56 vues

Transféré par naraNJOR

AIChE_Journal__Vol[1]._51__No._12___December_2005_

- Immunogenicity of H5N1 h5n1
- PotentialmedicinalplantsforimprovetheimmunesystemfromBorneoIslandandtheprospecttobedevelopedasnanomedicine
- Job Interview
- Vietnam Today Nov Dec 11 VI
- 5-150303023636-conversion-gate01
- NOVAVAX INC 8-K (Events or Changes Between Quarterly Reports) 2009-02-20
- Bird Flu
- Infectious_Disease_Notes
- Research Report on China's Animal Vaccine Industry, 2009
- 22.essentialsofimmunology
- Antiviral Therapy
- Homeostasis and Immunity Concepts
- Infectious Diseases
- Bird flu
- US Army: APHIS Veterinary Services Memorandum 580 18 20060425
- MatrudevobhavabyDrRAnjaneyuluMDDGO
- 20111129022500_Part 1 MRCSI regulations and g.pdf
- IMUNOLOGIA Antígenos e Anticorpos
- Bob Beck Protocol
- Model Set Mcq

Vous êtes sur la page 1sur 229

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:

Fundamentals

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.

The Future of Chemical Engineering Research

November 8, 2004

AIChE Annual Meeting, AUSTIN, TX

SESSION II: BIOLOGICAL ENGINEERING

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)

2005 American Institute of Chemical Engineers

DOI 10.1002/aic.10728

Published online October 28, 2005 in Wiley InterScience (www.interscience.

wiley.com).

3082

December 2005

AIChE Journal

Perspectives

RESEARCH: BIOLOGICAL ENGINEERING

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

Chemical Engineering Research

Michael W. Deem

Rice University, 6100 Main Street-MS 142, Houston, TX 77005

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

3083

Perspective

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).

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

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.

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.

3084

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.

engineering

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.

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.

2003;100(6):3071-6.

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.

1997;15(6):553-7.

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.

1999;96(6):2591-5.

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.

2005;350(4):806-16.

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;

63(6):1288-96.

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

3085

Perspective

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).

Executive Summary

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.

Response of the Immune System to a

Perturbation

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.

3086

December 2005

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

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).

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

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).

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

3087

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

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.

Summary

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.

Acknowledgments

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

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

(2005a).

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),

q-bio.BM/0503030.

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

3089

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

(2004).

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://

www.u.lanl.gov/.

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

(2005).

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

3090

December 2005

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

(2002).

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).

AIChE Journal

Perspective

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).

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.

AIChE Journal

December 2005

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

3091

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

3092

December 2005

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.

Acknowledgment

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

4.

5.

6.

7.

8.

9.

10.

11.

12.

1991:252;1675-1681.

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;

381.

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

13.

14.

15.

16.

17.

18.

19.

20.

21.

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;

11:392-396.

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.

3093

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

Newtonian viscous uids

Introduction

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

delapla@lille.inra.fr.

3094

December 2005

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

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.

Experimental

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

m.

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

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

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

3095

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

tank.

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.

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-

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

below).

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.

Theory

Dimensional analysis for an agitated vessel equipped

with an impeller vertically and centrally mounted in the

tank

Symbols refer to various impeller gyrational speeds tested.

Solid line is obtained by linear regression.

3096

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

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 ,

NR D

,

0

NG ds

(5)

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

pi-set

F 6 N pG, ReG ,

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

(1)

dependency between dimensionless numbers8:

F 2N p, Re, T/d, H/d, HL/d, C/d 0

(2)

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

NR

0

NG

(6)

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

(7)

(3)

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

(4)

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

of the classical mixing system and the TRIAXE planetary mixer when rotational speed is

set at zero.

Vol. 51, No. 12

3097

u impeller tipt

2NG

2NR

D

2

2NG

ds

2

D 2 2

cos 2NR t 22 NR NG ds D sin2NR t

2

1/ 2

(10)

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.

(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

(8)

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/

(9)

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.

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.

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-

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)

3098

December 2005

using the three pi-numbers dened in Eq. 6.

Vol. 51, No. 12

AIChE Journal

revolutionary speeds.

Acknowledgments

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

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

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

impeller.

Conclusion

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

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.

2000;28:170-182.

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

dOM

dt

dOO

dt

dOO

dt

dOO

dt

OO

dOM

dt

with

R/R

Scheme 1.

Vol. 51, No. 12

3099

dOO

dt

OM

OO

and

dOM

dt

dOM

dt

R

OM

R/R

with

dOM

dt

V a

D/2cos2NR t

OM

D/2sin2NR t

0

0

D/2cos2NR t

0

0

2NG D/2sin2NR t

0

0

2NG D/2cos2NR t

0

0 2NG ds /2

2

0

G

0

ds /2

0

D/22NR cos2NR t

2NG D/2cos2NR t

frame (X , Y , Z ) is given by

D/22NR sin2NR t

D/22NR cos2NR t

0

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

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/

2)]2sin(2NRt)}.

Consequently dw(t)/dt 0 when the following conditions

apply:

t

1

4N R

3

1

,

4N R 4NR

3

1

,

4N R 4NR

or at

or

arcsin

3

t

4N R

NR ds

,

NG d

arcsin

NR ds

NG d

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

NR ds

sin2NR t

NG d

3100

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

reached at

December 2005

AIChE Journal

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).

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

Introduction

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

robert.rundqvist@fcc.chalmers.se

AIChE Journal

December 2005

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

3101

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.

Experimental

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

section, 2. Aluminium pipe, 3. Removable pipe

section, 4. Venturi tube, 5. Plexiglas section, 6.

Etoile straightener, 7. Bellow, 8. Fan.

3102

December 2005

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

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.

Numerical

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

d

K(U U ) if k p

M kIi

gi

pi

(3)

For a dispersed ow, the generalised drag force per unit volume in a suspension of particles is written

d

n pF idr

M kIi

(4)

np

p

Vp

(5)

a spherical particle can be written

dr

F sp

d p2

1 sp

C d gU rU r

2

4

(6)

between the phases. The drag function used here for spherical

particles is that given by 32

24

(1 0.15(Red )0.687 ) if Red 1000

C Red

0.44

if Red 1000

sp

d

(7)

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

(1)

V ki

d

U U k

P MkIi

U

t k k ki

x j k k ki kj

xj ij

k k gi Cki Vki

(2)

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

d

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

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

(8)

relation

kij k

1 U kl

U kl

2 k S kij

x l ij

3 x l ij

(9)

S kij

1 U ki U kj

2 xj

xi

(10)

suspensions, the bulk viscosity of phase k, k , is often set to

Vol. 51, No. 12

3103

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

U

) gug i pup i

Pij K (U

gi

pi

xj

g

p

U

) pup i gug i

Pij K (U

pi

gi

xj

p

g

U pupi 0

x i p pi

(11)

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 k

gg ug i ug j g g gi

xj g gij

U

U

p pU

pi

t

x j p p pi pj

p

kuki tk

(14)

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

equations are written as

k k

0

U

t

x i k k gi

(18)

2

U

gi

gj

ij k

xj

xi

3

(19)

1

u u

2 gi gi

(20)

the turbulent kinetic energy k, and the rate of dissipation,

using the following relation

tg C

k2

(21)

particulate phase, according to 18, can be written

up i up j tp

2

U

U

pi

pj

pl

ij kp tp

xj

xi

3

xl

(22)

kp

(15)

1

u u

2 pi pi

(23)

trace, that is to ensure that Eq. 23 is valid.

The additional term in the drag force in Eqs. 16 and 17 that

U

U

g gU

gi

t

x j g g gi gj

3104

k

xi

becomes

0

U

x i p pi

(17)

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

(13)

k uki can be modeled using a gradient diffusion model

k

kk

k

k

k

(16)

pp up i up j p p gi

xj p pij

(12)

averaged, and the Favre averaged can be obtained

December 2005

AIChE Journal

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

20:

p p

.

K

18

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

t

xj

xj

xj

g g

(C P C 2) e p

k 1 k

AIChE Journal

U

gi

xj

ugiupi 2k

(32)

(28)

kp k

(33)

TL

m TL

(34)

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

U

2

P kp fU

gi

pi

(35)

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

(29)

December 2005

TL

m TL

particulate phase k p , is modeled as

(27)

P k ug i ug j

K

TL

m

k T L

g

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

(26)

Turbulence model

(31)

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

(25)

tg ( g gk)

g gk

g gk

U gj

t

xj

xj k

xj

(30)

and the strain associated with the uctuating slip velocity. A

derivation by 35 shows that e p can be modeled

ep

T L 0.41

K

ugi upi

g

xj xj

(24)

m

TL

Comparing the drag term in the model by 18 with the drag term

in Eq. 2, the particle response time can be written

K

1

U

)

ugiugi ugiupi

u (U

pi

g

p p gi gi

ep

tp tg

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

K 2

U gi U pi

g

(36)

3105

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.

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

Calculations

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

x2

(37)

d p

6 2 p

(38)

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.

cross-section.

3106

December 2005

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

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.

of further investigation.

Conclusions

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

simulation of a horizontal ow.

Acknowledgments

The authors would like to thank the Swedish Foundation for Strategic

Research (SSF) and SCA Research AB for nancial support of this work.

Notation

C1

C2

C sp

d

C

d

C fib

d

C ki

C

D

df

dp

ep

F dr

i

dr

F

dr

F sp

f

K

k

kp

L

d

M kIi

m

P

Pk

P kp

AIChE Journal

December 2005

TL

t

U gi

in

U

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

3107

Ur

ugi

upi

V ki

Vp

xi

y

p

p

1

2

ij

tg

tp

k

g

p

d1

d2

k

kij

m

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

15.

