Académique Documents
Professionnel Documents
Culture Documents
1
k(1=t 1 k=2)
e
k
t
^t
ge
l^t
;
^
t < t
1 g(1 e
ltk
)e
l^t
;
^
t > t
_
_
(15)
For times smaller than the bubble residence time,
^
t < t,
the solution involves two terms of exponential decay with
rates k=t and l, whereas for
^
t > t, the saturation dynamics is
determined by l only. The initial phase corresponds to the
establishment of the concentration gradient in the bubbling
Figure 4. Left: Cellular uptake rate and overall efciency: [ ^ q
cell
[ is determined by a biologically motivated increasing function [ ^ q
cell
[ = f (^ c
l
) (thin line) and the function
[ ^ q
cell
[ = (1 e
k
)(1 ^ c
l
) from the new PIDE model for gasliquid mass transfer (thick line). The resulting value of the cellular uptake rate (circle) increases with k.
Right: Identication of the overall efciency k from steady-state measurements of the scaled headspace and liquid concentrations, ^ c
h
and ^ c
l
. Parameter values identied from the
new model M
PIDE
grad
(solid line) and the literature models M
con
(dotted line) and M
eff
(dashed line) assuming a=0. Values for the STR under consideration (circle):
(1 ^ c
h
)=(1 ^ c
l
) = 0:71, k
PIDE
grad
= 1:24.
3002 Biotechnology and Bioengineering, Vol. 109, No. 12, December, 2012
system: The rst full bubbles enter and release gas to the
liquid, while empty bubbles are still present and take up
gas from the liquid. It remains to explain the (scaled)
saturation rate l and the coefcient g, which depend on the
parameters k and t: in general, the function
p(s) = s(st k)(st k kt) k
2
(1 e
(stk)
) (16)
has two real zeroes, s
1
= (k=t) and s
2
<0, the latter of
which can be determined only numerically. Now, l =[s
2
[
and g = k(lt k)=(lp
/
(l)). We note that l(0; 1) for all
k and t, cf. Figure 6 (top-left). As a consequence, even for
highly efcient mass transfer, that is, k , the saturation
rate l is nite, since in this case mass ow is the limiting
process.
Now, we can compare the solution for the liquid
concentration predicted by the new model M
PIDE
grad
with
the solutions predicted by M
con
, M
eff
, and M
ODE
grad
. That is, we
compare Equation (15) above with Equations (71), (77), and
(82) in Supplements S4S6:
M
con
: ^c
l
(
^
t) = 1 e
k^t
(17a)
M
eff
: ^c
l
(
^
t) = 1
l
e
l
^t
e
l
^t
(17b)
M
ODE
grad
: ^c
l
(
^
t) = 1 e
l
0
^t
(17c)
where
l
=
1 k kt
2t
1 k kt
2t
_ _
2
k
t
(18a)
l
0
= l[
t=0
= 1 e
k
(18b)
From Figure 5, we see that the solutions differ
considerably for the parameter values we have chosen.
(The laboratory-scale STR under consideration is charac-
terized by a 1, k ~1, and t ~5.) Obviously, M
con
is not a
good approximation for small ^c
l
, since this model assumes a
constant bubble concentration throughout the dynamics,
that is, ^c
b
= 1, which in particular implies d^c
l
=d
^
t(0) > 0.
Hence, this model overestimates the increase in ^c
l
in the
initial phase. On the other hand, M
eff
predicts d^c
l
=d
^
t(0) = 0
correctly such that it is a good approximation in the initial
phase. However, this model uses a single effective bubble con-
centration, that is, ^c
b
(t), which does not cover the gradient
from inow to outow. In particular, it underestimates the
increase in ^c
l
in the later phase since it overestimates the
outow into headspace. For the parameters chosen, M
ODE
grad
is
only a slightly better approximation than the structurally
equivalent M
con
(since t 1=k is not fullled). This model
considers the gradient in the bubble concentration ^c
b
(t; x),
however, it assumes that the gradient is established
instantaneously, which causes an initial overestimation.
