the percentage change in diffusion coefcient is found to
depend on temperature. Although there is some scatter in the
results, it is seen that, at each temperature, the trend is for the
diffusion coefcient to decrease as the clay content in the resin
increases. At room temperature, the addition of only 1 wt %
clay appears to reduce the diffusivity by 46%, whereas at the 5
wt % clay level, the reduction in diffusivity is almost 91%. The
incorporation of an additional amount of clay in the vinyl ester
may reduce the moisture diffusion coefcient further, but this
may also lead to deterioration in other properties. Thus, there is
likely to be an optimum in the amount of added clay; this was
Vol. 51, No. 12


Table 1. Measured Values of Moisture Diffusivity and

Equilibrium Moisture Content of Vinyl Ester Resin Samples
106 (mm2/s)

Content (wt %)






Neat resin
1 wt % Cloisite 10A
2 wt % Cloisite 10A
5 wt % Cloisite 10A





not determined as part of this study. It should be noted that the

extent of diffusivity reduction achieved using these transient
experiments seems to be quite large, but similar results have
been obtained by us and by others, as well.5,12
As opposed to the decrease in diffusivity with clay addition,
the equilibrium moisture content of the nanocomposite sample
was found to increase, albeit in a nonlinear fashion, with an
increase in the amount of clay; these results are also listed in
Table 1. Note that the standard deviation for all the diffusivity
data is of the order of 0.02 106 mm2/s, whereas for the
equilibrium moisture content it is 0.01 wt %. As noted in the
past,5 the reason for the higher equilibrium water content with
increasing amounts of clay is the tendency of the treated clay
to attract moisture. Even though the surface treatment that is
given to the clay tends to make it organophilic, it still retains its
hydrophilic character; when the treated clay was exposed to
75% relative humidity at 25C, for a period of 720 h, it
increased in weight by almost 19% as opposed to a 60%
increase in weight for nonsurface-treated clay.5 For the nanocomposite samples, it was also found that, at both 4 and 25C,
the time to reach equilibrium in these experiments was approximately 2 months. Thus, despite the generally higher diffusivity
values at the higher temperature, the timescales for diffusion
are essentially similar because of the correspondingly larger
equilibrium moisture contents at the higher temperature (see
Table 1).
Results of sorption experiments conducted at 42.5 and 62C
using nanocomposite samples were essentially similar to the
results on neat resin samples. At 42.5C, a slow increase in
sample mass was followed by a decrease in sample mass after
about 25 days of immersion, and this did not seem to be
inuenced by the percentage of clay in the sample. A visual
inspection of the samples revealed the presence of tiny craters
on the surface. At 62C, all the samples began to dissolve/
degrade after 4 days of water immersion. However, the samples
containing clay appeared to maintain their physical integrity
much better, and this effect was very pronounced at the highest
clay loading, that is, 5 wt % in our case. Even so, some kind of
chemical or physical modication of this polymer is necessary
to prevent polymer degradation if contact with hot water is
likely to occur.
In summary, then, adding clay to vinyl ester progressively
reduces moisture diffusivity as measured in transient sorption
experiments. However, the surface-treated clay retains its hydrophilic character and, as a result, the equilibrium moisture
content of the nanocomposites increases with amount of added
clay. Both these facts mean that the time to reach concentration
equilibrium can stretch into several months and, if unimpeded,

December 2005

water diffusion can go on for a very long time, even in thin


Sorption of brine and caustic soda into nanocomposites

Experiments, similar to those conducted using distilled water, were also carried out using 0.1 M NaCl solution and a
NaOH solution of pH 13. The diffusion coefcients were
calculated as before, and results for diffusivity and percentage
equilibrium weight gain for the salt solution are displayed in
Table 2 at temperatures of 4 and 28C. When one compares the
results presented in Table 2 with those given in Table 1, one
notices that (1) clay is much less effective in reducing the
diffusivity of brine than distilled water, and (2) brine adsorbs to
a much lesser extent on the clay particles. Indeed, at 4C, the
equilibrium brine content of the nanocomposite samples is, for
all practical purposes, independent of the amount of added
clay. In other words, there is no interaction between the ller
particles and the diffusing molecules at 4C.
If we accept the hypothesis that the mechanism by which
nanoparticles reduce the diffusion coefcient is the additional
path length encountered by the diffusing species, then, given
that the nanocomposite structure is not changed by the choice
of the diffusing molecule, the same percentage change in
diffusivity ought to be observed regardless of the chemical
nature and size of the diffusing molecule or the temperature of
measurement. These requirements are obviously not met by the
data in Tables 1 and 2. We also note that others too have
observed that, for the same amount of clay in a sample, the
sorption method yields a larger reduction in the diffusion
coefcient when the diffusing molecule adsorbs on the clay
than when it does not.12 An explanation is clearly in order.
From the various structural models available in the literature,3,4 we know that the predicted diffusion coefcient should
depend sensitively on the physical structure, or equivalently the
ller agglomeration level, of the nanocomposite; in particular,
there should be a substantial difference in the diffusion coefcient when the clay is exfoliated compared to when it is
intercalated. Unfortunately, this difference cannot be picked up
in sorption experiments when the diffusing molecule, such as
distilled water, adsorbs signicantly onto the clay surface, and
the amount of adsorbed moisture is comparable to or more than
the amount of moisture that can be dissolved in the polymer.
When moisture enters the nanocomposite, it is attracted to the
clay surface, and water molecules preferentially go to the
surface of the clay platelets that they encounter rst before
diffusing further into the polymer. It is only when the surface
of these ller particles is completely saturated with moisture
that additional water molecules seek out the next set of clay
Table 2. Measured Values of Brine (0.1 M NaCl) Diffusivity
and Equilibrium Weight Gain Values of Vinyl Ester Resin
106 (mm2/s)

Content (wt %)






Neat resin
1 wt % Cloisite 10A
2 wt % Cloisite 10A
5 wt % Cloisite 10A





Vol. 51, No. 12

AIChE Journal

clay. As mentioned earlier, the diffusion coefcient Ds can be

calculated from the measured steady-state mass ux as

Figure 4. Water permeation curve of postcured neat

resin at 25C.

platelets and so on. In this situation, it is clear that as long as

water molecules are not excluded from approaching the surface
of each clay platelet present in the nanocomposite, the rate of
water uptake will not depend on the level of exfoliation (or
dispersion) provided that the clay is uniformly distributed in
the polymer matrix. As a consequence, when the diffusing
molecule is strongly attracted to the clay surface, the results of
a sorption experiment will indicate the maximum possible
reduction in diffusivity, and this corresponds to complete ller
exfoliation. However, the observed reduction in diffusion coefcient will become progressively less as the level of interaction between the diffusing species and the ller particles is
reduced to zero. This is what is observed in Tables 1 and 2. We
note that untreated clay generally contains sodium or calcium
ions between the different clay layers, and these ions are
exchanged with organic cations during surface treatment. We
speculate that not all the sodium ions are replaced with organic
ions in this process, and the interaction between the sodium
ions in the clay and the sodium ions in the brine is responsible
for the reduced moisture absorption when nanocomposite samples are immersed in the brine. Thus, the results for brine
diffusion at 4C are more representative of the nanocomposite
structure than water diffusion results at room temperature, and
these indicate that the clay contained in the nanocomposite
samples used in the present research was not fully exfoliated.
Brine diffusion experiments were carried out at higher temperatures as well, but, as in the case of distilled water, the
samples began to degrade in about 6 to 8 days, and the
experiments were terminated. The same phenomenon was observed with the NaOH solution but at room temperature; at a
temperature of 42.5C, it took only 24 h for samples to degrade. Alkali diffusion data could be obtained at 4C and these
were almost identical to the NaCl data at the same temperature.7

Water permeation through nanocomposites

Periodic weight-gain data for the aluminum dish assembly
used for permeation experiments are shown in Figure 4 for neat
resin lms. It is seen that, in each of the four cases, a steady
state is attained after nearly 200 h. When corresponding runs
were carried out with nanocomposite lms, results were similar, but the time to achieve steady state increased with increasing clay amount, reaching 600 h for samples containing 5 wt %
AIChE Journal