16.

17.

18.

19.

20.

21.

22.

23.

24.

25.

k time-averaged variable for phase k

26.

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:

173186.

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.

1996;27:4159.

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.

1997;93:253260.

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

3108

December 2005

27.

28.

29.

30.

31.

32.

33.

34.

35.

36.

37.

38.

39.

Serizawa A, Fukano T, Bataille J, eds. Amsterdam: Elsevier Science;

1995.

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,

1995.

Ljus C, Johansson B, Almstedt AE. Turbulence modication by particles in a horizontal pipe ow. Int J Multiphase Flow. 2002;28:1075

1090.

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;

1975.

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.

AIChE Journal

SEPARATIONS

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

Introduction

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

gchase@uakron.edu.

AIChE Journal

December 2005

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

3109

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

reservoir.

Number of drops of size x in the inlet

size x measured in the outlet stream as

e x 1

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

tank.

Samples of the water-in-oil emulsion are taken at intervals of

10 to 25 min from the ow system at sample points upstream

3110

December 2005

(1)

n O x

n I x

(2)

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

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.

AIChE Journal

Separation Efciencies

Filter

Glass-Fiber

Mass (g)

Binder*

Mass

(g)

Nanober

Mass (g)

Total Filter

Sample

Mass (g)

A

B

C

D

E

0.500

0.500

0.500

0.500

0.500

0.325

0.334

0.361

0.359

0.374

0

0.007

0.014

0.020

0.039

0.825

0.841

0.875

0.879

0.913

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

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.

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

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

AIChE Journal

December 2005

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

drops.

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

3111

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.

Conclusions

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.

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.

3112

December 2005

added to the glass-ber lters.

Acknowledgments

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.

2001;187/188:469-481.

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:

299-319.

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.

2003;30:79-88.

14. Hazlett RN. Fibrous bed coalescence of water: Role of a sulfonate

surfactant in the coalescence process. Ind Eng Chem Fundam. 1969;

8:633-640.

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:

697-713.

17. Chase GG, Beniwal V, Venkataraman C. Measurement of uni-axial

ber angle in non-woven brous media. Chem Eng Sci. 2000;55:21512160.

AIChE Journal

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

Butterworths; 1990.

24. Brown RC. Air Filtration: An Integrated Approach to the Theory

and Applications of Fibrous Filters. New York, NY: Pergamon;

1993.

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.

3113

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

volatility, hydrocarbons, application

Introduction

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.

3114

December 2005

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

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

CAMD.

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.

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

CH2

I aromatic carbon ring attachment, such as ACH

H substituted aromatic carbon ring attachment, such as ACCl

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 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

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.

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.

As we know, relative volatility (which is consistent with

selectivity) is the most important among all the explicit propVol. 51, No. 12

3115

for Screening Solvents

Property

Relative volatility at

innite

dilution20,31

Selectivity at innite

dilution20,31

Solubility capacity20

Molecular weight

Normal boiling point

Critical properties

Vapor pressure

Azeotropic judgment

Method

ij

i Pi0

j Pj0

i

j

1 MWi

SP

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

method32,33

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

Sij

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.

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.

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,

COOH, OH, CH2CN, CH2NH2, ACH, ACCH3, ACNH2,

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

distillation.3

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

3116

December 2005

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

No.

Molecular Structure

MW

T b (K)

12

S

12

SP

1 and Solvent

2 and Solvent

1

2

3

4

5

6

7

8

9

10

DMSO

CH2CNOCH2CN

CH3COOOCH2OCH2CN

DMF

CH2NH2OCH2CN

NMP

CH2NH2OCH2OCOOH

ACN

CH3COOOCH2OCH2OCH2CN

OHOCH2OCH2CN

78.1

80.0

113.0

73.1

70.0

99.1

89.0

41.0

127.0

71.0

462.2

495.3

474.3

426.15

442.9

477.2

486.3

354.8

497.2

462.5

9.782

8.759

7.026

5.767

5.604

5.552

5.242

5.134

5.086

5.067

19.938

17.853

14.321

11.754

11.422

11.317

10.685

10.465

10.366

10.328

0.318

0.312

0.354

0.686

0.468

0.756

0.543

0.621

0.417

0.210

yes

yes

no

yes

yes

no

no

yes

no

yes

no

no

no

no

no

no

no

yes

no

no

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%.

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,

COOH, OH, CH2CN, CH2NH2, H2O, CH3OH, (CH2OH)2,

AIChE Journal

December 2005

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

3117

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

Azeotropic Judgment

No.

Molecular Structure

MW

Single solvent

80.0

1

CH2CNOCH2CN

2

ACN

41.0

3

CH3COOOCH2OCH2NH2

103.0

113.0

4

CH3COOOCH2OCH2CN

5

NMP

99.1

6

CH3COOCH2OCH2NH2

87.0

97.0

7

CH3COOCH2OCH2CN

8

CH3COOOCH2OCH2OCH2CN

127.0

9

CH3COOOCH2OCH2OCH2NH2

117.0

10

CH3COOCH2OCH2OCH2NH2

101.0

70.0

11

CH2NH2OCH2CN

12

CH3COOCH2OCH2OCOOCH3

130.0

13

CH3COOOCH2OCH2OCOOCH3

146.0

111.0

14

CH3COOCH2OCH2OCH2CN

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

15

H2O

18.0

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

16

H2O

18.0

Tb

(K)

12

S

12

SP

1 and

Solvent

2 and

Solvent

495.3

354.8

421.9

474.3

477.2

417.5

470.0

497.2

444.8

440.4

442.9

449.0

453.3

492.8

1.802

1.704 (1.620)

1.634

1.629

1.620

1.576

1.542

1.523

1.505

1.486

1.484

1.480

1.468

1.458

2.172

2.055

1.970

1.964

1.954

1.900

1.859

1.836

1.815

1.791

1.790

1.784

1.770

1.758

0.142

0.267

0.451

0.149

0.241

0.406

0.170

0.181

0.456

0.420

0.187

0.190

0.182

0.197

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

373.2

1.846 (1.750)

2.226

0.248

no

no

373.2

1.969

2.374

0.175

no

no

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.

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

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

No.

Molecular Structure

Tb

(K)

12

S

12

SP

1 and

Solvent

2 and

Solvent

495.3

354.8

421.9

474.3

477.2

417.5

470.0

497.2

444.8

440.4

442.9

449.0

453.3

1.780

1.684 (1.696)

1.614

1.609

1.601

1.557

1.524

1.505

1.487

1.468

1.466

1.462

1.451

2.172

2.055

1.970

1.964

1.954

1.901

1.859

1.836

1.815

1.791

1.790

1.784

1.770

0.111

0.238

0.477

0.122

0.225

0.414

0.143

0.157

0.491

0.437

0.161

0.168

0.161

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

373.2

495.3

462.5

399.2

352.8

390.4

337.8

443.5

351.5

1.824 (1.880)

1.692

1.649

1.640 (1.725)

1.631 (1.707)

1.626 (1.787)

1.610

1.607

1.579 (1.713)

2.226

2.065

2.012

2.002

1.990

1.984

1.965

1.961

1.927

0.209

0.223

0.219

0.245

0.251

0.238

0.242

0.225

0.250

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

no

373.2

1.945

2.374

0.139

no

no

MW

Single solvent

80.0

1

CH2CNOCH2CN

2

ACN

41.0

3

CH3COOOCH2OCH2NH2

103.0

113.0

4

CH3COOOCH2OCH2CN

5

NMP

99.1

6

CH3COOCH2OCH2NH2

87.0

97.0

7

CH3COOCH2OCH2CN

8

CH3COOOCH2OCH2OCH2CN

127.0

9

CH3COOOCH2OCH2OCH2NH2

117.0

10

CH3COOCH2OCH2OCH2NH2

101.0

70.0

11

CH2NH2OCH2CN

12

CH3COOCH2OCH2OCOOCH3

130.0

13

CH3COOOCH2OCH2OCOOCH3

146.0

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

18.0

14

H2O

15

CH2CNOCH2CN

80.0

16

OHOCH2OCH2CN

71.0

17

CH3OCH2OCH2OCH2OCH3COO

116.0

18

CH3COOCH2OCH3

72.0

19

CH2NH2OCH2NH2

60.0

32.0

20

CH3OH

21

OHOCH2OCH2NH2

61.0

46.0

22

CH3OCH2OOH

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

23

H2O

18.0

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

(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

2

V2 V3

RT ln S23 1p

(3)

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

2

1p

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

Groups from Eight Kinds of Different Groups (Excluding

Molecular Groups) for Single Solvents

Without any

constraint

After step 2

After step 3

AIChE Journal

Case 1

Case 2

Case 3

2994

112

6

2994

112

12

2994

112

11

December 2005

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

3119

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).

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.

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

1-butene system.

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

Conclusions

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.

3120

December 2005

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

reliable.

15.

16.

17.

18.

19.

20.

21.

22.

23.

Acknowledgments

This work is nancially supported by the National Nature Science

Foundation of China under Grant 20406001.

24.

Literature Cited

25.

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.

2003;32:121-213.

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;

589:396-414.

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:

S615-S618.

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;

22:747-763.

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

26.

27.

28.

29.

30.

31.

32.

33.

34.

35.

36.

37.

Chem Eng. 1999;23:1395-1414.

Sinha M, Achenie LEK, Ostrovsky GM. Environmentally benign

solvent design by global optimization. Comput Chem Eng. 1999;23:

1381-1394.

