Delft University of Technology

Dynamic modeling of syngas fermentation in a continuous stirred-tank reactor

Multi-response parameter estimation and process optimization
de Medeiros, Elisa M.; Posada, John A.; Noorman, Henk; Filho, Rubens Maciel

DOI
10.1002/bit.27108
Publication date
2019
Document Version
Final published version
Published in
Biotechnology and Bioengineering

Citation (APA)
de Medeiros, E. M., Posada, J. A., Noorman, H., & Filho, R. M. (2019). Dynamic modeling of syngas
fermentation in a continuous stirred-tank reactor: Multi-response parameter estimation and process
optimization. Biotechnology and Bioengineering, 116(10), 2473-2487. https://doi.org/10.1002/bit.27108

Important note
To cite this publication, please use the final published version (if applicable).
Please check the document version above.

Copyright
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent
of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Takedown policy
Please contact us and provide details if you believe this document breaches copyrights.
We will remove access to the work immediately and investigate your claim.

This work is downloaded from Delft University of Technology.

For technical reasons the number of authors shown on this cover page is limited to a maximum of 10.
Received: 15 March 2019 | Revised: 11 June 2019 | Accepted: 3 July 2019

DOI: 10.1002/bit.27108

ARTICLE

Dynamic modeling of syngas fermentation in a continuous

stirred‐tank reactor: Multi‐response parameter estimation
and process optimization

Elisa M. de Medeiros1,2 | John A. Posada1 | Henk Noorman1,3 | Rubens Maciel Filho2

1
Department of Biotechnology, Delft
University of Technology, Delft, Abstract
The Netherlands Syngas fermentation is one of the bets for the future sustainable biobased economies
2
Laboratory of Optimization, Design and
due to its potential as an intermediate step in the conversion of waste carbon to
Advanced Control (LOPCA), School of
Chemical Engineering, University of Campinas ethanol fuel and other chemicals. Integrated with gasification and suitable downstream
(UNICAMP), Campinas, São Paulo, Brazil
processing, it may constitute an efficient and competitive route for the valorization of
3
DSM Biotechnology Center, Delft,
The Netherlands various waste materials, especially if systems engineering principles are employed
targeting process optimization. In this study, a dynamic multi‐response model is
Correspondence
Elisa M. de Medeiros, Department of presented for syngas fermentation with acetogenic bacteria in a continuous stirred‐
Biotechnology, Delft University of Technology, tank reactor, accounting for gas–liquid mass transfer, substrate (CO, H2) uptake,
van der Maasweg 9, 2629HZ, Delft,
The Netherlands. biomass growth and death, acetic acid reassimilation, and product selectivity. The
Email: E.MagalhaesdeMedeiros@tudelft.nl unknown parameters were estimated from literature data using the maximum
Funding information likelihood principle with a multi‐response nonlinear modeling framework and
BE‐Basic metaheuristic optimization, and model adequacy was verified with statistical analysis
via generation of confidence intervals as well as parameter significance tests. The
model was then used to study the effects of process conditions (gas composition,
dilution rate, gas flow rates, and cell recycle) as well as the sensitivity of kinetic
parameters, and multiobjective genetic algorithm was used to maximize ethanol
productivity and CO conversion. It was observed that these two objectives were clearly
conflicting when CO‐rich gas was used, but increasing the content of H2 favored higher
productivities while maintaining 100% CO conversion. The maximum productivity
predicted with full conversion was 2 g·L−1·hr−1 with a feed gas composition of 54% CO
and 46% H2 and a dilution rate of 0.06 hr−1 with roughly 90% of cell recycle.

KEYWORDS
dynamic model, ethanol, multiobjective optimization, parameter estimation, statistical analysis,
syngas fermentation

1 | INTRODUCTION route. It employs anaerobic bacteria called acetogens, which are

capable of autotrophically metabolizing CO, H2, and CO2 into cell
Gas fermentation is a promising biotechnological process that has mass, acids (e.g., acetate), and solvents (e.g., ethanol and butanediol).
gained attention due to its potential as a versatile waste‐to‐fuels The microbial substrate is, therefore, a gas with various possible

---------------------------------------------------------------------------------------------------------------------------
This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium,
provided the original work is properly cited.
© 2019 The Authors. Biotechnology and Bioengineering Published by Wiley Periodicals, Inc.