December 2005

Mass transfer rate Sample thickness

Sample area


where, in the calculation of the concentration gradient, the

moisture concentration in the polymer lm in equilibrium with
77% relative humidity at 25C is used in place of the concentration difference because one side of the lm (the inner surface) is maintained at zero moisture content; this equilibrium
concentration was found to be 11.96 mg/cm3 for neat vinyl
The calculated diffusivities are listed in Table 3, and
whereas the diffusion coefcient value for diffusion of water
through vinyl ester is comparable to that measured using sorption experiments, such is not the case for the nanocomposites.
Although, as expected, the numbers in Table 3 decrease as the
montmorillonite content in the vinyl ester is increased, showing
the benecial effect of clay addition, the percentage decrease is
much smaller than what would be anticipated based on Eq. 7
and the diffusion coefcient results seen earlier in Table 1.
Indeed, Table 3 shows that the percentage reduction in diffusion coefcient on adding clay is closer to the percentage
reduction in the diffusion coefcient of brine at 4C. In this
regard, we note that the moisture content in the nanocomposite
lms in equilibrium with 77% RH at 25C was slightly higher
than 11.96 mg/cm3, and it increased with increasing clay content.7 Had we used these values in the computation of the
steady concentration gradient in Eq. 4, the percentage reduction
in the diffusion coefcient values would have approached the
4C brine values. We also note that results of permeation
experiments available in the literature show relatively modest
fractional reductions in the diffusion coefcient on incorporating layered silicates within polymers.12-16 We believe that this
is explained, in part, by the fact that the clay is not properly
dispersed, and there are short circuits available to the diffusing
All of the above suggests that barrier properties of nanocomposites should be measured using steady-state permeation experiments because this is clearly a more severe test of improvement in barrier properties; transient sorption experiments are
evidently not very sensitive to the extent of clay exfoliation
when the diffusing molecules can interact strongly with the
clay. In other words, the use of sorption results may lead to
erroneous conclusions concerning the improvement of barrier
properties upon clay addition. In closing, we also note that in
our most recent work, and using the same system in conjunction with commercial watervapor permeation equipment (Mocon, Inc., Minneapolis, MN), we have found that17 measured
diffusivities are independent of the imposed relative humidity
Table 3. Measurements of Water Vapor Permeation at 25C
through Vinyl Ester Resin Films

Diffusion Coefcient 106


Neat resin
1 wt % Cloisite 10A
2 wt % Cloisite 10A
5 wt % Cloisite 10A


Vol. 51, No. 12


gradient and, to within experimental error, the temperature of

measurement does not inuence the relative decrease in diffusivity obtained on adding clay. Clearly, expectations associated
with a tortuous path being the cause of diffusivity reduction are
met in this situation, and the immobilization of moisture on the
clay surface does not inuence the steady permeation results.


The results of this research demonstrate that montmorillonite
can reduce moisture transport through polymers such as vinyl
ester. The measured reduction in moisture diffusivity on clay
addition, though, appears to be larger in sorption experiments
than that in permeation experiments, and this seems to be
related to the fact that results of sorption experiments are not
very sensitive to the extent of ller exfoliation when the diffusing molecule is strongly adsorbed on the ller surface.
Permeation experiments are thus preferable to sorption experiments, and these might even be used to estimate the effectiveness of a given processing technique to exfoliate clay platelets.



This research was funded by the U.S. Department of Transportation and
the Federal Highway Administration under contract number DTFH61-C00021. Professor Sridhar was on sabbatical leave at West Virginia University when this research was done.


Literature Cited
1. Freed AD, Gupta RK, GangaRao HVS. Minimizing voids in pultruded
polymer composites. Proc of the Society of Plastics Engineers, ANTEC, May 4 8, Nashville, TN; 2003:2105-2109.
2. Kajorncheappunngam S, Gupta RK, GangaRao HVS. Effect of aging
environment on degradation of glass-reinforced epoxy. ASCE J Compos Construct. 2002;6:61-69.
3. Cussler EL, Hughes SE, Ward WJ, Aris R. Barrier membranes. J
Membr Sci. 1988;38:161-174.
4. Yano K, Usuki A, Okada A, Kurauchi T, Kamigaito O. Synthesis and


December 2005


properties of polyimide clay hybrid. J Polym Sci Part A: Polym

Chem. 1993;31:2493-2498.
Shah AP, Gupta RK, GangaRao HVS, Powell CE. Moisture diffusion
through vinyl ester nanocomposites made with montmorillonite clay.
Polym Eng Sci. 2002;42:1852-1863.
Carrado KA. Polymer clay nanocomposites. In: Shonaike GO, Advani SG, eds. Advanced Polymeric Materials. Boca Raton, FL: CRC
Press; 2003:349-396.
Rana HT. Moisture Diffusion through Neat and Glass-Fiber Reinforced Vinyl Ester Resin Containing Nanoclay. MS Thesis. Morgantown, WV: Dept. of Chemical Engineering, West Virginia University;
Wu Z, Zhou C, Qi R, Zhang H. Synthesis and characterization of nylon
1012/clay nanocomposite. J Appl Polym Sci. 2002;83:2403-2410.
Zhou G, Lee LJ, Castro J. Nano-clay and long ber reinforced composites based on epoxy and phenolic resins. Proc of the Society of
Plastics Engineers, ANTEC, May 4 8, Nashville, TN; 2003:20942098.
Crank J. The Mathematics of Diffusion. 2nd Edition. Oxford, UK:
Clarendon Press; 1975.
Kumar A, Gupta RK. Fundamentals of Polymer Engineering. 2nd
Edition. New York: Marcel Dekker; 2003.
Tortora M, Vittoria V, Galli G, Ritrovati S, Chiellini E. Transport
properties of modied montmorillonite poly(-caprolactone) nanocomposites. Macromol Mater Eng. 2002;287:243-249.
Petrovicova E, Knight R, Schadler LS, Twardowski TE. Nylon 11/
silica nanocomposite coatings applied by the HVOF process. II. Mechanical and barrier properties. J Appl Polym Sci. 2000;78:2272-2289.
Manias E, Touny A, Wu L, Strawhecker K, Lu B, Chung TC. Polypropylene/montmorillonite nanocomposites. Review of the synthetic
routes and material properties. Chem Mater. 2001;13:3516-3523.
Chang YW, Yang Y, Ryu S, Nah C. Preparation and properties of
EPDM/organomontmorillonite hybrid nanocomposites. Polym Int.
Bhardwaj RK, Mehrabi AR, Hamilton C, Trujillo C, Murga M, Fan R,
Chavira A, Thompson AK. Structureproperty relations in crosslinked polyester clay nanocomposites. Polymer. 2002;43:3699-3705.
Bhardwaj M, Gupta RK. Water vapor diffusion through glass-ber
reinforced nanocomposites: A quantitative approach. Proc of the Society of Plastics Engineers ANTEC, May 1-5, Boston, MA; 2005:

Manuscript received Apr. 14, 2004, and revision received Apr. 4, 2005.

Vol. 51, No. 12

AIChE Journal

Interfacial Convection during Evaporation of

Binary Mixtures from Porous Obstacles
S. Post, I. Urukova, and E. Tsotsas
Thermal Process Engineering, Otto-von-Guericke-University, Universitatsplatz 2, D-39106 Magdeburg, Germany
DOI 10.1002/aic.10557
Published online September 1, 2005 in Wiley InterScience (www.interscience.wiley.com).

A novel experimental approach is proposed to investigate the inuence of interfacial

convection on liquid-side mass transfer. Specically, a binary liquid mixture (water and
isopropanol) is evaporated from a porous obstacle into inert gas (air). The change of liquid
bulk molar fraction with changing amount of liquid (concentration curve) is observed. The
porous obstacle is a plate with parallel planar gaps, or a packed bed of spherical particles,
both completely lled with the liquid. Without interfacial convection, that is, for the undisturbed state of the system, the liquid is quiescent and evaporation should be nonselective.
Consequently, measured selectivities not only prove the existence of interfacial convection, but
also give the opportunity of evaluating liquid-side mass transfer coefcients and enhancement
factors for the disturbed state of the system. This evaluation is conducted with the help of a
one-dimensional model for mass and heat transfer. If applied to the undisturbed state of the
system, the same model provides driving density and molar fraction differences in the form of
Rayleigh and Marangoni numbers. It is shown that the extent of interfacial convection can be
controlled by operating parameters such as the gas bulk temperature, whereas it also depends
on structural parameters such as gap width or particle diameter. Thermally driven and molar
fraction driven Rayleigh and Marangoni convection may coexist. With the described method
much higher enhancement factors than ever previously reported could be detected. The
dependency of such enhancement factors on the Rayleigh or Marangoni number is considerably stronger than reported in the literature. Some aspects concerning the consideration of
liquid-side heat transfer or permeability and structure of the medium are discussed, and
nonselective experiments corresponding to stable congurations are identied. First attempts
toward correlation of the results are outlined, along with the potential of the method for future
research. 2005 American Institute of Chemical Engineers AIChE J, 51: 32573274, 2005
Keywords: interfacial convection, enhancement of mass transfer, planar gaps, packed
beds, evaporation, binary liquid mixture, selectivity