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:

1349-1360.

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.

1998;37:4876-4882.

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.

2002;41:1678-1688.

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;

26:729-742.

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.

3121

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).

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

Introduction

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

guiochon@utk.edu.

3122

December 2005

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

demonstrated.

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.

Theory

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

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

e

C

2C

C

u

eD L

t

x

x2

1 3 k

e,i

C C p,i r R i

Ri

ext

(1)

of the solid phase

p

q i

C p,i

1

C p,i

r2

1 p

D eff 2

t

t

r r

r

(2)

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

relationship

e,i 1 1 e

Vi

Va

(3)

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

zero.

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

1

Vol. 51, No. 12

L

Cttdt

0

Ctdt

u0

(4)

3123

Ctt 1 2dt

Ctdt

(5)

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

H HL Hk HD

2u 01 e P 1 P K 2 R

R2

2

3k ext 15D eff

0

(6)

1 e P 1 P K 2

02

eD L 02

u 02

(7)

H0

R2

f d

R

A

2

3k ext 15D eff

0

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

(8)

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

(9)

0 e 1 e P 1 P K

(10)

12

2

(11)

L L 2

2

N

1

(12)

2u 0 e

L f d

02

(15)

(16)

(17)

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

K

p

o e

1 e1 p 1 p

(18)

dispersion coefcient and the external mass transfer

coefcient

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.

(19)

the external mass transfer coefcient, to the Schmidt (Sc) and

the Reynolds (Re) numbers. These numbers are dened by

(13)

December 2005

eD L

2 L

2 H0

12 2u 0

u0

follows

(14)

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

H

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

parameters, A and H 0

2L

f d 2 inj 2 sys

u0 L

3124

2 eD L

u0

Re ud/

(20)

Sc / D m

(21)

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

e

d

2/3

(23)

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

A

(24)

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

2

e

eD L

B1 2 2

du 0

2

Re Sc

(25)

with

B

Re Sc2

Re Sc

1 p2

4 1 e

1614 1 e 2

2

1

p1 p3 exp

4 1 e

1

p1 pRe Sc

2

1

(28)

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

(29)

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).

Experimental

Equipment

(26)

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

(27)

stream percolating through the bed and the axial dispersion

coefcient is the following equation elaborated by Chung and

Wen16

AIChE Journal

(22)

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.

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

3125

Description

3 [m]

5 [m]

10 [m]

15 [m]

50 [m]

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

3.74

10%

1.57

103

383

17.35

3.48

0.603

0.362

0.378

5.21

6%

1.49

98

398

18.13

3.57

0.592

0.369

0.353

10.6

6%

1.59

101

386

17.07

3.03

0.61

0.371

0.38

14.92

7.5%

1.54

107

388

15.56

2.96

0.642

0.379

0.423

47.1

1.82

0.612

0.37

0.384

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

[m].

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.

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

3126

December 2005

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.

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

solute.

ISEC

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

Inuence of the particle-size distribution on the second

moments

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

polystyrene standards.

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

samples.

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

porosity.20,21

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

equation

2,d2 2,d1

2,d1

(30)

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

2,PSD

(31)

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

3127

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

[m].

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.

coefcient

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).

3128

December 2005

0.8

0.071

46.49

1

0.063

1.4

0.056

12.7

1

0.028

5.16

1

0.0098

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

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.

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

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

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

D L [cm2/s]

Ethylbenzene

d [m]

Exp

Eq. 25

5

3.74

4.28 10

5.21

6.30 105

10.6

8.13 105

14.92

1.34 104

47.1

1.17 103

Propylbenzene

b

Exp

5

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

5

4.23 10

5.96 105

8.88 105

2.02 104

1.27 103

Butylbenzene

b

Exp

5

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

5

3.53 10

7.15 105

9.29 105

1.46 104

1.28 103

Eq. (28)b

5

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

3129

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

diffusivity.6,13

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

H0 AR

3k ext

15

R2

D

1 pRD sK R

R

m p

(32)

Transfer-Coefcient and the Surface Diffusivity

(assumed)

0.66 (E)

0.26

0.66 (P)

0.162

0.66 (B)

0.103

Tortuosity factor from Eq. 35

0.66 (E)

0.202

0.66 (P)

0.139

0.66 (B)

0.1

1 (E)

0.086

1 (P)

0.044

1 (B)

0.096

1 (E)

0.14

1 (P)

0.136

1 (B)

0.127

D m pR

1 pRD sK R

(33)

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

(34)

p 1.51 p

(35)

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]

Experiment

Theoryb

3.74

5.21

10.6

14.92

47.1

0.2

0.162

0.101

0.0806

0.0377

0.070.094

0.0570.082

0.0510.039

0.0310.041

0.0140.019

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

3130

December 2005

D s [cm2/s]

Fisher

Test

6.34 107

4.926 107

5.04 107

1041

1281

1012

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

1197

1464

1076

1023

1002

498

3321

6678

1 104

The values of the external mass-transfer coefcient were calculated for a

particle diameter equal to 5.12 m.

inversely proportional to the particle dia.,12 and k ext D m /R p .

Taking all that into account, it was assumed that

contributions of the pore diffusivity and the surface diffusivity

D s , to the effective diffusivity are parallel. Thus,

D eff

k ext

[cm2/s]

k ext

R p

(36)

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

similar.

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

From this theory, the following relationship can be derived

k ext

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

1.

Dm

(37)

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.

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

diffusivity.

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

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

i/j

2/3

1.06

1.45

1.95

1.03

1.07

1.39

0.46

1.07

1.13

2.02

0.89

1.1

D mi /D mj

E/

P

P/

B

E/

B

Molecular diffusivities from Eq. 24.

Tortuosity parameter from Eq. 35.

b

c

3131

as a function of the particle radius.

mass resistance to band broadening increases with increasing

mobile phase velocity.

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

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

undertaken.

Conclusion

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

correlation.

Our experimental results suggest, by contrast, that the exterTable 7. Ratios of Effective Surface to Pore Diffusivities and

of Surface to Molecular Diffusivities

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].

3132

December 2005

Solute

D e,s /D p

D s /D m

Ethylbenzene

Propylbenzene

Butylbenzene

0.67

1.12

2.04

4.04 102

3.97 102

4.09 102

Tortuosity parameter was calculated from Eq. 35.

AIChE Journal

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.

1

2

viscosity

absolute moment

second central moment

uid density

Subscripts

a denote adsorbent

i i-th fraction of adsorbent particle

s solid phase

Superscript

R means that parameter depends on used particle radius

Acknowledgment

Literature Cited

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.

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.

Notation

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.

3133

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

control

Introduction

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

g.pannocchia@ing.unipi.it.

3134

December 2005

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

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

guarantees.

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

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 . . . .

In this article, we consider a discrete-time linear time-invariant plant

xk 1 Axk Buk Edk

(1a)

zk C zxk

(1b)

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 ,

(2)

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)

rank

I C A B0 n p

(3)

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

1

The results in this article can easily be extended to the case with mixed constraints

on the state and input.

3135

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

(4a)

k3

xk ,

uk ,

k .

(4b)

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

w

denotes the i th row of H, then

min Hw :

w

min H1 w min Hq wT

w

w

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

(5)

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).

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

r} if r 0.

3136

December 2005

d k 1 d k xk x k

Vol. 51, No. 12

(6a)

(6b)

AIChE Journal

subject to

I C A B0 ux 0I

z

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

(7a)

zk k

(7b)

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

2

(8a)

December 2005

I I

0

0 0 I s

(8b)

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,

u,

x, x,

x,s

u, u, u,s s

(9)

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)

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.

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

(10)

around the target calculator and therefore it is important to

verify that the closed-loop system (observer target calculaVol. 51, No. 12

3137

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

(11)

then

:

(12)

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

(13)

u s

(14)

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,

(15a)

zk k,

(15b)

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.

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

3138

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.

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

(16)

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

(17)

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

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.

15.

Proof. See the Appendix.

for which there exists a controller state such that the augmented

state is in , as

X 0 : x n x, d 2n such that

(18)

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.

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

(19)

nite horizon optimal control problem (dened later) is then a

nite sequence of input perturbations that guarantees robust

N , s : v mN

AIChE Journal

cost function.

Under the control in Eq. 19, the augmented state dynamics in

Eq. 7 become

k 1 k k dk sk

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

dened.

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

(20)

(21)

3139

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

(22)

x x*, s Qx x*, s u

N1

V N , s, v :

i0

u *, sT Rui u *, s xN x *, sT PxN x *, s

(23)

with the matrices Q nn , R mm and P nn

positive denite. The vectors x i n and u i m are dened

as

x 0 x I n 0

x i1 Ax i Bu i x x d ,

uk k sk *0 k, sk

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

(24a)

i 0, . . . , N 1

X Nv : x x, d 2n such that

N , s A for all s

(24b)

u i u * , s K x i x * , s i,

i 0, . . . , N 1

(24c)

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

(25)

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

3140

(26)

December 2005

(27)

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

satisfy

v

XNv

X 0 X 1v XN1

(28)

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

section.

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

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 ( ,

s)}.

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

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

expression for N ( , s) in terms of a nite number of afne

inequality constraints

N , s v mN Fv c H Hs s

(30)

c : b min Gd d min Gs s

(31)

where

d N

s N1

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

k3

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

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

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

(32)

N1

v*, s argmin

v

iT Wi Fv c H Hs s

i0

(33)

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

3141

computed from the following strictly convex QP

,v

x ,d

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

N1

x(0)) dT d)

T

i

i0

v

Figure 2. Domain of attraction (XN

) for different xed

horizons.