Finally, we study the validity of the ODE models as
approximations of the PIDE model for wide ranges of the
crucial parameters k and t. To this end, we contrast the
exponential term e
l^t
, which determines the long-term
dynamics of M
PIDE
grad
with the corresponding terms e
k ^t
, e
l ^t
,
and e
l
0
^t
in the solutions of M
con
, M
eff
, and M
ODE
grad
. This boils
down to a comparison of the saturation rate l with k, l
,
and l
0
. In Figure 6, we plot the saturation rate l as well as
the resulting relative errors (k l)=l, (l
l)=l, and
(l
0
l)=l over the (k, t)-plane. The most simple ODE
model M
con
is a valid approximation of the PIDE model only
if k _0.1, that is, in a small region of the (k, t)-plane. On the
other hand, M
eff
is a valid approximation only if k _0.1 or
k _10. Hence, M
eff
is valid in a larger region of the (k, t)-
plane than M
con
. Finally, M
ODE
grad
is a valid approximation
only if kt _0.1. This shows that t 1=k is not only a
sufcient, but also a necessary condition for the quasi
steady-state approximation of the PIDE model.
Parameter Identication from Time-Series Data
After predicting the liquid concentration over time for given
parameters k and t (and a 1), we solve the inverse
problem of identifying parameters from time-series mea-
surements of the liquid concentration. In particular, we are
interested in the overall efciency k (which contains k
L
a).
For the laboratory-scale STR under consideration, k is not
the only unknown parameter. For example, also t (which
contains the gas holdup V
b
) and the probe response time t
p
have to be identied. For a consistency check, we determine
the gas holdup and the probe response time from indepen-
dent experiments and nd t ~4.7 (from V
b
~90 mL) and
t
p
~40 s. We note that probe response is typically modeled
as a rst order process (Merchuk et al., 1990), where the
measured liquid concentration ^c
l;ms
depends on the actual
value ^c
l
via an additional ODE:
d^c
l;ms
dt
=
1
t
p
(^c
l
(t) ^c
l;ms
(t)) (19)
Figure 5. The liquid concentration as a function of time in the cell-free case
(gassing out experiment). Predictions by the new model M
PIDE
grad
(thick line), and the
literature models M
con
(dotted line), M
eff
(dashed line), and M
ODE
grad
(thin line).
Parameter values: a=0, k =1, t =5.
Muller et al.: Highly Efcient Mass Transfer in Aerated Bioreactors 3003
Biotechnology and Bioengineering
In the standard method for parameter identication
from gassing out data, linear regression is applied to the
transformed data ln(1 ^c
l
(t)). Thereby, one makes the
following assumptions: (i) the actual data ^c
l
can be obtained
from the measured data ^c
l;ms
by means of numerical
differentiation, cf. Equation (19); (ii) the transformed data
follow a linear model, that is, ln(1 ^c
l
(t)) = d k t; and
(iii) for a given dynamic model, the unknown parameter k
can be calculated from the (scaled) rate k t
s
. For the
laboratory-scale STR under consideration, all three assump-
tions are problematic: (i) since probe response time and
bubble residence time are of comparable size, the
establishment of the concentration gradient can be observed
only with a large uncertainty arising from numerical
differentiation of noisy data; (ii) the transformed data do
not exhibit one particular rate k, rather there is a fast increase
from zero to a maximum rate within one bubble residence
time followed by a slower decrease (to the reciprocal of the
headspace residence time); and (iii) not all parameters
which determine the rate constant in an individual model
are known, for example, t is needed to solve l(k; t) = k t
s
in
the new PIDE model. For comparison with other methods,
we take the maximum rate k =0.04 s
1
and use t =4.7.
By equating k t
s
with l in the PIDE model and k, l , and l
0
in the ODE models, we identify the values for k listed
in Table I (Gassing out/Linear regression). We note that
l
(k; t) = k t
s
cannot be solved for k; hence, there is no
result for M
eff
.