Biotechnology and Bioengineering. 2019;116:2473–2487. wileyonlinelibrary.com/journal/bit | 2473

2474 | DE MEDEIROS ET AL.

origins; it may be, for example, (a) syngas produced via gasification of intervals. The model was then used to study the effects of different
a wide range of feedstocks, including municipal solid waste and process conditions (i.e., gas composition, dilution rate, gas residence
lignocellulosic biomass; (b) off‐gas from steel production and cement time [GRT], and cell recycle), as well as the sensitivity of the kinetic
industries; (c) CO2 captured from power plants blended with H2 from parameters, and a multiobjective optimization was conducted for
renewable electricity, generated via electrolysis; and (d) reformed maximization of productivity and conversion. Although similar
biogas (Liew et al., 2016). There has been a great expansion of gas studies exist for other process, such as acetone–butanol–ethanol
fermentation technology over the last few years; at least three (ABE) fermentation (see e.g., Buehler & Mesbah, 2016), to our
commercial‐scale ethanol plants are currently under construction knowledge, there are no previous studies contemplating upon
(LanzaTech, 2018) or have started operation (China News Service, parameter estimation, statistical treatment, sensitivity analysis, and
2018), and many pilot plants have already been in operation for long multiobjective optimization of syngas fermentation; therefore, this
periods of time (Liew et al., 2016). Different studies indicate that the work was devised to fill this lacuna.
process can play an important role in the development of a
sustainable bioeconomy, being comparable to other lignocellulosic
processes in terms of cost, energy efficiency, and environmental 2 | MODEL DESCRIPTION
impact, while also permitting feedstock flexibility (Liew et al., 2016;
de Medeiros, Posada, Noorman, Osseweijer, & Filho, 2017; Pardo‐ The dynamic model developed in this study describes a stirred tank
Planas, Atiyeh, Phillips, Aichele, & Mohammad, 2017; Roy, Dutta, & with continuous supply of syngas and batch or continuous flow of
Deen, 2015). From the point of view of process systems engineering, liquid, with or without cell recycle. It accounts for two phases (G/L)
however, there is still vast room for improvement, from strain and seven species—CO, H2, CO2, ethanol (C2H6O or EtOH), acetic acid
enhancement and efficient product separation, to the integrated (C2H4O2 or HAc), water, and biomass–comprising 13 state variables
optimization of process parameters. With that in mind, in this article, which are the concentrations of the six chemical compounds in the gas
we address specifically the syngas fermentation bioreactor, coveting CG,j (mmol/L) and in the liquid CL,j (mmol/L)—where j = CO, H2, CO2,
the presentation and analysis of a model that can be useful in EtOH, HAc, H2O—as well as the concentration of biomass in the liquid
optimization frameworks. CX (g/L). Two types of input are provided to the modeling framework
Models to describe syngas fermentation are still scarce in the (a) kinetic parameters, which define the relations between biochemical
literature, and only a few authors have attempted to adjust kinetic reaction rates and concentrations of chemical species and cells—these
expressions to experimental data. Younesi, Najafpour, and Mohamed parameters are estimated in this study; and (b) operating conditions,
(2005) and Mohammadi, Mohamed, Najafpour, Younesi, and Uzir (2014) such as gas flow rate, dilution rate, agitation rate, and syngas
adjusted logistic curves to the growth of Clostridium ljungdahlii on composition—these are specified for each of the cases analyzed in
artificial syngas using experimental data from batch fermentation essays this work and their effects are further evaluated.
in serum bottles. Mohammadi et al. (2014) were also able to fit Fitting the model parameters with literature data turned out to
Gompertz equations to their experimental profiles of product formation, be a challenge due to several reasons. First, the number of
and uptake rate equations for CO, presenting estimations of kinetic experimental papers on syngas fermentation is relatively small
parameters that were later adopted by Chen, Gomez, Höffner, Barton, compared with other types of fermentation; and an even smaller
and Henson (2015) in their dynamic Flux Balance Analysis (FBA) model number provides data without coproduction of other chemicals, such
of a syngas fermentation bubble column. The latter was the first as butanol and butanediol. Among these, some provide exploratory
application of FBA in a dynamic model for syngas fermentation and the data of very long cultures in which several accidents or interventions
first spatiotemporal model of this process, but it was not compared with occur, and others fail to provide clear information about the process
experimental data. The same group also published an improved version conditions (e.g., often the gas flow rates are omitted from the text,
of their model, applied for CO fermentation with Clostridium auto- probably because they were not fixed during the experiment). In the
ethanogenum and considering uptake parameters obtained and pro- present work, the model parameters were estimated for five
tected by LanzaTech (Chen, Daniell, Griffin, Li, & Henson, 2018). different case studies from three different papers (C1; Phillips,
Furthermore, Jang, Yasin, Park, Lovitt, and Chang (2017) simulated CO Klasson, Ackerson, Clausen, & Gaddy; Phillips, Klasson, Clausen, &
fermentation in a batch culture of Eubacterium limosum KIST612 using a Gaddy, 1993); (C2; Gaddy et al., 2007); (C3A,B,C; Maddipati, Atiyeh,
dynamic model with kinetic parameters previously estimated by Chang, Bellmer, & Huhnke, 2011). These case studies have in common the
Kim, Lovitt, and Bang (2001), but this process results in the formation of use of continuous supply of syngas mixtures in stirred tanks and the
acetic acid as the only product, which has lower a value than ethanol. formation of acetic acid and ethanol as the only products. Table 1
In the present study, a dynamic model was constructed for syngas presents the main differences between the five scenarios, apart from
fermentation with ethanol production in a continuous stirred tank the liquid medium composition, which is omitted due to space
reactor (CSTR). The unknown model parameters were estimated with limitations. It is worth noting that C2 actually consists of 35 steady‐
a multi‐response minimization framework using experimental culture state points obtained under different conditions of gas composition,
data from the literature and the significance of parameters was flow rates and agitation, while C1 and C3A,B,C comprise dynamic
assessed with statistical analysis and generation of confidence data. C3 is one case study subdivided in three, that is, all of the
DE MEDEIROS ET AL. | 2475