Gradients of temperature or molar fraction (more general:
electrochemical potential) along the boundary of a liquid
phase, or between this boundary and the bulk, can give rise to
Present addresses: S. Post, BASF Aktiengesellschaft, GIC/PQ920, 67056 Ludwigshafen, Germany (stephan.post@basf-ag.de); I. Urukova, Applied Thermodynamics, University of Kaiserslautern, Erwin-Schrodinger-Strae, D-67663 Kaiserslautern,
Germany (urukova@mv.uni-kl.de).
Correspondence concerning this article should be addressed to E. Tsotsas at

2005 American Institute of Chemical Engineers

AIChE Journal

December 2005

spontaneous ow, which is called interfacial convection or

interfacial turbulence. This presupposes that either density or
interface tension gradients result from the gradients of the state
variables, so that we usually distinguish between Rayleigh and
Marangoni convection, respectively. Marangoni convection
may be driven by temperature differences (thermocapillarity)
or by molar fraction differences, and the same is true for
Rayleigh convection.
Stability analysis is one important instrument for the study of
such phenomena. Since the derivation by Lord Rayleigh, which
refers to a liquid layer with temperature-dependent density
Vol. 51, No. 12


heated from below and cooled at the boundary to a gas phase,

a vast amount of literature has appeared in this eld. Recent
publications expand the classical analysis by considering such
phenomena as waves of the deformable interface1 or the temperature dependency of physical properties other than density.2
Combined, thermally driven Rayleigh and Marangoni convection is treated by Bragard and Velarde3 with respect to the form
of the resulting cells, and Birikh et al.4 investigate boundary
conditions different from those in the original formulation of
the problem. Thermally driven Marangoni convection is discussed with respect to modulation by weak nonplanar shear by
Or and Kelly,5 spatially limited gas phase and pattern formation by Golovin et al.,6 and deep liquid layers by Hashim and
Wilson.7 Boeck and Thess8 apply direct numerical simulation
to study pattern formation, whereas Li et al.9 investigate
bridges, consisting of a cylindrical liquid phase surrounded by
a thin annulus of a second liquid. Approximate solutions for
deep liquids with various boundary conditions are provided by
Tan and Thorpe.10 Marangoni convection driven by molar
fraction gradients and reversible surface reaction is analyzed by
von Gottberg et al.11 and Mendes-Tatsis and Perez de Ortiz.12
Kang et al.13 apply the propagation theory to analyze numerically the stability during desorption at the interface without
gas-side mass transfer resistance, whereas Skarda and McCaughan14 approach analytically molar fraction or thermally
driven Marangoni convection in multicomponent mixtures. Control concepts with feedback strategies are discussed by Bau.15
Stability analysis for liquidliquid systems has been initiated
by the work of Sternling and Scriven,16 and developed further
by many investigators, including Perez de Ortiz and Sawistowski,17 Imaishi and Fujinawa,18 Reichenbach and Linde,19
and Wassmuth et al.20 Parallel to this has been the development
and application of experimental techniques for the visualization
and measurement of velocity and concentration elds (various
optical, Schlieren, and laser methods, holographic interferometry; see the recapitulation by Wolf21 and the recent reports by
Chai and Zhang22 and Guzun-Stoica et al.23).
However, stability analysis and visualization only partially
answer the question, how driving potentials for interface convection emerge during technical processes with mass and heat
transfer such as extraction, absorption, or distillation, and they
do not answer the still more important question about the
enhancement of such processes by the resulting interface convection. Experimental work coupled with empirical correlations and simplied models is necessary to this purpose. Such
work referring to liquidliquid extraction has been recapitulated by Pertler et al.24 and Wolf,21 and will not be discussed in
more detail here. It should just be noticed that planar liquid
layers, quiescent or mechanically agitated emulsions, stratied
channel ows, falling lms, and irrigated packed beds have
been investigated, with a clear trend to single-drop experiments
that promise a more fundamental observation and understanding of the relevant transport phenomena (see, for example,
Henschke and Pfennig25).
As to liquid gas or liquidvapor systems, the augmentation
of mass transfer by interfacial convection has been discussed
by Semkov and Kolev,26 Proctor et al.,27 and Wu et al.,28 who

December 2005

also summarize the results of older experimental investigations

on absorption or distillation conducted mainly in packed-bed
columns. Such work reveals, on the one hand, considerable
increases of separation efciency that can be attributed to
interfacial phenomena, but also clearly shows the difculties
related with an accurate evaluation of column experiments
arising from the inuence of axial dispersion, liquid distribution, wettability, and droplet formation. To avoid these difculties several authors29-31 followed the so-called spreading
method introduced by Ruckenstein and coworkers32 for the
experimental investigation of the impact of interfacial convection on liquid-side mass transfer during absorption. In the
spreading method, a solute with low surface tension is injected
from a capillary tube on the liquid gas interface, ows as a
thin lm in radial direction, induces Marangoni convection,
and enhances the rate of absorption of the gas in the liquid. A
theoretical treatment is possible, however, only after a number
of simplifying assumptions for the system that is complex
(three components in the liquid in the vicinity of the interface)
and, in fact, dissimilar to the case of driving potentials evolving
during mass transfer operations.
Along with a number of correlations, several more or less
simple models based on surface renewal, penetration, or cell
rotation concepts have been proposed in the literature by,
among others, Ruckenstein and Berbente,33 Linde and
Schwarz,34 Ostrovskii et al.,35 Imaishi et al.,36 and Pikkov and
Rabinovich.37 In these correlations and models the dependency
of enhancement factors on the Marangoni (or Rayleigh) number is described with exponents ranging between 0.25 and 0.67,
with the majority opting for an exponent near 0.5. Enhancement factors have values of up to 5, usually lower. Still more
unclear is the inuence of the Schmidt number. In a recent
publication Sun et al.38 point out the fortuity of knowledge
about how Rayleigh and Marangoni convection combine with
each other, and approach this topic on the basis of experiments
with absorption or desorption of a gas to/from a liquid lm.
Addressing the evident need for further data on the inuence
of interfacial convection on mass transfer, a novel experimental
method that can be used to this purpose will be presented. The
method is based on the evaporation of a binary liquid mixture
into inert gas, a process that has been thoroughly discussed by
many authors, especially by Schlunder and coworkers.39 The
goal of such investigations has been to nd out how the
selectivity of the evaporation depends on mass transfer. As we
will see later on, the analysis predicts vanishing selectivities in
the case of control by liquid-side mass transfer. However,
attempts to realize experimentally this limiting case by articially inhibiting the liquid-side mass transfer with the help of a
porous obstacle have rarely been successful, which was attributed to interface convection.40 In the present investigation, the
objective of previous work is inverted: that is, from understanding the inuence of mass transfer on selectivity to deriving from measured selectivities conclusions about the inuence
of interfacial convection on the liquid-side mass transfer.
Specically, the purpose of the present paper is to:
Present experimental apparatuses designed to enable the
determination of residual concentration curves during the batch
Vol. 51, No. 12

AIChE Journal

The discussion of the above-mentioned aspects will reveal

values of the enhancement factor in a range that has never been
investigated before, and lead to scarcely treated topics, such as
interfacial convection in porous media by a combination of
temperature and molar fraction proles.43 Although the
present, experimentally oriented work will, at best, only touch
such topics, we do hope that it might provide some additional
momentum and data for a complex and interesting eld of
engineering research. Although rst correlation attempts will
be outlined, they are only preliminary and not the focus of this
investigation. Completely outside of our present scope are ow
visualization, stability analysis, and two-dimensional modeling
of the transport phenomena in the porous obstacle. The same
holds for the possible extrapolation of the results to situations
without porous obstacle, such as those appearing in distillation
or absorption columns.