Figure 6. Saturation rate l as function of parameters k and t (top-left); relative errors of k, l , and l
0
with respect to l (top-right and bottom); errors cut off at 10% and view
points chosen for best visibility. In the cell-free case with negligible feedback from the headspace, that is, for ^ q
cell
= 0 and a 1, the long-term dynamics of mass transfer is
determined by l in M
PIDE
grad
, k in M
con
, l in M
eff
, and l
0
in M
ODE
grad
.
Table I. Overall efciency k identied for different dynamic models and from different data/with different methods.
Gassing out Continuous culture
Linear regression Optimization a=0 a=0.025
M
PIDE
grad
(0.51) 1.08 1.24 1.21
M
con
(0.31) 0.20 0.71 0.71
M
eff
() 1.57 2.46 2.27
M
ODE
grad
(0.38) 0.24 1.24 1.21
The new PIDE model yields the most consistent identication, whereas the ODE models from literature considerably under- or overestimate k. The linear
regression approach for gassing out data is problematic, cf. main text following Equation (19), and its results are put into brackets. When data from
continuous culture are used, k is identied for negligible headspaceliquid mass transfer a=0 and the realistic value a=0.025.
3004 Biotechnology and Bioengineering, Vol. 109, No. 12, December, 2012
Alternatively, we take an optimization approach to
parameter identication: We vary k (and the other
unknowns a, n, t, and t
p
) and minimize the least-squares
distance between model predictions and data. This method
is preferable, since numerical differentiation is avoided, each
dynamic model is directly t to the data, and all unknowns
are accounted for. On the other hand, it is computationally
more costly, since it involves the numerical solution of the
respective model and the evaluation of the data mismatch in
each optimization step. In Figure 7, we plot the optimal
solutions predicted by the new model M
PIDE
grad
(9) and the
literature models M
con
(8), M
eff
(4), and M
ODE
grad
(10). Clearly,
the new PIDE model gives the best t. In Table I (Gassing
out/Optimization), the identied values for k are listed. We
have already mentioned that M
con
overestimates the initial
increase in ^c
l
; this effect is compensated by an underesti-
mation of k. The same argument holds for M
ODE
grad
. On the
other hand, M
eff
underestimates the long-term increase in ^c
l
which is compensated by an overestimation of k.
Finally, we compare the parameter values identied from
the cell-free gassing out experiment and the (periodically
oscillating) continuous culture, cf. Table I. For the culture-
based method, we use a=0 as in Steady-State/Continuous
Culture Section as well as a =0.025, which has been
consistently identied from the gassing out data using the
optimization approach. Clearly, the new PIDE model yields
the most consistent identication of k, and we conclude
that k =1.21 in continuous culture. It remains to compute
k
L
a = k=t
s
using the minimal saturation time t
s
=7.65 s, cf.
Equation (7). The k
L
a values identied for the different
dynamic models and from the two datasets are summarized
in Table II. The new PIDE model yields k
L
a =569 h
1
in
continuous culture.
Biological Enhancement of Mass Transfer
The concentration gradient in the bubbling system is
neglected in standard dynamic models of gasliquid mass
transfer, which can result in a signicant under- or
overestimation of the volumetric mass transfer coefcient
k
L
a, depending on whether the model assumes a constant or
effective bubble concentration. Notably, the underestima-
tion by the textbook model M
con
(and the overestimation by
M
eff
) also depend on the method of parameter identica-
tion. The method based on measurements of cellular
exchange rates in continuous culture typically yields a higher
k
L
a than the gassing-out method performed in cell-free
medium thus confounding the determination of the
biological enhancement of mass transfer (Garcia-Ochoa
and Gomez, 2009; Ju and Sundarajan, 1992). In order to
estimate the biological enhancement factor E, we use the k
L
a
values identied from continuous culture and the gassing
out experiment:
E =
k
L
a
continuous culture
k
L
a
gassing out
(20)
The resulting values are given in Table II. Since standard
models lead to errors in the identication of k
L
a, these errors
propagate to the estimation of E, which in these cases must
be called an apparent biological enhancement factor. For
the culture under consideration, the textbook model M
con
yields an apparent enhancement factor of E =3.55, whereas
the new dynamic model M
PIDE
grad
yields a realistic value of
E =1.12.