0.7776 0.0045 x k

xx kk 11 26.6185

1.8555 x k

0.0004

0.0002 0.0893 d k

0.2907 uk 0.1390 1.2267 d k

1

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

2

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

x

0.5

15 u 15,

0.5

5 x 5 ,

2

0.1

dd 0.12 ,

1

1 s 1

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

controllers.

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

3142

December 2005

horizon controllers (linear plant): controlled

variable (top) and input (bottom).

Vol. 51, No. 12

AIChE Journal

t (min)

d

s

[0, 4)

[2

0.1] T

0

[4, 8)

[2

0.1] T

0

[8, 12)

[2

[12, 16)

0.1] T

1

0.1] T

[2

1

[16, 24)

[2

0.1] T

1

[20, 24]

[2

0.1] T

1

(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.

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.

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

standard robust receding horizon controllers:

controlled variable (top) and input (bottom).

AIChE Journal

December 2005

Conclusions

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-

3143

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;

76:6172.

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.

AIChE Journal

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;

16011607.

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:

317367.

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;

1972.

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.

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

(A3)

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

(A4)

where the last step comes from Eq. 10. It is important to notice

that Eq. A4 and Assumption 4 imply that

x x

(A5)

s C zx C zx lim k

k3

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

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 ,

k3

(A1)

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

(A2a)

x Ax Bu x x d

(A2b)

d x x d

(A2c)

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

3145

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

v

v

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

k3

3146

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

that

lim k d s d s

k3

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.

(A6)

k3

December 2005

AIChE Journal

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

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

Introduction

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

ydstie@andrew.cmu.edu.

AIChE Journal

December 2005

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

3147

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

(2003).13,14,15

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

3148

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

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

n2

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

3NN!

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

1

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

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

right.

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

that

S 2 S 1 c Vln

U2

V2

R ln

U1

V1

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

3149

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

equivalent.

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

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

equation.

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.

S Z 1 1 Z 2 S Z 1 S1 Z 2

Using homogeneity gives

dS

S Z 1 1 Z 2 SZ 1 1 SZ 2

It follows that S is concave. The reverse argument follows

immediately.

We now introduce the set of conjugate variables

w

S

Z

(1)

P

1

1

n

dU dV

dM 1, . . . , dMn

T

T

T

T

SZ w TZ

MZ 0

(4)

where

M ij

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

3150

(3)

Using Eq. 3, and the chain rule we can now write (S/Z)

(w/Z) Z w T . It follows from Eq. 1 that

we can make the assignment

1 P

1

n

w

, , ,...,

T T

T

T

(2)

December 2005

2 S

0

Zi Zj

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

(5)

(Sandler, 1999)45

Eulers theorem asserts that if f( x) f( x) then we must have f( x) (f/ x) x.

AIChE Journal

shown in Figure 2. From Eq. 8 and the Euler identity (Eq. 3) we

have

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

P

1

V d

U

T

T

1, . . . , d n M

n

M

d

T

T

AIChE Journal

dv

p

dt

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

(7)

n

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

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,

200147)

(6)

developments. The specic (or molar) properties are dened so

that

n T

, M

1, . . . , M

Z Z/M V , U

(8)

December 2005

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

f

m

i1

the rate of production. We then get the following (very nice)

decomposition for the thermodynamic process

vt v0

pds

3

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.

3151

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

f

z

t

x

to the separation of thermodynamics, transport and reaction engineering into distinct

subelds of process systems engineering.

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

inventory i so that

v it

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

x, tdx

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

i1

and pt

(14)

where

fconv uz

(15)

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

Vt

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

(13)

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

3152

December 2005

U

V

z ,

M1

2

M

U

V

f conv m

,

M1

2

M

Q

0

fdiff ,

FD1

FD2

u

u

2

x

x

u

x

1

2

(16)

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

4

The corresponding conservation law is given by ( V / x) (m

V / x)

(u/ x) which is true but uninteresting.

AIChE Journal

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

(17)

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

We can analyze stability by studying the trajectories of the

deviation system derived from Eq. 14 dened so that

f

z

t

x

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

T

f diff

X wT

t

x

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

A

1 z1 2 z2

w1 w2 T

t

t

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

variables (Eq. 19) we get

f

A

wT

t

x

w 1 w 2 k 0w 1 w 2

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

t

x

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

X

w

x

(18)

Wt

f f 1 f 2

1 2

5

In the general case the right hand side of the energy balance consists of additional

n

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

diffusion.

December 2005

A x, tdx 0

(21)

Corollary C1 (Dissipation equality for process systems):

dW

f Tw 0L

dt

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

deviation variables

AIChE Journal

(19)

fT X wT dx

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

3153

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

and

Kw

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

write

T

f diff

X XT LX 0 XT X

wT wT Kw 0

and

then have

dW

f Tw 0L 0

dt

X TX dx

1992)52,53 we can write

dW

2

f Tw 0L 0 2

dt

L

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

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

(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

2

kL 0

L2

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

Proof: The Poincare inequality for homogeneous boundary

conditions gives

2

dWt

f Tw 0L 0 0 2 Wt

dt

L

X TX dx 0

2

L2

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

3154

December 2005

w Tw dx

0 0

2

kL

L2

process system with homogeneous boundary conditions is stable if

k L 0

2

L2

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

uyds

different initial conditions, and the same

boundary conditions converge to a single trajectory in the mean square sense.

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

A TP PA ddT

PB C 0

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 unforced system. More generally and assuming that W(t) is

differentiable we have the following denitions for nonlinear

systems.

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

dW

u Ty 22

dt

.

The system S is

It is quite easy to develop counter examples to control. For

example, the linear transfer function

Gs

s 3s 0.5

s 10s 0.5s 4

(22)

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

in Eq. 22 generated by MATLAB.

Note the Bode stability criterion does not apply due to the

presence of the unstable pole.

3155

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

22

T d

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

dWt

u Ty 0

dt

w t, x Tw t, xdx

(23)

f t, 0

u

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

theory.

One aspect concerning passive systems is that they are easy

to control. This can be observed by dening a proportional

control

u Ky,

Wt W0

Ky 2dt 0

w Tw dt

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

dt

and

dW C

eT u CeT e

dt

where u is dened in Figure 7. By adding the two storage

functions, dening W WS WC and setting u d u we

get

dW

e Td CeT e

dt

We now integrate over the interval [0,T] to give

K0

us that the storage function decreases according to

3156

system S, and a strictly passive controller C.

December 2005

6

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.

AIChE Journal

Wt W0

e Td CeT edt

eT edt W0

eT ddt

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).

Hence

W0

eT edt

eT edt

proposed by Georgakis (1986).4

dT ddt

is passive.

Proof: Let

e Tedt/

d Tddt

1

C

Farschman et al. (1998)5 used the macroscopic balance

dv

m, z, d pm, z, d

dt

(24)

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

dv*

dt

(25)

easily veried. By differentiating V we get

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

up

1

v v* Tv v*

2

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

dv*

Ce

dt

(26)

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

3157

0. The dissipation rate is given by C Kc.

Proof: First dene the variables

s e

K c 2 K c D 2

e

s

V

2T

2

where dot denotes time derivative. We now dene

u Kc e

1

s D e

I

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

passive.

Result R5: The PID control

m, d p m, d

3158

1

dv*

Kc e

dt

I

edt D

Kc

es Kc D ee

I

(27)

V

Kc

ss K c Dee

T

(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

(29)

quantities, and is a vector of unknown parameters.

Result R6: The parameter adaptive controller

u m, z, d, t T

de

dt

December 2005

Hence

eu K ce 2

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

1

es D ee

I

dv*

Ce

dt

d

gm, z, d, te

dt

Vol. 51, No. 12

AIChE Journal

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 *

eu e T t e T t ev *

This we re-arrange to give

T te eu e T t ev *

(30)

V

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

1

t T t

2g

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

and

limE U* 0

t30

V g 1 t T t

t30

M

n

V t T m, z, d, te

and E U K P

i1

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

dx

dv*

dt

m, z, t pm, z, t

y M M*

dv*

Kc v v*v v*

dt

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

so that

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

dv*

C i e,

m, z, t pm, z, t

dt

AIChE Journal

i 1, . . . , N.

December 2005

Proof:

3159

YWV

where W is dened by Eq. 21 and V by Eq. 25. We get, using

Corollary C1 and Result R4

dY

f Tw 0L

dt

Application to 1D TRPs

fTX w T dx

v v* T p

dv*

dt

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

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

and

eT edt c1

control to a TRP described by the equation

c

c

2c

v

D 2

t

x

x

with t [0, ) and x (0, L), Hulburt boundary conditions

c0, t c 0t

cL, t

0

x

and undergoing a rst-order reaction of the form A 3 B with

reaction rate

kc

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

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

i

x

u f0 f L

e M M*

x 0, L

t3

get

limT1 T* 0

limP1 P* 0

limi1 *i 0

t3

t3

t3

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

3160

dx

(31)

is strictly passive if

i 1, . . . , n

limwx* wx, t 0

December 2005

D2

k 0

L2

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*

dv*

,

dt

Kc 0

(32)

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*

1

Kc (M M*)

c0 cL

dx b*

proles for Example 1.

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

*dx

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

relationship

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

1

1

Kc (M M*)

c0 cL

I

(M M*)dt

dx

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

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

3161

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

control for Example 1.

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

model.

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

1

Kc (M M*)

v

dx b*

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)).

3162

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).

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.

Conclusions

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-

proles for Example 3.