Experimental Apparatuses and Procedures

Figure 1. Apparatus for the investigation of the evaporation of a binary mixture from planar gaps into
inert gas.

evaporation of a binary liquid mixture from a porous obstacle

to inert gas in a reliable and reproducible way under welldened conditions.
Discuss experimental results and show that measured concentration curves and selectivities clearly prove the existence
and indicate the intensity of interfacial convection. Show that
the intensity of interfacial convection can be deliberately controlled by changing the bulk temperature of the gas phase, and
discuss the inuence of further operating parameters.
Present a simplied, one-dimensional model for the mass
and heat transfer and show, how liquid-side mass transfer
coefcients can be derived with its help from the measured
concentration curves. Such coefcients express the inuence of
interfacial convection on liquid-side mass transfer and can be
transformed to enhancement factors (Eh).
Show that the same model can be applied to the limiting
case without interfacial convection (undisturbed state) to derive
relevant driving forces (density and surface tension differences) in the form of Rayleigh (Ra) and Marangoni (Ma)
Discuss the phenomenology of the experimental data,
after their translation in Eh, Ra, and Ma. Show that four
different types of interfacial convection (Rayleigh convection
driven by temperature or molar fraction gradients, Marangoni
convection driven by temperature or molar fraction gradients)
can be responsible for the overall effects, in various combinations.
Slabs with a number of parallel planar gaps and packed beds
will be the porous obstacles corresponding to research work by
Post41 and Urukova.42 In both cases, the liquid system considered will be isopropanol and water, and the inert gas will be air.
AIChE Journal

December 2005

The experimental apparatus for evaporation from planar

gaps41 is depicted schematically in Figure 1. Its construction
fullls the following requirements:
Evaporation can take place only from the porous obstacle
that separates the liquid from the gas phase. To this purpose the
sliced, rectangular slab that constitutes the obstacle (Figure 2)
is inserted in a rectangular oater frame.
The oater is free to move within certain limits. Because
of this movability and initial ne adjustment with the help of
small weights, the planar gaps are completely lled with liquid
at any time, that is, the upper edge of the oater is identical
with the liquid surface. Consequently, the amount of liquid still
present after a certain evaporation time (Nl) can be calculated
from the position of the oater. The latter is measured with the
help of a laser sensor.
In spite of the freedom of movement of the oater, evaporation cannot occur through the gap between oater and the
cage of the device. Such bypass evaporation is suppressed by a
exible and air-tight bellows that separates the gas space above
the oater from the residual gas space of the apparatus. The
amount of vapor necessary to saturate the latter is negligible.
The temperature in the bulk of the gas phase above the
porous obstacle (Tg,B) is measured by a temperature sensor and
regulated to a prescribed, constant value by an external thermostat. Relatively large volume rates (V g ) of dried air (Y i,in
0) inserted in the apparatus by a distributor system create
well-mixed conditions in the gas bulk. Furthermore, they give

Figure 2. Cross-sectional view of the oater with insert,

geometrical parameters of the gaps.
Vol. 51, No. 12


Figure 3. Apparatus for the investigation of the evaporation of a binary mixture from packed beds
into inert gas.

rise to relatively large gas-side mass transfer coefcients and,

in combination with low evaporation temperatures, keep the
vapors dilute in the gas phase.
A pump recirculates in closed, tempered loop a certain
amount of the liquid. As a result of this recirculation, the molar
fraction is spatially uniform in the bulk of the liquid. The
respective value, x 1,B , is derived from density measurements
conducted in the recirculation loop by means of a highly
sensitive U-tube oscillatory device.
In spite of the mixing action of the recirculation in respect
to the bulk of the liquid phase, no forced convective mixing
takes place in the planar gaps of the oater insert. This has been
checked by variation of the recirculation rate, uid-dynamic
calculations, and tracer (color, salt) tests.
By temperature control of the recirculation circuit and of
the liquid-lled coat of the apparatus, the temperature in the
bulk of the liquid phase (Tl,B) can be kept to a prescribed,
constant and spatially uniform value. This value may be different from the temperature in the bulk of the gas phase.
The device for evaporation from a packed bed of particles42
is depicted schematically in Figure 3. It also fullls all the
discussed requirements. Major differences to the arrangement
according to Figure 1 are that:
The temperature in the bulk of the liquid phase (Tl,B) is set
to a prescribed value by putting the entire evaporation cell in a
heating chamber.
An external reservoir and a oater are used to constantly
keep the level of the liquid at the upper edge of the porous
The amount of liquid still present after a certain evaporation time (Nl) and the molar fraction in the bulk of the liquid
phase (x 1,B ) are calculated with the help of mass balances from
the molar fractions of the evaporating components in the outlet
gas measured by means of infrared spectroscopy. Precise calibration and consideration of cross-sensitivity provide an excellent accuracy of the IR readings.
As to the constructive details, the evaporation cell consists of
a cylindrical, stainless steel tube (inner diameter: 60 mm),
which iswith the exception of the connections for gas and
liquid hermetically closed. A second block of stainless steel
is placed into this tube. In this block, which is shorter than the

December 2005

outer tube, two cylindrical holes of equal diameters have been

machined. In this way, both the bulk of the liquid phase and the
space to be lled with the packed bed are properly dened. The
bed is supported by a thin-wire sieve, with apertures somewhat
smaller than the particle diameter. Advantages of the experimental design after Figure 3 are seen in the fact that the porous
obstacle is not moving, the liquid level is easier to adjust, and
the temperature control of the liquid phase is more efcient.
The desired output of the experiment is, in both cases, the
function Nl/Nl,0(x 1,B ) that describes the change of liquid bulk
molar fraction with decreasing amount of liquid, with reference
of the latter to the initial amount (Nl,0). This function is denoted
in the following as the concentration curve. Measurement of
the concentration curve up to very low values of Nl/Nl,0 with
the device of Figure 1 would require a large translatorial
movement of the oater in the vertical direction and a dilatability, which the used bellows cannot provide. Consequently,
the measurement of the entire concentration curve is not possible in one pass. Instead, the run has to be interrupted and the
device is lled with an additional amount of liquid (Nl) that
has the same molar fraction (x 1,B ) as the residual amount of
liquid at the specic time. After n repetitions of the procedure
the ordinate of the concentration curve is
N l,1
N l,2
N l,n1

N l,0 N l,0 N l,1 N l

N l,n N l


In the device of Figure 3 liquid is added in the external

reservoir. The molar fraction of residual liquid was determined
off-line before every relling of the reservoir by means of a
very accurate U-tube oscillatory densitometer. The resulting
values of x 1,B have always been in excellent agreement with
the values derived from the IR readings and the mass balances.
As a consequence of this way of measurement and of the very
moderate temperatures, the time necessary to determine one
complete concentration curve is long, up to one week. The
addition of liquid is conducted, depending on the specic
operating conditions, two or three times a day.
Although the long duration of one run is the price to be paid
for the segmentation of the experiment and for the moderate
temperature level, the latter hasin combination with the
relatively high gas ow rates used one major advantage: that
is, high dilution of the vapors in the gas phase, so that isopropanol and water diffuse simultaneously through the air in the
same way as the respective pure components would have done.
Consequently, it is not necessary to use the StefanMaxwell
equations for the gas phase, which facilitates the subsequent
evaluation. More important, the gas-side mass transfer coefcients of isopropanol and water can be determined by separate
evaporation experiments with the pure components, so that the
subsequent evaluation can concentrate on the liquid-side mass
transfer without any reconsideration or adjustment of gas-side
transport phenomena. The details and results of the derivation
of gas-side mass transfer coefcients and the Lewis exponent
(Eq. 14) from pure-liquid evaporation experiments are discussed by Post41 and Urukova,42 and will not be repeated here.
With reference to the same PhD theses, we also refrain from
repeating the correlations used for the determination of the
various relevant properties of the pure components and the
mixture, including binary diffusion coefcient, thermal conVol. 51, No. 12