Conclusions
This study provides a framework for rened modeling of
efcient aeration designs. Its main impact is the correct
identication of the volumetric mass transfer coefcent k
L
a,
which plays a crucial role in process monitoring and
development. In the presence of a concentration gradient in
the bubbling system, we advise to use the new PIDE model
Figure 7. Identication of the overall efciency k from time-series measure-
ments of the liquid concentration (gassing out experiment) by minimization of the
mismatch between model predictions and data (circles). Optimal solutions predicted
by the new model M
PIDE
grad
(solid line) and the literature models M
con
(dotted line), M
eff
(dashed line), and M
ODE
grad
(thin line). Parameters values identied from the individual
models listed in Table I (Gassing out/Optimization).
Table II. Volumetric mass transfer coefcient k
L
a (h
1
) identied for
different dynamic models and from different data/with different methods
(gassing out experiment using the optimization approach and continuous
culture using a=0.025, cf. Table I).
Gassing out Continuous culture
M
PIDE
grad
508 569 E =1.12
M
con
94 334 E =3.55
M
eff
739 1,068 E =1.45
M
ODE
grad
113 569 E =5.04
The resulting (apparent) biological enhancement factor E is given in the
rightmost column, cf. Biological Enhancement of Mass Transfer Section.
Dened glucose medium as described in Supplement S1. Experimental
conditions: liquid volume V
l
=0.650 L, aeration rate F
g
=0.150 L min
1
,
stirring with 750 rpm; temperature controlled at 308C, pH at 3.4. Average
dissolved oxygen in culture: 70%.
Muller et al.: Highly Efcient Mass Transfer in Aerated Bioreactors 3005
Biotechnology and Bioengineering
(9) or, depending on the reactor conguration, its ODE
approximation (10). To characterize an aerated STR,
we advise to use the dimensionless overall efciency
k = k
L
a k
H
(V
l
=F
g
) and the scaled bubble residence time
t. If k 1, then the gradient is negligible, and the standard
dynamic models (4) and (8) can be used. If t 1=k, that is,
the bubble residence time is much smaller than the time
scale for mass transfer, then the gradient in the bubble
concentration is established instantaneously, and the full
PIDE system can be approximated by an ODE system for the
liquid and headspace concentrations only.
Recently, a gradient has been considered in a model of
carbon dioxide uxes in a laboratory-scale photobioreactor
(Nedbal et al., 2010). The model has been derived based on
the assumption of a negligible bubble residence time, and we
note that it is equivalent to the ODE approximation of the
full PIDE model. While our considerations have been based
on experimental data for oxygen transfer, we note that
carbon dioxide has a much higher solubility (k
CO
2
H
= 0:74 vs.
k
O
2
H
= 0:029 at T=303 K), which implies that a gradient will
arise already in much less efcient bubbling systems, where
oxygen transfer may still be accounted for with standard
models.
Nomenclature
Variables:
c
b
concentration in the bubbles
c
h
concentration in the headspace
c
l
concentration in the liquid phase
q
cell
cellular mass exchange rate (conc./time)
Constants:
k
H
Henry constant (in conc./conc.)
k
L,b
mass transfer coefcient (between bubbles and liquid)
k
L,h
mass transfer coefcient (between headspace and liquid)
A
b
mass transfer area (between bubbles and liquid)
A
h
mass transfer area (between headspace and liquid)
V
b
bubble volume (gas holdup)
V
h
headspace volume
V
l
liquid volume
F
g
gas ow rate (aeration rate)
F
l
liquid ow rate
c
g,in
gas concentration at inow
c
l,in
liquid concentration at inow
t
s
minimal saturation time (reference time scale)
t
p
probe response time
Dimensionless Parameters:
a ratio of headspace and bubble k
L
a
n ratio of headspace and bubble volume
t scaled bubble residence time
k overall mass transfer efciency
E enhancement factor
We thank James Lu and Clemens Zarzer for fruitful discussions. R.M.
acknowledges nancial support by the Vienna Science and Technolo-
gy Fund WWTF, project MA07-30, the Marie Curie Actions Network
HARVEST, project 238017, and the research program Theoretical
Biology, SFB 618, HU Berlin. D.M. is grateful for funding from the
Japan Science and Technology Agency JST, Yamagata prefecture and
Tsuruoka city.