Note that the concentration prole adjusts so that molar

holdup of component A is kept constant.

AIChE Journal

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.

Acknowledgments

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;

1990:9-23.

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.

1997;216:398-420.

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:

715-725.

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:

213-227.

23. Hanczyc EM, Palazoglu A. Sliding mode control of nonlinear distributed parameter chemical processes. Ind. & Eng. Chem Res. 1995;34:

557-566.

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,

2004.

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.

3163

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.

1998;21:121-132.

57. Hangos KM, Alonso AA, Perkins JD, Ydstie BE. Thermodynamic

approach to the structural stability of process plants. AIChE J. 1999;

45:802-816.

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;

1995.

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.

1970;46:344-366.

67. Wei J. Irreversible thermodynamics in engineering. Ind and Eng

Chem. 1966;58:55-60.

Thermodynamics and Symmetry of the Linear

Laws

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

in the book by Kondepudi and Prigogine (1995).63

The entropy density is dened so that for any nonempty

volume V

sdV

s s z

z

wT

t

z t

t

By using the transport Eq. 14 we get

s

f

wT

t

x

w Tf

w

s

fT

wT

t

x

x

Entropy Flux

(A1)

Entropy Generation

easily derived from Eq. A1. It states that the rate of entropy

generation is positive so that

S f TX w T 0

(A2)

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

T

T

S f conv

X fdiff

X wT 0

V

T

f conv

X u

M1

2

M

1/T

d

P/T

dx 1 /T

2 /T

By writing this expression out term by term and using the fact

we get

that u M

Xm

U

f conv

d 1

d P

d 1

1

V

M

dx T

dx T

dx T

2

M

Vol. 51, No. 12

d 2

dx T

AIChE Journal

T

X0

f conv

S f diffXT X wT 0

(A3)

w T

1 u 1 u

P

P

T x T x

T 2 T1 xu

nr

i1

The rst two terms represent reversible work and they cancel.

From Eq. A3 we, therefore, get

T 2 T1 xu 0

nr

S f diffXT X

i1

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

momentum.

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

that

1 u

S f diffXT X TR m 2

T x

1

nr

,...,

T

T

(A4)

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

(A5)

R Km

(A6)

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

1 u

T x

The afnity of reaction j is dened so that

v T

n

Aj

ij

j 1, . . . , r

(A9)

i1

(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

nr

nr

kj j

j1

kj j

ij

j1

(A7)

forces and intensive variables it is, therefore, necessary and

AIChE Journal

(A8)

i1

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

where

m

v s 0,

n

December 2005

7

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.

3165

kvv m

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

nr

T

j j j

(A10)

S dX TL

d S

wT K

dx X

w

dx

j1

T

kvv

nr

T

j j j

(A11)

j1

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

have

dLX

Kw

dx

(A12)

S

X TL LXT

X

hence

dX TL LXT

d S

dx X

dx

(A13)

S

w TK KwT

w

(A14)

Likewise

3166

December 2005

d S

S

w TKL 1 L LT

dx X

w

(A15)

S

d S

0

dx X

w

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.

AIChE Journal

Emulsion Polymerization

David Edouard, Nida Sheibat-Othman, and Hassan Hammouri

69622 Villeurbanne Cedex, France

DOI 10.1002/aic.10559

Published online September 13, 2005 in Wiley InterScience (www.interscience.wiley.com).

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

Introduction

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.

AIChE Journal

December 2005

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

3167

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

initiator.

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

3168

December 2005

constructed using this model.

Modeling

Population balance equations

The general population balance equation is given by the

following hyperbolic differential equation:

, t

( , t)

d

dt

generation , t

(1)

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):

t

r

n (r, t)

desorption 2

termination

entry

t

(2)

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

entry , desorption , and

termination are the total rates of entry, desorption, and termination within a single particle of size between r and r r,

respectively.

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

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

al.18)

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)

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)

t

r

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)]

t

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

particles.

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

(4)

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

nuct

Gr nuc

(5)

with

jcrit1

nucmic

em,i

[Micelle][IM i]V aq

iz

nuchom(t) k

aq

p,( jcrit1)

aq

(6)

(3)

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

(7)

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

studied.

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

IM E I

jcrit1

i1

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:

QI

1 dIV aq

aq k dI

V aq

dt

V

Figure 1. 0 1 model.

AIChE Journal

December 2005

(8)

3169

kd

I 2I

I M O

IM1

aq

kPI

aq

kpi

IMi M O

IMi1

Propagation

IMi T

Termination

ktaq

I T O

inactive polymer chain

ET

ktraq

IMi M O

kem,i

IMi Micelle O

new particle z i jcrit 1

Micellar nucleation

aq

kpj

crit 1

Homogeneous nucleation

new particle

IMj crit1 M O

IMi particlen O

particlen1

ke,i

keE

E particlen O

particlen1

1 dI V aq

aq

2k dI k pI

M aq k taqTI

V aq

dt

(9)

1 dIM iV aqt

aq

aq

k p,i1

M aqIM i1 k p,i

M aq k taqT

V aq

dt

k traqM aq k em,iMicelleIM i

1 dIM 1V aq

V aq

dt

aq

aq

k pI

M aqI k p,1

M aq k taqT k traqM aqIM 1

(10)

1 dIM iV aqt

aq

aq

k p,i1

M aqIM i1 k p,i

M aq k taqT

V aq

dt

k traqM aqIM i, i 2, z 1

1

V aq

Parameter

Description

kp MP MWm

4r2 dp NA

k des

3Dw Msat

aq

rs2 MPsat

k e,i

rs

4NaDw

k eE

k em,i

rs

3170

rmic

k e,inrdrIM i

i z, j crit 1

(12)

1 dEV aqt

1

aq

aq

V

dt

V

rnuc

(11)

1

V aq

k eErEnrdr k taqET

(13)

rnuc

4 r s NaD w

ow rate (mol/s), and [I] is the initiator concentration (mol/

dm3).

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.

phases

4NaDw

rnuc

G(r)

z i jcrit 1

z i jcrit 1

dm

dm MPsat MWm

1/3

December 2005

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

dt

(14)

RPt

4

m p(t) d pN A

3

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

rnuc

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.

(15)

dN St

Q St

dt

m P(t)

N m(t) N (t)

MW m

T

m

T

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

NA

(16)

reactor.

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

NSaq

adsorbed on the surface of particles and droplets as given by

the following thermodynamic relationships:

N SP

nr, t N Adr

(17)

rnuc

n t

S parb sS aq

a spN A1 b sS aq

(21)

S par 4 N A

nrr 2dr

(22)

rnuc

n r, tnr, tdr

rnuc

(18)

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:

sat

M aq

The concentration of micelles ([Micelle]) is then calculated

using the following expression:

nr, tdr

rnuc

dm

K daqMW m

where [M]sat

aq is the concentration of monomer in the aqueous

phase under saturation (mol/dm3):

sat

M sat

P M aq K paq

[Saq ] CMC

nagg

(23)

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

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

AIChE Journal

(20)

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

following:

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:

N TPt

(19)

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

3171

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.

(26)

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

t

r

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

t

t

deriving Eq. 24 with respect to time:

n r, t

(24)

Grn1p r, t

nr, t

nuc r rnuc

t

r

(28)

k desrn 1mr, t

r

n(r, t)

nr, t

(29)

(25)

growing, that is,

suggested by Crowley et al.,18 the following algebraic equation

is obtained from Eq. 3:

n 1mr, t

Grn 1mr, t

0

r

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:

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

form:

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

(30)

PSD of a 0 1 system is therefore the following:

(a)

n(r, t)

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

nuc (r rnuc )

t

r

(b)

n (r, t)

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

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

t

r

n(r, t)

kdes r

nr, t

Nr, t

Nr, t

t

t

2

n r, t

n r, t

r, t1 2n r, t

t

3172

nr, t

following equation can be derived:

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)]:

(27)

(31)

ktr MP n r, t

entm r, tn r, t 1

December 2005

AIChE Journal

for the 0 1 system

Numerical solution

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.

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

literature.

1700 s (without desorption).

nr nuc, t

nuct

Gr nuc

and

n rnuc , t 1

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

3173

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

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

3174

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)

dt

2

d (t)

(N N m)

dt

k p[M] P m

(32)

1700 s (with desorption).

Vol. 51, No. 12

AIChE Journal

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 .

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

AIChE Journal

December 2005

3175

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

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:

3176

December 2005

AIChE Journal

Second, uncertainties are considered in the model of

n (r, t).

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)

xt

x k1(t)

N pT(t)

x k2(t)

nuc(t)

x k3(t)

(t)

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

d

dt

dt

rnuc

nuct Gr n r , tnr , t

(33)

(t), and A(r) is given by

ak ak1

0

0

ak1 ak2

0

0

a1

Ar 0

0

0

0

0

1

0

1

0

0

0

0

0

0

1

0

particles size does not exceed it, then the following assumption can be made:

Gr n r , tnr , t 0

observability, which allows us to construct a high-gain observer (see, for example, Farza et al.35) as given by the following system:

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

nuct

dt

(34)

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:

y(t) Cx(t) [0 0100] x(t)

(t) N A

N TP(t)

n (t)

(35)

AIChE Journal

December 2005

a i Gr nuc i rn r nuc i r, t/ r

(t) N A

y(t)

n (t)

(36)

(37)

where

3

53

5

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

3177

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

3178

December 2005

AIChE Journal

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

0.2

0.2

AIChE Journal

December 2005

3179

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.

3180

December 2005

AIChE Journal

into account any uncertainties in the model of n (r, t).