AIChE Journal

ductivity, density, and surface tension of the liquid phase in

dependency of molar fraction and temperature. Good knowledge of the material properties has been one reason for the
choice of isopropanol (index: i 1) and water (index: i 2)
as the liquid system for the present investigation. The differences in density and surface tension of the pure components, as
a presupposition for the investigation of respective instabilities
are the second reason. Remember that isopropanol has both
lower density and surface tension (l,1 781.3 kg/m3, l,1
0.02333 N/m at 25C) compared to those of water (l,2 997.0
kg/m3, l,2 0.07540 N/m at 25C), and that both the density
and the surface tension decrease with increasing temperature
over the entire molar fraction range.
As to the porous obstacles, the sliced slabs according to
Figure 2 consisted of a specialty polyurethane polymer (Prolab
50, Axson Ltd.) with nonporous and homogeneous structure,
low density, and very low thermal conductivity. The parallel,
planar gaps have been cut out very accurately by means of a
laser manufacturing technique. Slabs of different gap widths
(GW 0.5, 1.0, and 1.5 mm) and different gap depths (GD
1.0, 2.0, 4.0, and 8.0 mm) have been used. The length of the
gaps varied between 16 and 70 mm. Because this dimension
has not been found to be of signicance, it will not be stated in
the subsequent presentation of experimental results. The number of the gaps varied between 24 and 144 and is also without
signicance for the results. However, a sufcient number of
parallel gaps is necessary to keep the time necessary to complete one experiment within tolerable limits. The investigated
packed beds were made of polished glass spheres. Five closely
monodispersed fractions with average particle diameters of dp
0.23, 0.57, 1.30, 2.60, and 3.50 mm have been used. The bed
height h varied in value from 5 to 25 mm, with bed porosities
ranging in value from 0.35 to 0.38. The permeability of the
beds, dened by

d p2 3
150 1 2


ranged in value from 4.02 1011 to 9.30 109

m2. The particle density was determined as p 2470 kg/m3
and the particle thermal conductivity was, according to the
producer, p 0.8 W m1 K1.
Without spontaneous convection, the evaporating components must overcome the entire porous obstacle by molecular
diffusion before reaching the interface. The gaps or beds may
thus be considered as the analogon of a very thick laminar
boundary layer according to lm theory. This provides a clear
reference value for the liquid-side mass transfer coefcient,
with consequences that will be discussed in the next section.
Prediction of the limiting case without a porous obstacle (see
next section) provides an overall validation of the experimental
technique. The reproducibility of measured concentration
curves was excellent.

Experimental Results for Evaporation from Planar

Gaps and Their Interpretation in Terms of
Interfacial Convection
In a total of 29 experiments conducted with sliced slabs as the
porous obstacle, the liquid bulk temperature and the gas ow
rate were kept constant at Tl,B 298.15 K and V g,0 8 L/min
AIChE Journal

December 2005

Figure 4. Concentration curves for the evaporation of a

binary mixture from planar gaps into inert gas;
variation of temperature in the bulk of the gas
phase, Tg,B, at x1,B,0 0.70; measurements,
limiting cases, and calculations (solid lines) by
punctual adaptation of the liquid-side mass
transfer coefcient l.

(referring to standard conditions, that is, to 0C and 101,300

Pa), respectively. The initial molar fraction in the bulk of the
liquid phase x 1,B,0 was either approximately 0.70 or 0.30, lying
at one or the other side of the thermodynamic azeotrope. The
temperature in the bulk of the gas phase Tg,B, the width of the
gaps GW, and the depth of the gaps GD have been varied.
Measured concentration curves resulting from this variation are
presented in Figures 4 to 7. The data for all experiments are
documented by Post.41
All the plots have two asymptotic limiting cases in common,
that is, for very small and for very large liquid-side mass
transfer coefcients (l 3 0 and l 3 , respectively). At l
3 0 the process is controlled by liquid-side mass transfer.
Because the mixture is binary in the liquid phase, this transport
step is nonselective. Consequently (see also next section), the
selectivity of evaporation is zero at l 3 0, which corresponds
to a vertical line in the plots. On the other hand, the selectivity
is maximal for l 3 , determined by thermodynamic equilibrium and mass transfer in the gas. Because of the presence of
air, the latter is selective in favor of water, that is, of the smaller
molecule. The limiting case l 3 can be realized experimentally by removing the porous obstacle and evaporating the
well-agitated mixture from the free liquid surface. Under the
already discussed conditions of dilution in the gas phase, and
after the separate determination of gas-side mass transfer coefcients from pure-liquid experiments, the concentration
curves can be predicted at the limit l 3 . As shown in
Figures 4 and 5, the agreement of these predictions with the
corresponding measurements is excellent.
With porous obstacle, but without spontaneous interfacial
convection, the value of the liquid-side mass transfer coefcient would be
Vol. 51, No. 12


Figure 5. Concentration curves for the same conditions

as in Figure 4, although at x1,B,0 0.30.

l,0 D l/G D


where Dl is the binary diffusion coefcient in the liquid. For

every used depth of the gaps this value is so small that it may
be considered to very closely represent the limiting case l 3
0. Consequently, all concentration curves of the Figures 4 to 7
should correspond to vertical lines without any measurable
selectivity. In other words, every binary liquid of any initial
concentration should behave as an azeotropic mixture. Vice
versa, every deviation from the described behavior is an unequivocal proof for the existence of spontaneous, interfacial
convection in the gaps. As a glance at Figures 4 to 7 reveals,
such convection is the rule, not the exception. Furthermore, the
extent of the deviation from the vertical line toward the opposite limit of l 3 , that is, the selectivity of the process, is a
measure for the intensity of convection in the gaps.
In light of the foregoing general remarks, the phenomenology of the experimental ndings can be discussed in some
more detail. Starting with an initial molar fraction in the bulk
of the liquid phase of x 1,B,0 0.70, that means at the righthand side of the thermodynamic azeotrope, as in Figure 4, any
existing selectivity can only be in favor of water. Because
water is the component with the higher density and surface
tension, depletion of the interface from water molecules means
a decrease of density and surface tension of the interface
compared to the density in the bulk of the liquid and the surface
tension that liquid with the bulk molar fraction would have.
This is a stable conguration with respect to both molar fraction driven Rayleigh and molar fraction driven Marangoni

December 2005

convection. Consequently, the interfacial convection evident

by three of the four curves of Figure 4 can only be thermally
driven, a fact additionally supported by the nding that its
extent can be controlled by changing the temperature in the
bulk of the gas phase, Tg,B. Indeed, the intensity of convection
increases with decreasing gas bulk temperature until the limit
of l 3 is approximated. Vice versa, microconvection slows
down with increasing gas bulk temperature and is not detectable at Tg,B 328.15 K. This behavior can be explained by the
fact that with decreasing gas bulk temperature the temperature
of the liquid gas interface also decreases until it becomes
smaller than the temperature in the bulk of the liquid phase.
Then, both density and surface tension at the interface become
larger than the values corresponding to the state of the bulk
liquid. As soon as some threshold is exceeded, interfacial
convection is triggered and intensied by any additional decrease of the gas bulk temperature.
At the lefthand side of the azeotrope (x 1,B,0 0.30; Figure
5), not only low gas bulk temperatures, but also the selective
evaporation of alcohol can lead to instability, so that temperature-driven and molar fraction driven Rayleigh and Marangoni convection may coexist. In spite of this, the measured
concentration curves stay at a considerable distance from the
limiting case l 3 . The inuence of the gas bulk temperature
is not so pronounced as in Figure 3, and the nonselective
limiting case (l 3 0) cannot be approached. The ndings
indicate nonlinearities in the combination of the mentioned
four possible types of interfacial convection.
Space restriction in the lateral direction inhibits the mixing
inuence of micro convective ow, as the results of Figure 6
from experiments with different gap widths (GW), under otherwise the same conditions, show. For the operating conditions
of Figure 6 selectivity and interfacial convection could be
completely suppressed by use of the narrowest gaps with GW
0.5 mm. The same trend as in Figure 6 has been observed for
every investigated gap depth.
Less clear, however, is the inuence of variation of the gap
depth GD. As the data of Figure 7 show, the intensity of
interfacial convection appears to increase for a gap width of GW
1 mm when increasing the gap depth (GD) values from 1 to
4 mm, reach a maximum, and decrease again when approaching GD 8 mm. The same behavior has been found with GW
1.5 mm, but not with GW 0.5 mm. In the latter case the
selectivity was observed to decrease continuously when increasing the gap depth values from 1 to 8 mm, having no
detectable value at GD 4 mm (compare with Figure 6) and
GD 8 mm. Penetration of the forced convective ow from the
bulk of the liquid in shallow gaps cannot be the reason for the
outlined behavior. Not only has such penetration been excluded
by various tests (see Experimental), but it would also lead to
the opposite behavior than observed in Figure 7 for a change of
gap depth values from 1 to 2 mm. A satisfactory explanation
for the inuence of gap depth on the intensity of microconvection is still missing.