T A B L E 1 Case studies used for the estimation of kinetic parameters

Case C1 C2 C3A,B,Ca
Microbe C. ljungdahlii C. ljungdahlii Clostridium strain P11
Number of experiments under different 1 35 3
conditions
Number of points per experiment (NE) 24 1 17
Number of types of responses (NR) 5 5 3
Type of data Dynamic liquid concentrations and Steady‐state liquid concentrations Dynamic liquid
gas conversions and gas conversions concentrations
Gas composition [H2:CO:CO2:inert] 20:55:10:15 20:65:10:5 5:20:15:60
16:27:6:51
50:45:0:5
b
Gas residence time (GRT; min) 33–100 4.25–30 20
Dilution rate (Drate; h−1) 0.0035–0.012 0.018–0.083 0
Agitation [rpm] 300–450 750–900 150
Cell purge fraction (XP)c 0.1 0.3–1 1
Reference Phillips et al. (1993) Gaddy et al. (2007) Maddipati et al. (2011)
a
Cases C3A to C3C differ in the amount of yeast extract (YE) or corn steep liquor (CSE), respectively, 1 g/L YE, 10 g/L CSE, and 20 g/L CSE.
b
Liquid volume divided by inlet gas flow rate.
c
Fraction of cells that are not recycled to the reactor vessel (i.e., XP = 1 when there is no cell recycle).

process conditions are the same, except for the concentration of inside the cell. In C. ljungdahlii, Richter et al. (2016) reported that the
yeast extract or corn steep liquor. enzymes needed for the synthesis of ethanol were always available in
The next three subsections present the modeling approach for excess and, as reducing equivalents are constantly being provided by
the specific production/consumption rates of species due to cell the oxidation of CO and H2 (see Equations (1) and (2) catalyzed by
fermentation (Reaction rates); the mass balance equations consider- carbon monoxide dehydrogenase and hydrogenase, respectively), the
ing in/out flows, gas–liquid mass transfer, fermentation, and cell authors suggest that ethanol is formed as soon as undissociated
recycle (Mass balance equations); and the calculation of special terms acetic acid and reducing equivalents reach a threshold concentration
that appear in the mass balance equations (Calculation of special required to make the reduction of acetic acid thermodynamically
terms). feasible.

CO + H2 O → CO2 + 2H+ + 2e− (1)

2.1 | Reaction rates

H2 → 2H+ + 2e− (2)
C. ljungdahlii and other acetogens assimilate CO, H2, and CO2 through
the Wood–Ljungdahl pathway to produce acetyl‐CoA, which is then With that in mind, we propose a kinetic model following the
used to produce cell biomass and products, as schematized in stoichiometry of the reactions presented in Equations (3)–(6), which
Figure 1. intend to generally represent the chemical reactions catalyzed by
In theory, acetyl‐CoA reduction towards ethanol is possible with the cell. The model accounts for the following assumptions: (a) The
aldehyde dehydrogenase, but this route is always thermodynamically uptake rates of CO and H2 follow Monod kinetics with inhibition by
less favorable and actually infeasible if H2 is the electron donor substrate and product (Equation (7)); (b) acetic acid and ethanol inhibit
(Bertsch & Müller, 2015). Indeed, Richter et al. (2016) found with substrate uptake with standard inhibition kinetics (Equation (7b)), but
proteome analysis of C. ljungdahlii that ethanol was produced ethanol inhibition is only activated after a threshold concentration is
exclusively through the AOR route. Ethanol production is favored achieved; (c) CO inhibits the uptake of H2 but not CO (Equation (7c))—
when acetate accumulates inside the cell due to growth limiting this was decided after preliminary estimation routines showed that a
conditions (i.e., biomass cannot be produced) or due to low CO inhibition constant for CO uptake could not be estimated with the
extracellular pH (Richter et al., 2016). In the latter case, undisso- experimental data used here; (d) biomass growth is a function of the
ciated acetic acid, which is prevalent under pH lower than 4.76 uptake rates of CO and H2 (Equation (8)) and cell death (Equation (9)),
(acetic acid pKa), diffuses freely through the cell membrane due to its and its composition is assumed constant; (e) acetic acid is produced
neutral charge; however it dissociates again in the cytosol where the from CO (Equation (3)) and H2/CO2 (Equation (4)); (f) ethanol is
pH is close to neutrality and it cannot be exported through the cell produced exclusively through reduction of acetic acid (Equations (5)
membrane without active transport processes (i.e., using cellular and (6)), with reaction rates that are hyperbolic functions of the acetic
energy), thus leading to the accumulation of acetate and protons acid concentration, also mimicking Michaelis–Menten kinetics
2476 | DE MEDEIROS ET AL.