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,

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:

n r, t

r, t1 2n r, t

t

r

n(r, t)

Ar,t

k des(r)

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)

(38)

Br, t

concentration of radicals in the aqueous phase (IM3 and IM4)

and the concentration of micelles ([Micelle]). To calculate

AIChE Journal

December 2005

should be accounted for. These terms thus imply an important

calculation time. However, a rapid numerical analysis of these

Vol. 51, No. 12

3181

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

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:

3182

December 2005

AIChE Journal

t

r

n(r, t)

(39)

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.

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

AIChE Journal

December 2005

3183

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.

Conclusions

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

particles).

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.

Notation

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

3184

December 2005

dp polymer density (1.044), kg/dm3

Dw diffusion coefcient for monomer in water (1.5 107),

dm2/s

[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

aq

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

ical (280), dm3 mol1 s1

aq

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),

t

dm3 mol1 s1

ktr coefcient of transfer to monomer in the polymer particles

(9.3 103), dm3 mol1 s1

aq

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

[M]sat

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

AIChE Journal

rnuc

rd

rs

RP

nuc

nuc hom

nucmic

[Saq]

Spar

[T]

Vaq

Vd

z

(r, t)

entr(r, t)

entm(r, t)

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

14.

15.

16.

17.

18.

19.

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:

1878-1896.

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

13.

December 2005

20.

21.

22.

23.

24.

25.

26.

27.

28.

29.

30.

31.

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;

2000.

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;

58:3681-3698.

Semino D, Ray WH. Control of systems described by population

balance equationsI. Controllability analysis. Chem Eng Sci. 1995;

50:1805-1824.

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;

37:875.

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.

3185

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,

2005

packed bed, rotating trickle-bed reactor

Introduction

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

dprao@iitk.ac.in.

3186

December 2005

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

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.

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

kinetics.8

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

Experimental

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.

Vol. 51, No. 12

3187

Mean diameter, mm

Packing density

External porosity

Internal porosity

Particle density

Pore volume

External surface area

2.2, 1.5

797.6 kg/m3

38.18%

62.80%

1.17 g/cm3

0.5366 cm3/g

1690, 2489 m2/m3

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.

Setup

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

Property

RECEMAT (2733)

RECEMAT (1723)

Specic surface area, m2/m3

Material

Dimensions of the rings

ID, mm

OD, mm

Density, g/cm3

Porosity

0.6

2500

Nickelchromium

0.9

1700

Nickelchromium

102

149

0.60

0.9

102

149

0.65

0.9

Metal foams

3188

December 2005

AIChE Journal

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.

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

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,

The reaction rate was determined from the rate of change of

the concentration of cumene in the charge with time using the

equation

AIChE Journal

December 2005

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.

3189

rpm).

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.

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

3190

December 2005

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

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.

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

possible.

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.

Conclusions

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

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

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

step.

Acknowledgments

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

3191

10.

11.

12.

13.

14.

15.

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:

424-426.

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

3192

December 2005

16.

17.

18.

19.

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;

1995.

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.

AIChE Journal

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).

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

Karman vortex

Introduction

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

vlachos@che.udel.edu.

AIChE Journal

December 2005

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

3193

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.

Microdevices

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

Re

dU

(1)

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

St

fd

U

(2)

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

3194

December 2005

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.

interscience.wiley.com.]

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

ows.

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.

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

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

www.interscience.wiley.com.]

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

performed.

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)

processor.

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

3195

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

(3)

a rate given by

kmol m2 s1 148CA

(4)

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

(5)

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

(6)

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.

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

3196

December 2005

(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

www.interscience.wiley.com.]

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

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).

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.

wiley.com.]

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

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

www.interscience.wiley.com.]

3197

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

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.

wiley.com.]

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

introduced

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

3198

December 2005

y i 0.5VdS

VdS

(7)

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.]

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.

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

3199

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

3200

December 2005

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

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.

interscience.wiley.com.]

AIChE Journal

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

communication.

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).

Improvements

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

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 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

Pe

UL

D

(8)

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

3201

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.

Conclusions

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.

3202

December 2005

Acknowledgments

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;

285:83-85.

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;

9:86-92.

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.

2004;4:278-286.

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:

829-834.

17. Norton DG, Wetzel ED, Vlachos DG. Fabrication of single-channel

AIChE Journal

18.

19.

20.

21.

22.

23.

24.

25.

26.

27.

28.

29.

30.

31.

32.

33.

34.

35.

36.

37.

38.

39.

40.

41.

42.

43.

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.

2003;58:3961-3971.

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:

L483-L491.

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:

113-133.

Li J, Chambarel A, Donneaud M, Martin R. Numerical study of

laminar ow past one and two circular cylinders. Comput Fluids.

1991;19:155-170.

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:

619-640.

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:

166-173.

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

44.

45.

46.

47.

48.

49.

50.

51.

52.

53.

54.

55.

56.

57.

58.

59.

60.

61.

62.

63.

64.

65.

66.

67.

68.

69.

70.

71.

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.

1995;123:309-316.

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:

151-164.

Thomas DG, Kraus KA. Interaction of vortex streets. J Appl Phys.

1964;35:3458-3459.

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.

2004;362:987-1000.

Yang Z, Matsumoto S, Goto H, Matsumoto M, Maeda R. Ultrasonic

micromixer for microuidic systems. Sens Actuators A Phys. 2001;93:

266-272.

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.

2004;101:403-407.

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.

3203

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.

1972;4:313.

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-

3204

December 2005

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.]

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.

AIChE Journal

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

autoclave reactor

Introduction

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

ray@engr.wisc.edu.

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

3205

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

Homopolymerization of Ethylene

Modeling the LDPE Autoclave Reactor*

Initiation:

feff kd

2R decomposition of initiator)

IO

fast

P1

R MO

Chain Propagation:

kp

Pn1

Pn MO

Termination by combination:

Quantity

Value

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

1

2000 atm

21386 cal/mol

30200 cal/mol

*Note: Values for physical constants are taken from Read et al.22

ktc

Dnm

Pn Pm O

Ethylene decomposition:*

Cmon1.89kmd1 kpd1 0.0714kpd2

M O

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:

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.

the LDPE Autoclave Reactor

2

R heat,decomp HdecompCmon

1.89kmd kpd1 0.0714kpd2Cmon

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

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

Reaction

Units

k0

E A (cal/mol)

V A (cal/atm-mol)

Source

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 )

L/(mol-s)

s1

L/(mol-s)

L/(mol-s)

L/(mol-s)

s1

1.14 107

1.06 1016

3.00 109

4.00 1019

1.59 1020

4.39 1020

7091

35560

2400

65000

65000

65000

0.477

0.0605

0.3147

0.1937

0.3218

0.1937

30

4

30

27

27

31

*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

3207

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.

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

reactors.12

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

later.

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.

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

3208

December 2005

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.

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.

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

compartment.

The feedstream is introduced at 300 K with 50 ppm of

the TBPOA initiator.

3209

Figure 6. Procedure for selecting and using a compartment model of the LDPE autoclave reactor.

on reactor performance may be found elsewhere.28

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.

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

3210

December 2005

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.

efciency

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

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.

can be dened as follows

V HR

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.

nzones

f rV

p,i

Rp,max

i1

(2)

n zones

i1

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

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

(3)

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

considered as an inverse measure of initiator efciency.

December 2005

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.

3211

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.

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

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

decomp.39

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.

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.

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-

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.

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

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.

3213

les, which show only a very small volume being used for

reaction at the highest residence time (328 s).

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

region.

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

3214

December 2005

The compartment volumes are dened as: V i V 1 i1 ,

where the growth factor is xed for this work at 1.2.

Ncompt

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

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

reactors.

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

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

Advantages:

Very simple and fast computation, including stability

analysis

No adjustable ow/volume parameters

Does not require knowledge of the ow eld

Disadvantages:

Does not offer stabilization from poor initiator usage or

variable effective reaction volume

Very limited stability region

Model IIThree-compartment model technique:

Advantages:

Simple and fast to compute, including stability analysis

Includes the stabilizing effect of poor initiator mixing,

greatly expands stability region compared to perfect

mixing

Only requires rough knowledge of ow eld

Disadvantages:

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

Model IIICFD:

Advantages:

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

Disadvantages:

Time consuming to set up and compute; difcult to use in

stability analysis

Requires detailed knowledge of reactor geometry

Model IVCombined CFD/Compartment model:

Advantages:

Retains spatial detail of CFD solution

Accurately represents stability region

Much faster than CFD, allowing stability analysis

No free parameters for ows and volumes

Disadvantages:

Requires at least one CFD simulation or detailed

knowledge of ow/reactive eld

Model VNew plume mixing model:

Advantages:

Can represent stability region fairly well

Simple to compute, allowing stability analysis

Does not require CFD or detailed knowledge of ow/

reactive eld

Disadvantages:

Loses some knowledge of spatial details

Requires specication of free parameters for

ows/volumes

3215

region.

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.

Acknowledgments

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.

Notation

CFD

Cmon

Cp

Dn

EA

feff

fji

frV

fvol,i

H

k0

kd

kmd

kp

kpd1

kpd2

ktc

LDPE

M

MWI

MWmon

Ncompt

nzones

P

Pn

Qn

Rgas

R

R

3216

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

RPM

Rheat,decomp

Rp

T

Tk

Tf

TBPOA

VA

Vtot

Vn

wI,f

wIk

wkp

xp

I

Hpoly

Hdecomp

0

k0

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,

mol/L

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,

1989:273282.

6. Gemassmer AM. Autoclave Process for the High Pressure Polymerization of Ethylene. Erdol und Kohle-ErdgasPetrochemie. 1978;31:

221228.