References
Agrawal P, Lee C, Lim H, Ramkrishna D. 1982. Theoretical investigations of
dynamic behavior of isothermal continuous stirred tank biological
reactors. Chem Eng Sci 37:453462.
Alves S, Maia C, Vasconcelos J. 2004. Gasliquid mass transfer coefcient in
stirred tanks interpreted through bubble contamination kinetics. Chem
Eng Process 43:823830.
Dhaouadi H, Poncin S, Hornut J, Wild G, Oinas P, Korpijarvi J. 1997. Mass
transfer in an external-loop airlift reactor: Experiments and modeling.
Chem Eng Sci 52:39093917.
Dhaouadi H, Poncin S, Midoux N, Wild G. 2001. Gasliquid mass transfer
in an airlift reactorAnalytical solution and experimental conrma-
tion. Chem Eng Process 40:129133.
Dhaouadi H, Poncin S, Hornut J, Midoux N. 2008. Gasliquid mass transfer
in bubble column reactor: Analytical solution and experimental con-
rmation. Chem Eng Process 47:548556.
Finn R. 1954. Agitationaeration in the laboratory and in industry. Bacteriol
Rev 18:254274.
Garcia-Ochoa F, Gomez E. 2009. Bioreactor scale-up and oxygen transfer
rate in microbial processes: An overview. Biotechnol Adv 27:153176.
Gourich B, Vial C, Azher NE, Soulami MB, Ziyad M. 2008. Inuence of
hydrodynamics and probe response on oxygen mass transfer measure-
ments in a high aspect ratio bubble column reactor: Effect of the
coalescence behaviour of the liquid phase. Biochem Eng J 39:114.
Han L, Al-Dahhan MH. 2007. Gasliquid mass transfer in a high pressure
bubble column reactor with different sparger designs. Chem Eng Sci
62:131139.
Heinzle E, Dunn I, Furukawa K, Tanner R. 1983. Modelling of sustained
oscillations observed in continuous cutlure of Saccharomyces cerevisiae.
In: Halme A, editor. Modelling and control of biotechnical processes.
London: Pergamon Press. p 5765.
Ju LK, Sundarajan A. 1992. Model analysis of biological oxygen transfer
enhancement in surface-aerated bioreactors. Biotechnol Bioeng 40:
13431352.
Marques M, Cabral J, Fernandes P. 2010. Bioprocess scale-up: Quest for the
parameters to be used as criterion to move from microreactors to lab-
scale. J Chem Technol Biotechnol 85:11841198.
Merchuk J, Yona S, Siegel M, Zvi A. 1990. On the rst-order approximation
to the response of dissolved oxygen electrodes for dynamic K
L
a
estimation. Biotechnol Bioeng 35:11611163.
Monod J. 1949. The growth of bacterial cultures. Annu Rev Microbiol
3:371394.
Murray D, Beckmann M, Kitano H. 2007. Regulation of yeast oscillatory
dynamics. Proc Natl Acad Sci USA 104:22412246.
Nedbal L, Cerveny J, Keren N, Kaplan A. 2010. Experimental validation of a
nonequilibrium model of CO
2
uxes between gas, liquid medium, and
algae in a at-panel photobioreactor. J Ind Microbiol Biotechnol
37:13191326.
Strassle C, Sonnleitner B, Fiechter A. 1988. A predictive model for the
spontaneous synchronization of Saccharomyces cerevisiae grown in
continuous culture. I. Concept. J Biotechnol 7:299317.
Whitman W. 1923. Preliminary experimental conrmation of the two-lm
theory of gas absorption. Chem Metall Eng 29:146149.
Wu L, Lange H, Van Gulik W, Heijnen J. 2003. Determination of in vivo
oxygen uptake and carbon dioxide evolution rates from off-gas mea-
surements under highly dynamic conditions. Biotechnol Bioeng 81:
448458.
3006 Biotechnology and Bioengineering, Vol. 109, No. 12, December, 2012