Derivation of Liquid-Side Mass Transfer

Coefcients and Enhancement Factors
To quantify the ndings about the inuence of interfacial
convection on the liquid-side mass transfer, a one-dimensional
Vol. 51, No. 12

AIChE Journal

transfer coefcients g,i are determined with the help of pureliquid evaporation experiments.
Contrary to the treatment of the gas phase, Stefan uxes
must not be neglected in the liquid, so that the relationship
N i n lA Ph lr iln

ri xi,Ph
ri xi,B



applies for liquid-side mass transfer kinetics. At the limit l 3

0, r i x i,B 0 is obtained from Eq. 8. Because every
component evaporates in the same proportion as it is present in
the mixture (r i x i,B ), the selectivity of the process becomes
zero, and the liquid composition does not change during the
process (x 1,B x i,B,0 ). The presence of a third species in the
liquid would radically change this situation. Surplus of a component with high boiling temperature (such as glycerin) would
kinetically favor the evaporation of the smaller molecule (water) in the liquid phase, in the very same way as air favors the
evaporation of water in the gas phase (Eq. 7). In general,
StefanMaxwell equations must be written for mass transfer in
multicomponent liquid mixtures, and the condition of vanishing selectivity at l 3 0, which is typical for the binary liquid
mixture, no longer applies.
At the phase boundary equilibrium according to
Figure 6. Concentration curves for the evaporation of a
binary mixture from planar gaps into inert gas
at different widths of the gaps (GW); measurements, limiting cases, and calculations (solid
lines) by punctual adaptation of the liquid-side
mass transfer coefcient l.


y i,Ph i x i,Ph p i0

x i,Ph


is assumed and evaluated with the help of Antoine equations

for the saturation vapor pressures, p 0i , and the UNIQUAC
method for the activity coefcients, i.41
Notice that the above system of equations has been discussed

model for mass and heat transfer in the gas and the liquid phase
has been used. The respective mass balances for the liquid are
N1 N2 N



N x
dt l i,B



with i 1, 2 the component index and N i the molar evaporation rates. For the gas it is
N i N gY i,out


with N g the molar ow rate of the inert carrier (air), and Y i,out
the molar content of vapors at the outlet of this stream (Y i,in
Because of the already discussed high dilution, gas-side
mass transfer kinetics may be linearized and expressed in terms
of molar content; (for y i 1 it is: Y i y i ). For the ideally
backmixed gas phase it, then, holds
N i n gA Ph g,iY i,Ph Y i,out


Herein APh is the gasliquid phase contact area (crosssectional area of the gaps). It has been pointed out that the mass
AIChE Journal

December 2005

Figure 7. Concentration curves for the evaporation of a

binary mixture from planar gaps into inert gas
at different depths of the gaps (GD); measurements, limiting cases, and calculations (solid
lines) by punctual adaptation of the liquid-side
mass transfer coefcient l.
Vol. 51, No. 12


in detail by previous authors,39 including the denition and use

of dimensionless separation factors for thermodynamic equilibrium, gas-side kinetics and liquid-side kinetics, limiting
cases, simplied analytical solutions, and the behavior of socalled pseudoazeotropic points.
The energy balances for the gas and for the liquid boundary
layer are

N h T

N gh g,in h g,out

i g,i






N h T



c p,l,iT Ph T l,B Q




which means by assuming equally thick boundary layers for

temperature and concentration in the liquid and by neglecting
heat transfer through the solid.
In this way, the only unknown in the system of Eqs. 4 to 16
is the liquid-side mass transfer coefcient, l. This is calculated
by tting the solution of the model to the experimental results
for every concentration curve with measurable deviation from
the limiting cases l 3 and l 3 0. To this purpose, the
measured concentration curve is discretized in the direction of
its abscissa with a short x 1,B step, and the tting is repeated
at every point, leading to a trajectory of l for every experiment, and to the solid lines located between the limiting cases
in Figures 4 to 7. Every value of l derived by this method is
a quantitative expression for the inuence of interfacial convection on the liquid-side mass transfer. It characterizes the
disturbed state of the system, is larger than the respective value
of l,0 for the undisturbed state (only molecular diffusion; see
Eq. 3), and may be transformed by reference to the latter to an
enhancement factor

g ) and
respectively. The heat ow rates coming from the gas (Q

from the liquid (Q l ) are calculated by the kinetic relationships

g gK A,gA PhT g,B T Ph


l lK A,lA PhT l,B T Ph



The dimensionless quantities KA,g and KA,l, so-called Ackermann corrections that account for the temperature change of
the evaporation streams in the boundary layers, are for the
conditions of the present investigation approximately equal to
The gas-side heat transfer coefcient g is calculated according to the Lewis analogy

n gc p,gLein


ng cp,g Dg,i



from the gas-side mass transfer coefcients g,i and the exponent n, as determined separately from pure-liquid evaporation
experiments. In Eq. 15 Dg,i are the binary diffusion coefcients
of isopropanol (index: 1) and water (index: 2) in air, g and c p,g
are the thermal conductivity and the molar specic capacity of
the gas mixture. In this context, notice that material properties
referring to the phase bulks (index: B) or to the gasliquid
interface (index: Ph) are always calculated at the local values
of concentration and temperature, whereas material properties
referring to the boundary layers are calculated at the respective
arithmetic averages.
The liquid-side heat transfer coefcient l has been related
to the liquid-side mass transfer coefcient l according to


l l

December 2005




that expresses the intensication of liquid-side mass transfer by

interfacial convection. By inserting Eq. 3 in Eq. 17 it is obvious
that the enhancement factors could also be interpreted as liquid-side Sherwood numbers. The solution of Eqs. 4 to 16 is
iterative (because of multiple coupling between the individual
equations and the dependency of material properties on temperature and concentration). Brents method (reported in Press
et al.44) has been used for the described optimization.
Before proceeding in the next section with the discussion of
enhancement factors, some characteristics of the thermal part
of the model should be claried. To this purpose, concentration
curves that have been calculated with the complete model
(curves 1, 2, and 3) are compared in Figure 8 with curves
obtained by considering only the mass transfer (curves 4, 5, and
6). The same temperature has been assumed to prevail in the
bulk of the gas and the liquid in all calculations (Tg,B Tl,B
298.15 K). Because the liquid-side heat transfer coefcient is
proportional to l (Eq. 16), the consideration of heat transfer
has no inuence at the limit l 3 ; (in any case it is: TPh
Tg,B Tl,B). Isothermal conditions are also valid at nite values
of l, if heat transfer is neglected (curves 4, 5, and 6). However,
by consideration of heat transfer at nite l, Tad TPh Tl,B
Tg,B is obtained. Because of lower interface temperatures
and evaporation rates, somewhat higher selectivities are predicted when heat transfer is taken into account. The curves
spread out in this range (compare curve 2 with curve 5), to fall
once again together in the limiting case of l 3 0. Although
TPh reaches its minimal value (the adiabatic, wet-bulb temperature Tad), the direct inuence of l dominates at l 3 0.
Figure 9 shows the temperature at the interface (TPh) for
those calculations from Figure 8 that have been conducted with
consideration of heat transfer. One interesting message of Figures 8 and 9 is that the consideration of heat transfer is not very
important at relatively large values of l and l, which means
for the disturbed state with considerable microconvection. In
this case the difference between interface temperature TPh and
liquid bulk temperature Tl,B is relatively small (Figure 9), and
Vol. 51, No. 12

AIChE Journal

Figure 10. Interface molar fractions corresponding to

the results of Figure 9.

Figure 8. Concentration curves for the evaporation of

isopropanolwater into air, as calculated for
different liquid-side mass transfer coefcients
with () and without () consideration of heat
transfer from the liquid phase.

the change of selectivity arising from deviations from isothermal conditions is moderate (Figure 8). Consequently, inaccuracies in the calculation of the liquid-side heat transfer coefcient l according to Eq. 16 are not expected to be crucial for
the described determination of enhancement factors. This sit-

Figure 9. Interface temperatures corresponding to the

calculations of Figure 8 that have been conducted with consideration of heat transfer
from the liquid phase.
AIChE Journal

December 2005

uation changes, as soon as the undisturbed state of the system

is considered.
The interface molar fractions x 1,Ph , resulting from the same
calculations as in Figure 9, are plotted in Figure 10. Water is
preferentially removed within the depicted region, so that the
concentration of isopropanol is always larger at the interface
than in the bulk of the liquid.
Finally, some values of the adiabatic temperature (Tad) are
illustrated in Figure 11 over the entire concentration range. The
solid lines correspond to l 3 0 and l 3 0, whereas the
broken curves have been calculated with l 3 and l 3 0,
inconsistent with Eq. 16. Given the fact that isopropanol has
the higher vapor pressure, but the lower evaporation enthalpy,
the adiabatic temperatures of the pure components are not very
different from each other. The minimal value of Tad is obtained
at the pseudoazeotropic point of the system; (compare with the
thermodynamic azeotrope at approximately x l,B 0.66).