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

408.

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.

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;

39:111126.

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.

1997;A34:10171043.

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;

1995.

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.

1996;50(2):215243.

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;

1992.

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

1778.

41. Versteeg HK, Malalasekera W. An Introduction to Computational

Fluid Dynamics: The Finite Volume Approach. London: Longman

Group; 1995.

AIChE Journal

December 2005

Appendix

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

1

dt

Vk

d 0k

1

dt

Vk

k

n

n

I

(A1)

n0

f Q 2k f C k

n

k

n

n

0

d eff

tc

(A2)

n0

f Q w k C

Ncompt

k

n

n

p

0 MWmon

mon

n0

f Q T H

Ncompt

k

n

Ncompt

dw pk

1

dt

Vk

1

dT k

dt

Vk

f Q w k C MW

Ncompt

k Cmon0

Cp

poly p

n0

(A3)

2

Hdecomp[Cmon

(1.89kmd kpd1 ) 0.0714kpd2 Cmon]

Cp

(A4)

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

(A5)

the recycle ratio (R), and are dened as follows

f 01 1

Vol. 51, No. 12

3217

f 12 1

V1

f 23 1

V tot

N compt1

i0

R

f

31 R

1

3

f 32

V i V i1; 1 i N compt

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

2R

31 R

f ii 1, i 1 . . . 3

f ij 0, all other cases

(A6)

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

3218

(A7)

December 2005

f ii 1, i 1 . . . Ncompt

i

f Ncompt

RVi

; 1 i Ncompt

R 1Vtot VNcompt

(A8)

Manuscript received Nov. 24, 2004, and revision received Mar. 10, 2005.

AIChE Journal

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).

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

matching

Introduction

Dynamic models for heterogeneous reactors are composed

of distributed parameter, ordinary, and partial differential equa-

aticas Aplicadas y Computacion, Instituto Mexicano del Petroleo.

lvarez-Ramrez

Correspondence concerning this article should be addressed to J. A

at jjar@xanum.uam.mx.

AIChE Journal

December 2005

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

3219

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

available.

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

3220

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.

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

(1)

nondynamic reactive concentration, k is the reaction constant,

and

2

2

2

x 12 x 22 x 32

simply connected. The boundary conditions for Eq. 1 are given

by

c x c b

c x is finite

for all x D

for all x D

(2)

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:

def

D kc dV

D kc bdV

(3)

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

kL

D

def

(4)

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,

tanh

2I1

I0

1

1

2

3

tanh

for cylindrical geometry

(5)

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.

Let

def

def

R b

kc bdV

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.

conditions

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

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)

conditions.

(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

Rs

kC x, sdV

(7)

problem (notice that sc is the Laplace transform of c/t):

DC kC sC

kc xdV

(6)

(8)

C x, s C bs

for all x D

Cx, s is finite

for all x D

and

AIChE Journal

December 2005

(9)

3221

follows:

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)

s : R bs 3 Rs

(10)

relationship

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.

sn

(14)

normalizing the time variable as td. In fact, td 1

can be

d

interpreted as a characteristic diffusion timescale. For (s) k

s the dynamic Thiele modulus can be written as

s n 2 s n

(15)

expressions for common geometries are the following:

s n,

tanhsn

sn

2I1 sn

sn I0 sn

1

1

3

sn tanhsn sn 2

for cylindrical geometry

(16)

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

(11)

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

modulus

def

s L 2

D

(12)

an expression for the DEF is obtained by a formal substitution

x (). That is, the DEF is obtained as

s s

(13)

(0) to nondynamic conditions.

3222

s

d

December 2005

explicit. In this way, (sn, ) can be seen as a family of linear

operators parameterized by the classical Thiele modulus .

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

Thiele modulus .

Observe the high frequency convergence of the Fickian diffusion behavior (n, ) 1/in.

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

few slow modes that can capture, to some degree, the dominant

asymptotic spatiotemporal behavior of the overall reaction rate

R(s).

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

mechanism.

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

conditions of Figure 1.

Notice that (n, ) is a decreasing function of the frequency n.

3223

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

practice.

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

(17)

where

def

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

(18)

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

2

s n,

(19)

2

because the equation is transcendental, the part e sn

2sn

e

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.

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,

roots of the characteristic equation

Es n, 2 s n e sn e sn

2

3224

December 2005

ns n 1

time constant. Given that R(sn) h(sn, )Cb(sn) and h(0, )

Vol. 51, No. 12

AIChE Journal

retain the steady-state characteristic of the exact dynamics

R(sn) g(sn, )Cb(sn), the equality

kV

(20)

Imn ,

kV

ns n 1

(21)

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,

n,

ns n 1

(22)

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 ,

(24)

def

f n,

ni n 1

This function can be written in terms of its real and imaginary parts as follows:

f n,

n n

i

n 2 n2 1

n 2 n2 1

(27)

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

Im0,

(28)

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:

(25)

(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

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:

(23)

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:

(26)

and

hs n,

n 2 n2 1

Ren ,

December 2005

Im0,

Im, Im0,

102 sufces to obtain a stable nite-difference approxiVol. 51, No. 12

3225

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

(30)

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

following:

(1) Same step as in the rst-order model case.

(2) In this case,

f n,

1 n 2 n2 i2 n n

(31)

yields

f n,

Figure 3. Derivatives Re[(0, )] and Im[(0, )] of the

DEF (n, ) around steady-state condition n

0.

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

i

n,

n ,

(32)

(3) The model-matching problems are obtained as follows:

Ren ,

1 n 2 n2

n ,

(33)

and

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,

3226

Vk

2 2

n s n 2 n s n 1

(29)

December 2005

model from a model matching problem around

steady-state condition.

Vol. 51, No. 12

AIChE Journal

the reaction diffusion process does not exist.

{[ ()]/[n()sn 1]} dynamic EF.

Imn ,

2n n

n ,

(34)

steady-state condition n 0 is computed. In Figure 3 one can

observe that Re[(0, )] 0, which yields the approximation

1

Re0, n2 2n 2 n2

2

(35)

following time-constant estimate is obtained:

1

Re0,

4

(36)

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

c

Dc fc

t

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

3227

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

dt

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:

dc

t fc

dt

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

system.

It is interesting to note that when the method described

above is used in the linear case, the approximate dynamics are

given by

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.

Simulations

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

dc

3c bt c 2c

dt

c 2c

2c

t n x 2

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

3228

December 2005

(37)

c x, t n c bt n

cx, tn

0

x

for x 1

(38)

for x 0

Eq. 7)

Rt n Q 2

c x, t ndx

(39)

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

(40)

Thiele modulus has to be specied. Because both the exact and

Vol. 51, No. 12

AIChE Journal

the approximate solutions for two different

values of the Thiele modulus.

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

1.0 and n 0.75.

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

3229

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.

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.

Conclusions

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

effectiveness.

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

3230

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.

1997;216:398-420.

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.

AIChE Journal

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

Introduction

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

bala@uh.edu.

AIChE Journal

December 2005

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

3231

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

length.

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.

acidization experiments.

observed

in

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.

AIChE Journal

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

17

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

expensive.

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

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

uids.32

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

3233

The Darcy scale model equations are given by

1

K P

(1)

U0

t

(2)

C f

UCf De Cf kc a Cf Cs (3)

t

k cC f C s RC s

RC sa

t

s

(4)

Cs

(5)

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

3234

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

Cf

1

ks

kc

(6)

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.

(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

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

by

Sh

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

K o o o1

rp

ro

K o

K o

(7)

(8)

and

a r o

a o or p

(9)

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

dh

Sh 0.35

Dm

x

0.5

Rep1/ 2 Sc1/3

(10)

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

(11)

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

2

above, the two contributions are equal when 7 Re1/

3 or

p

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

3235

reaction.

(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

Udh

Dm

(12)

pore hydraulic diameter. For a well-connected pore network,

random walk models and analogy with packed beds may be

used to show that

DX

D eX

os XPe p

Dm

(13)

DT

D eT

os TPe p

Dm

(14)

and

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

1

D eX

3Pep

os Pe pln

Dm

6

2

Regime

December 2005

7 104

7 103

0.7

1.4 10

1.4 103

0.14

Cases

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

x

x

,

L

rp

,

ro

a

,

ao

y

,

L

z

,

L

cf

2k sr o

,

Dm

Da

U

,

uo

t

,

L/uo

K

,

Ko

(16)

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

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.)

r

D eX

os Pe p2.

Dm

Face

Wormhole

Uniform

(15)

and

3236

Injection Rates

k sa oL

,

uo

Cf

,

Co

N ac

Pe L

cs

Cs

,

Co

P Pe

u oL

Ko

Co

,

s

u oL

,

Dm

2r o

,

L

H

L

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

p

p

,

x

y

u,

(17)

0

t x y

DaA cf

c f uc f c f

t

x

y

x

2 r

1

Sh

X ur

cf

x

y

os Da

2

os Da

cf

T ur

2

y

DaN acA c f

t

2r

1

Sh

(18)

(19)

p

1@x0

x

p0@x1

(22)

p

0 @ y 0, 0

y

(23)

cf 1 @ x 0

(24)

c f

0@x1

x

AIChE Journal

(21)

(25)

December 2005

and y o

(26)

cf 0 @ t 0

(27)

x, y, a, l 0 f a,l @ t 0

(28)

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

by

(20)

equations are given below

c f

0@y0

y

PV FaceD

s1 o 1 o

C o o

N ac o

(29)

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

2r

s 1

Sh

(30)

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

3237

t bth

s

k sC o

2r

Sh

(31)

high injection rates is given by

PV UniformD

t bthu o

su o

oL

k sC o a o oL

2r

Sh

1

DaN ac o

2r

Sh

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.

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

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

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-

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.

3239

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

direction is given by the dispersion-reaction equation

D eT

The length scale l T in the transverse direction is given by

lT

(32)

lT

0.1 to 1 f Wormhole

lX

(33)

lT

O1 f uniform dissolution

lX

(34)

and

C f

keff Cf

x

(35)

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

(37)

l T k effDeT

lX

utip

(38)

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

aD

k s k c eT

u tip

(39)

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

convection-reaction equation

u tip

lT

O1 f Face dissolution

lX

2C f

k effCf

y 2

u opt

k ck s

a D k effD eT

k s k c eT

(40)

dispersion coefcient. The parameter in Eq. 39 can be

written in terms of dimensionless parameters Damkohler number Da and Peclet number Pe L as

1

1

1

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

Da

2r

1

Sh

A D T 1/ 2M

o A DT 1/ 2 M

(41)

(42)

where

u tip

lX

keff

3240

(36)

December 2005

M u o/u tip

Vol. 51, No. 12

AIChE Journal

reduces to

lT 1 DT

L A

breakthrough versus the parameter 1

for

o

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.

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

o

Da Da

Pe L

k sa oD m

u o2

(43)

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

lT

L

AIChE Journal

D eT

k effL2

r

Sh

2

DT

A

1/ 2

(44)

December 2005

1/ 2

(45)

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

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.]

3241

face dissolution.

(c) Reaction Regime. The magnitude of 2 r/Sh or k s /k c in

the denominator of the local equation

Cs

as function of the reciprocal of Damkohler

number.

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.

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-

Cf

ks

1

kc

Cf

2r

1

Sh

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

model.

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

Acid

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

3242

2 [r o 1 m20 m]

2 /Sh

C s /C f

5.3 105

0.00170.034

0.00060.0113

0.990.98

4 106

4.8 105

0.00240.048

0.00080.016

0.990.98

0.00460.092

0.00150.0306

0.990.97

0.010.2

1.1122.2

0.0030.06

0.377.4

0.990.94

0.730.135

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

AIChE Journal

om P T/Pe L

a oD m2

2u o2r o

(49)

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

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.

PT ShA

c f uc f c f

cf

t

x

y

r

x

x ur

os PT

m2

os PT

cf

y ur

2

y

m

cf

x

y

Da

A D 1/ 2M

Pe L T

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

where

PT

a oLD m

2u or o

(46)

controlled reactions dened as the ratio of convection time to

diffusion time and

m2 P TPe L

a oL 2

2r o

(47)

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

P T ShA D T

Pe L

r

1/ 2

M om

ShA D T

r

1/ 2

(48)

The x-coordinate is independent of reaction rate (parameters: N ac 0.1, o 0.2, f [0.15, 0.15]).

where

AIChE Journal

varied to simulate breakthrough curves of kinetic (2 0.001, 0.07) and mass transfer

(2 10, 100) controlled reactions.

December 2005

3243

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

u

c f

DaA cf

c f

2

x

y

x

r

1

Sh

os Da

cf

x ur

2

os Da

cf

y ur

2

(50)

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).

and

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-

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

2

and 2 for kinetically controlled reactions and om and m

for

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

3244

December 2005

DaA c f

2r

1

Sh

3

o

K

exp b

Ko

o

1

(51)

(52)

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

The initial value of porosity o is equal to 0.36.

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

for low values of b and steep for large values

of b.

AIChE Journal

December 2005

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

3245

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

salt

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 ,

2

m

, 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

2

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

3246

December 2005

experimental data for different structureproperty relations.

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

different structure-property relations.

Vol. 51, No. 12

AIChE Journal

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.

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

pursued in future work.

Acknowledgments

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.;

1992.

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;

2229.

14. Buijse MA. Understanding wormholing mechanisms can improve acid

treatments in carbonate formations. SPE Prod. & Facilities. Aug. 2000;

168.

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

3247

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:

213.

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.

3248

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.

AIChE Journal

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).

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

Introduction

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

rakesh.gupta@mail.wvu.edu.

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

3249

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.

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

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.

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

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

3251

2c

c

D 2

t

x

(1)

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

Mt

1

M

8

D(2n 1)2 2 t

exp

(2n 1) 2 2

4l2

(2)

the sample thickness. At the initial stages of diffusion, the

solution for Ficks law at small times reduces to

Dt

Mt

4

M

(2l) 2

1/ 2

(3)

initial slope of Mt/M vs. t1/2/2l using short-time water-uptake

data.11

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

3252

December 2005

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

1

c

nx

1

t

x

(4)

now let the ux nx equal (D/)(c/x), Eq. 4 becomes

c D1 2c

t

1 x 2

(5)

Vol. 51, No. 12

AIChE Journal

D1

1

(6)

D s/D t 1

(7)

Dt

Finally, then

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

molecules.

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

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.

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

3253

Equilibrium Moisture Content of Vinyl Ester Resin Samples

Diffusion

Coefcient

106 (mm2/s)

Equilibrium

Moisture

Content (wt %)

System

4C

25C

4C

25C

Neat resin

1 wt % Cloisite 10A

2 wt % Cloisite 10A

5 wt % Cloisite 10A

0.14

0.11

0.11

0.07

0.74

0.40

0.28

0.06

0.60

0.85

0.90

0.97

0.66

1.13

1.34

2.08

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,

3254

December 2005

samples.

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

Samples

Diffusion

Coefcient

106 (mm2/s)

Equilibrium

Moisture

Content (wt %)

System

4C

28C

4C

28C

Neat resin

1 wt % Cloisite 10A

2 wt % Cloisite 10A

5 wt % Cloisite 10A

0.12

0.12

0.08

0.08

0.66

0.43

0.37

0.26

0.58

0.66

0.59

0.61

0.68

0.88

0.95

1.09

AIChE Journal

calculated from the measured steady-state mass ux as

Ds

resin at 25C.

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

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

Sample area

Concentration

(8)

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

ester.

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

molecules.

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

System

(mm2/s)

Neat resin

1 wt % Cloisite 10A

2 wt % Cloisite 10A

5 wt % Cloisite 10A

0.75

0.70

0.67

0.64

3255

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.

5.

6.

7.

Conclusions

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.

8.

9.

10.

11.

12.

13.

Acknowledgments

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.

14.

15.

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

3256

December 2005

16.

17.

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;

2003.

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.

2002;51:319-324.

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:

3332-3336.

Manuscript received Apr. 14, 2004, and revision received Apr. 4, 2005.

AIChE Journal

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).

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

Introduction

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

evangelos.tsotsas@vst.uni-magdeburg.de.

AIChE Journal

December 2005

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

3257

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

3258

December 2005

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

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.

Figure 1. Apparatus for the investigation of the evaporation of a binary mixture from planar gaps into

inert gas.

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)

numbers.

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

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

geometrical parameters of the gaps.

Vol. 51, No. 12

3259

Figure 3. Apparatus for the investigation of the evaporation of a binary mixture from packed beds

into inert gas.

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

obstacle.

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

3260

December 2005

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

Nl

N l,n1

N l,n N l

(1)

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

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

1

150 1 2

(2)

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.

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

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.

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

3261

as in Figure 4, although at x1,B,0 0.30.

l,0 D l/G D

(3)

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

3262

December 2005

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.

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

Ni

ri

N

(8)

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.

Ki

y i,Ph i x i,Ph p i0

x i,Ph

p

(9)

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

Ni

dN1

dt

d

N x

dt l i,B

(4)

(5)

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

(6)

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

0).

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

(7)

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

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

3263

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

2

N gh g,in h g,out

i g,i

Ph

g0

Q

(10)

i1

and

N h T

2

l

Q

v,i

g

c p,l,iT Ph T l,B Q

Ph

(11)

i1

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

g gK A,gA PhT g,B T Ph

Q

(12)

Q

(13)

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

unity.41

The gas-side heat transfer coefcient g is calculated according to the Lewis analogy

g

n gc p,gLein

g,i

(14)

g

ng cp,g Dg,i

(15)

with

Lei

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

3264

l l

Dl

(16)

December 2005

Eh

l

l,0

(17)

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

the results of Figure 9.

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-

calculations of Figure 8 that have been conducted with consideration of heat transfer

from the liquid phase.

AIChE Journal

December 2005

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

3265

Ra

l gGD3

l Dl

(19)

l GD

l Dl

(20)

Ma

(18)

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-

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.

Convection and their Inuence on Enhancement

Factors

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

to

l,0

3266

l

GD

December 2005

AIChE Journal

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

10538.55

10543.43

x 0.08358

1 exp

0.067567

with

1

w 2 g lG D l/G D

2 l

(21)

x

Ra Ma

Re 2

Sc

2

eq

(22)

Eh 5 10 3

(23)

Sc

l

Dl

(24)

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 undisturbed

(26)

wGD

Reeq

l

(25)

December 2005

Ra Ma

DSc

4/5

(27)

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

3267

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.

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

(28)

(29)

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

(34)

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

l

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

Q

(30)

l lA 0T l,B T Ph

Q

(31)

l l

D l,bed

(35)

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

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

/1

(32)

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

l,0

D l,bed

h

(33)

with

AIChE Journal

December 2005

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

3269

Mas

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

l,0

bed

h

(36)

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

Ras

3270

l gh

l Dl,bed

(37)

December 2005

l

l Dl,bed h

(38)

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

bulk.

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