Figure 11. Adiabatic temperatures (l 3 0) of the mixture isopropanolwater in air of various bulk
temperatures for l 3 0 and l 3 .
Vol. 51, No. 12



l gGD3
l Dl


l GD
l Dl




and are assigned to the respective enhancement factors. This

procedure enables one to interrelate driving potentials obtained
from the undisturbed state of the system with enhancement
factors derived from the experimental data and allocated to
what is here described as the disturbed state. To control the
validity of this allocation would require a rigorous, two-dimensional and dynamic treatment of transport phenomena in the
porous obstacle, including momentum transport. Although
such a theoretical approach may provide access to the dynamic
sequences of instability development and destruction by developing interfacial convection, it is outside the scope of the
present investigation.
In Eqs. 19 and 20 g is the gravity acceleration and l is the
dynamic viscosity of the liquid. The latter, as well as the
diffusion coefcient Dl, have been calculated at temperatures
and concentrations corresponding to the arithmetic averages
between interface and liquid bulk.
The discussed evaluation is exemplied by selected results
in Figures 12 and 13. Figure 12 refers to the experimental data
of Figure 4 and presents enhancement factors against the sum
of the Rayleigh and the Marangoni number. In this sum only
positive values of Ra and Ma are considered; negative values
are set equal to zero. Such negative values would mean that the
respective driving potential for interfacial convection does not
exist, that is, that the conguration is stable with respect to
either density or surface tension. This is indeed the case for
curve 4 from Figure 4, as the separate presentation of Ra and
Ma in Figure 13 reveals. The enhancement factor for this case
is equal to unity, and is not plotted in Figure 12. Decreasing the
gas bulk temperature to Tg,B 308.15 K, which means going
to curve 3 in Figures 4, 12, and 13, does not change the fact of
negative Marangoni numbers (Figure 13). However, the Rayleigh number becomes now positive, which is accompanied by
microconvection (Figure 4) with enhancement factors some-

The quantity l,0 is also very small, so that the interface

temperature TPh is expected to lie in the vicinity of the adiabatic temperature Tad (lowest curve in Figure 9, solid lines in
Figure 11). It has been decided to conduct the calculations for
the undisturbed state of the system at the limit l 3 0 at not
only approximately, but exactly adiabatic conditions. Under
such conditions both differences (TPh Tl,B) and (x 1,Ph x 1,B )
attain their maximal possible values (see Figures 9 and 10).
Some critical remarks on this assumption will be given later on.
The calculation for the undisturbed state is conducted at the
same points along the abscissa of the concentration curve,
which means at the same values of x 1,B , as the previous
calculation for the disturbed state. In this way, densities and
surface tensions are calculated at the interface (l,Ph resp. l,Ph).
With the differences of these values to those corresponding to
the conditions of the liquid bulk (that is with l l,Ph l,B,
l l,Ph l,B), Rayleigh and Marangoni numbers are
determined according to the denitions

Figure 13. Separate presentation of Rayleigh and Marangoni numbers from Figure 12.

Figure 12. Enhancement factors vs. the sum of the Rayleigh and Marangoni number for the experimental data of Figure 4, that is, at different
temperatures in the bulk of the gas phase for
x1,B,0 0.70.

Dimensionless Driving Forces of Interfacial

Convection and their Inuence on Enhancement
Values of l and Eh calculated according to the previous
section express the inuence of interfacial convection on liquid-side mass transfer. To also quantify the potentials giving
rise to microconvection, the same model (Eqs. 4 15) has been
applied to the undisturbed state of the system, which means to
the state with only molecular transport mechanisms in the
porous obstacle. In this case, the liquid-side mass transfer
coefcient is known and equal to l,0 after Eq. 3. As already
pointed out, l,0 is very small, corresponding, in terms of
selectivity, to the limit l 3 0. Concerning the heat transfer
from the liquid, combination of Eq. 3 with Eq. 16 would lead




December 2005

Vol. 51, No. 12

AIChE Journal

what smaller than 10 (Figure 12). Because for x 1,B,0 0.70 it

always is x 1,Ph x 1,B (stable situation in respect to concentration differences; compare with Figure 10), the reason for this
enhancement of liquid-side mass transfer is thermally driven
Rayleigh convection. This type of convection is intensied by
further decrease of the gas bulk temperature, and is then
accompanied by thermally driven Marangoni convection, either from the very beginning of the experiment (curve 1 in the
lower part of Figure 13), or from some later point in the course
of evaporation (curve 2 in the same plot). The enhancement
factors increase to values between 20 and 40, far above the
enhancement factors of maximally 5 that have been previously
reported in literature (see Introduction).
A similar evaluation can be conducted for the data of Figures
5 to 7. It has shown enhancement factors of up to 60, regions
with Ma Ra and very large values of the sum Ra Ma, and
regions with coexisting molar fraction and temperature-driven
convection. On the other hand, considerable overlap has been
observed between curves in the Eh vs. (Ra Ma) plots,
indicating inuence by additional dimensionless quantities,
nonlinearities in the combination of the driving potentials, but
also some inuence by the consideration or not of gap width
GW and depth GD, in the denitions of Ra, Ma, and Eh after
Eqs. 19, 20, 17, and 3.
In a rst attempt to correlate the results of the evaluation,
which are all documented by Post,41 it has been assumed that
the dynamic energy created by density and surface tension
differences between the interface and the bulk of the liquid is
transformed to kinetic energy. Using the same length scale (the
gap depth GD) in both the density and surface tension dynamic
energy terms

Figure 14. Measured and calculated enhancement factors for interfacial convection in planar gaps.

In the next step, the equivalent Reynolds number is considered to be proportional to the enhancement factor, the Schmidt
number is replaced by a more general damping function (DSc),
and factor and exponent of Eq. 22 are given free for tting.
Several trials41 have led to the damping function
D Sc

x 0.08358
1 exp


w 2 g lG D l/G D
2 l


and, after rearrangement,

Ra Ma
Re 2

Eh 5 10 3


and the Schmidt number




This is analogous to the usual derivation of simplied correlations for heat or mass transfer by combined forced and free
convection, shows that the use of the sum Ra Ma is not
completely arbitrary, and introduces Sc as the additional inuence parameter. On the other hand, it can only be coarsely
approximate. Indeed, even in the case of conventional forced
and free convection, nonlinear, supporting, or competitive
combinations are well known.45
AIChE Journal

ScPh ScB disturbed

ScPh ScB undisturbed


and to the correlation

is obtained with the equivalent Reynolds number



December 2005

Ra Ma



Although the exponent of Ra or Ma is 0.5 after Eq. 22, it

increases to 0.8 after tting Eq. 27. The former is in line with
the majority of existing literature; the latter isto our knowledgethe highest value ever proposed (see Introduction). This
is not necessarily unreasonable because both the enhancement
factors and the dimensionless driving potentials lie in a region
of high intensication of liquid-side mass transfer by interfacial
convection that has never been investigated before. In fact, the
exponent of the Reynolds number shows in the case of conventional forced convection a similar trend of increase with the
respective argument.
The results of the correlation are compared with measured
enhancement factors in Figure 14, plotted against the equivalent Reynolds number after Eq. 22 without the constant coefcient. Most points can be reproduced with deviations of less
than 100%. Although this is not too disappointing with
respect to the complexity of the involved phenomena and the
lack of any previous data within the investigated region, it
Vol. 51, No. 12


Figure 15. Concentration curves for the evaporation of a

binary mixture from a packed bed into inert
gas at different temperatures in the bulk of
the gas phase (Tg,B) and x1,B,0 0.70; measurements, limiting cases, and calculations
(solid lines) by punctual adaptation of the liquid-side mass transfer coefcient l.

certainly allows substantial room for improvement. In this

context, it should be noticed that there is a ne structure in the
data, which is not reected by the correlation. Additionally, the
proposed damping function requires calculations for both the
disturbed and the undisturbed state (Eq. 26), is very sharp (Eq.
25), and arbitrary. That the value Eh 1 is reached at nite
positive values of the argument in Figure 14 is reasonable,
given that the addressed problem is a stability problem. However, no comparison with the total of three experiments without
measurable selectivity has yet been conducted in the sense of
threshold values of the dimensionless driving forces.

tion is possible, arising from preferential evaporation of alcohol, and may assist the thermally initiated interfacial convection. Gas bulk temperature loses its prevailing role as the
operating parameter controlling selectivity and convection.
This is similar to the observations with planar gaps (compare
with Figure 5). A further similarity exists in the inhibition of
the mixing inuence of micro convective ow by lateral space
restriction. The latter has been realized in the case of packed
beds by systematic decrease of the particle diameter dp. As the
data of Figure 17 show, the evaporation was nonselective for
the smallest particles with dp 0.23 mm, which is equivalent
to the complete absence of interfacial convection. An inhibition
of the mixing inuence of interfacial convection has also been
observed by increasing the bed height h, as the data of Figure
18 exemplify. This trend was valid for any combination of the
remaining operating and geometrical parameters. In this respect, a difference of behavior appears to exist between packed
beds and planar gaps because the impact of gap depth on
interfacial convection has been found to be ambiguous in the
case of planar gaps (see Figure 7).
Although the further evaluation of measured data was in
principle the same as that for planar gaps, some modications
have been carried out in the kinetic equations of the model, and
in the denition and calculation of dimensionless numbers.
These modications will be discussed in the following.
Concerning the model, not the cross-sectional area of the
gaps, but the total cross-sectional area of the space lled with
the packed bed (A0) is used in the kinetic equations for gas-side
mass transfer, liquid-side mass transfer, gas-side heat transfer,
and liquid-side heat transfer leading, after omission of the
Ackermann factors, to
N i n g A 0 g,iY i,Ph Y i,out
N i n lA 0 lr iln

ri xi,Ph
ri xi,B


Experimental Results for Evaporation from

Packed Beds and their Evaluation
A total of 36 experiments have been conducted with liquid
mixtures and a packed bed as the porous obstacle, varying the
temperature in the bulk of the gas phase Tg,B, the particle
diameter dp, and the height of the bed h.42 Gas ow rate, liquid
bulk temperature, and initial molar fraction in the bulk of the
liquid phase were the same as with sliced slabs. Measured
concentration curves are presented in Figures 15 to 18 along
with the limiting cases of l 3 0 and l 3 .
Figure 15 indicates the same inuence of gas bulk temperature (Tg,B), as already observed with planar gaps. However,
comparison with the corresponding plot (Figure 4) reveals a
lower intensity of interfacial convection in the packed bed than
in the gaps at the same gas bulk temperature. Lower permeability of the packed bed is the reason for this behavior.
At the left-hand side of the azeotrope (Figure 16, x 1,B,0
0.30), molar fraction driven Rayleigh and Marangoni convec3268

December 2005

Figure 16. Concentration curves for the same conditions as in Figure 15, although at x1,B,0 0.30.
Vol. 51, No. 12

AIChE Journal

D l,bed D l1 1


which means they are calculated with the effective diffusion

coefcient of the bed (Dl,bed), instead of the binary liquid
diffusion coefcient (Dl). Equation 34 is a well-known empirical correlation for diffusion in granular or porous media with
compact and inert solids, but also for thermal, electrical, or
other conduction at the limit of nonconducting particles.47 It
includes the inuence of tortuosity, which is not present in the
previously investigated planar gap geometry. That both tortuosity and space restriction are accounted for in the kinetic
coefcient, and not explicitly in Eq. 29, is usual practice in the
treatment of transport phenomena in porous media from the
macroscopic perspective.
The liquid-side heat transfer coefcient l may, again, be
assumed proportional to l according to the relationship

Figure 17. Concentration curves for the evaporation of a
binary mixture from packed beds into inert
gas at different particle diameters (dp); measurements, limiting cases, and calculations
(solid lines) by punctual adaptation of the liquid-side mass transfer coefcient l.

g gA 0T g,B T Ph


l lA 0T l,B T Ph


l l
D l,bed


Comparison with Eq. 16 reveals that the thickness of the

assumed molar fraction boundary layer is smaller for the
packed beds than for the planar gaps. Because the thermal
boundary layer is supposed to have the same thickness, liquidside heat transfer coefcients after Eq. 35 are larger than those
after the previous Eq. 16, indicating that the formerly neglected
heat conduction through the solid phase is nowat least indirectlyaccounted for. However, this is not very important for
the determination of enhancement factors because for the relatively large values of l and l of the disturbed system the
difference between interface temperature TPh and liquid bulk
temperature Tl,B is small (Figure 9), and the isothermal limiting
case is closely approximatedmore closely for the packed

instead of Eqs. 7, 8, 12, and 13. A correction factor () is

applied to the total cross-sectional area A0, to obtain the effective evaporation area in Eq. 28. This factor has been derived
from evaporation experiments with the involved pure liquids to



The same separate experiments42 provide the gas-side mass

transfer coefcients, g,i. Equations 28 and 32 dene an evaporation area that is smaller than the total cross-sectional area
A0, but larger than the cross-sectional area (A0) occupied by
the liquid phase in the interior of the bed. This accords well
with the fact that the local porosity of the packed bed increases
as its free surface is approached, given that it also increases in
the vicinity of rigid walls in tubular packed-bed reactors.46
Fitting l to the measured concentration curves of experiments with interfacial convection leads to the solid lines between the limiting cases of l 3 and l 3 0 in Figures 15
to 18. Enhancement factors are obtained from Eq. 17. However, the liquid-side mass transfer coefcients of the undisturbed system are now set to


D l,bed


AIChE Journal

December 2005

Figure 18. Concentration curves for the evaporation of a

binary mixture from packed beds into inert
gas at different bed heights (h); measurements, limiting cases, and calculations (solid
lines) by punctual adaptation of the liquidside mass transfer coefcient l.
Vol. 51, No. 12



Figure 19. Enhancement factors vs. the sum of the Rayleigh and Marangoni number for the experimental data of Figure 17 (different particle
diameters, dp).

beds than for the planar gaps. Thus, the approximate character
of Eq. 35 will not bias the evaluation of the disturbed state of
the system.
More critical is the consideration, or not, of liquid-side heat
transfer for the undisturbed state of the system. Respective
calculations have been carried out at the adiabatic limit of l 3
0 (as for the planar gaps), but also with heat transfer coefcients after




In Eq. 36 the thermal conductivity of the quiescent bed (bed)

is calculated according to Zehner and Schlunder (see Tsotsas47). Equation 34 is derivable from the same model by
inserting p 0, that is, by assuming nonconducting particles.
However, application of Eq. 36 resulted in computing negative
density and surface tension differences (stable congurations)
for some experiments that were clearly accompanied by interfacial convection.42 Such discrepancies never occurred when
calculating at the adiabatic limit, so that the nal evaluation for
the undisturbed state of the system has been conducted at
adiabatic conditions. Although this leads to a coherent interpretation of all experimental ndings, the assumption of l 3
0 is more questionable for packed beds than for the sliced slabs.
The reason for the partial failure in the application of Eq. 36 is
seen in the fact that especially in geometrically complex
media such as packed bedslocal temperature differences in
both axial and lateral direction may induce interfacial convection, but cannot be described with our present model, which is
homogeneous and one-dimensional.
Finally, the density and surface tension differences obtained
from the undisturbed-state calculations for the packed beds
were transformed to so-called DArcy-modied Rayleigh and
Marangoni numbers dened with the permeability , and the
bed diffusion coefcient Dl,bed, as

l gh
l Dl,bed

December 2005

l Dl,bed h


The use of the permeability (Eq. 2) in these relationships

enables one to account for the lateral space restriction in the
porous obstacle, which is not possible with the previous Eqs.
19 and 20. In the latter, the gap depth GD is the only length
scale. The dynamic viscosity l and diffusivity Dl (see Eq. 34)
of the liquid are calculatedas beforeat the average values
of temperature and molar fraction between interface and liquid
Figure 19 exemplies the evaluation for packed beds by
presenting enhancement factors derived from the experiments
of Figure 17. As expected, the enhancement factor decreases
with decreasing particle diameter dp, and increases with increasing sum of Ras and Mas. Because of the consideration of
the permeability of the packed beds , in the denition of Ras
and Mas, the curves do not overlap. The separate presentation
of Ras and Mas in Figure 20 indicates Rayleigh and Marangoni
convection, which is thermally driven, because the initial molar
fraction of the liquid lies at the right-hand side of the azeotropic
point (x 1,B,0 0.70). For the smallest particle diameter (dp
0.23 mm, lowest permeability), the evaporation is nonselective
(Figure 17), and the enhancement factor is equal to unity.
Howeverand in contrast to all nonselective experiments with
planar gapsthe values of the Rayleigh and Marangoni numbers are not negative for the experiment with dp 0.23 mm. In
fact, Ras and Mas have small positive values, as stated in Table
1. In this table all nonselective experiments with packed beds
(5 out of 36) are summarized. In four of these experiments
interfacial convection has been absent (Eh 1), although Ras
and Mas had positive values. This nding accords well with the
stability character of the problem, and with the fact that threshold values of the driving forces must be exceeded before
convective ow can take place. It also indicates that the experiments with packed beds come closer to such criti