1 1
IE = CL,EtOH
, IA = CL,HAc
1+ 1+ (7b)
KIE KIA

1
ICO , j = H2 = CL,CO
, ICO , j = CO =1 (7c)
1+ KI,CO

The specific biomass growth rate μ (h−1) is then calculated from

these uptake rates via yield coefficients YX,j (g/mol) for both
substrates as shown in Equation (8). Although H2 is not a source of
carbon, it is coupled with the consumption of CO2 and it has also
been shown to be associated with the growth rate (Mohammadi
F I G U R E 1 Schematic representation of syngas fermentation et al., 2014). The death rate rd is a function of cell concentration as
metabolism in Clostridium ljungdahlii under acidic pH, including acetic
shown in Equation (9), where kd is the death constant estimated in
acid diffusion through the cell membrane. In this study, ethanol
formation is considered possible only via the AOR pathway. ALDH, this study. It is worth noting that, with this equation, the growth rate
aldehyde dehydrogenase; AOR, aldehyde ferredoxin oxidoreductase is also affected by the concentration of inhibitors (ethanol, acetic
acid, and CO), and the effects of other nutrients and maintenance
issues are expressed in the yield coefficients and the death constant.
(Equation (10)); (g) the effects of pH are not directly included in the
model, but it is assumed that the estimated values of the kinetic μ = −νCO ⋅ YX ,CO − νH2 ⋅ YX ,H2 (8)
parameters associated with acetic acid uptake and reduction will
reflect the pH conditions adopted in the experiments used for the rd = kd ⋅ X (9)
parameter estimation. With these assumptions, we may calculate the
specific reaction rates νkR (mmol·g−1·hr−1)—where k indicates The reaction rates of acetic acid reduction (AcR), that is, νkR for
the reaction’s equation number, that is, Equations (3–6)—and the k = 5 and 6, are calculated with Equations (10a) and (10b), where the
specific consumption/production rates of species j, νj (mmol.g−1.hr−1), AcR
parameters νmax, AcR
j and Ks, j (j = CO, H2) are estimated in this study.
where a negative sign in the value of νj indicates that the species is The condition in Equation (10a) should be read as “for j = CO and
consumed, otherwise it is produced. k = 5, or for j = H2 and k = 6.” The expressions FAcR,j are only used to
make the equations clearer; they are not model parameters. The idea
4CO + 2H2 O → C2 H4 O2 + 2CO2 (3)
behind this set of equations is that acetic acid is reduced with

4H2 + 2CO2 → C2 H4 O2 + 2H2 O (4) hyperbolic kinetics limited by its concentration (Equation (10b)), and
the consumption rate of CO or H2 necessary to provide reducing
C2 H4 O2 + 2CO + H2 O → C2 H6 O + 2CO2 (5) equivalents to these reactions is bounded by the total uptake rates
previously calculated from Equation (7); thus it can be easily verified
C2 H4 O2 + 2H2 → C2 H6 O + H2 O (6) that νkR (k = 5, 6) tends to the expression FAcR,j when the uptake of
CO or H2 is significantly larger than FAcR,j, whereas it tends to −νj/2
Ethanol inhibition in acetogens is still a research gap in the when |νj/2| is smaller than FAcR,j (the division by 2 is due to the
literature, but there is evidence that it occurs in a similar fashion to stoichiometric coefficient of CO and H2 in Equations (5) and (6)).
what is observed in the ABE fermentation, for example, Ramió‐Pujol,
Ganigué, Bañeras, and Colprim (2018) observed that ethanol had 1 ⎛ 2FAcR, j ⎞
νkR = ⎛ ⎞ ⎜ ⎟ ⋅ νj , (j, k ) = (CO, 5) , (H2 , 6) (10a)
inhibitory effects on C. ljungdahlii, though much milder than butanol, ⎝ 2 ⎠ ⎝ 2FAcR, j + νj ⎠
but the authors were not capable of achieving the critical
j
concentration for full inhibition. The experimental data from case νmax,AcR ⋅ CL,HAc
FAcR, j = , j = CO, H2 (10b)
study C1 show an immediate decrease in gas conversion after the KSj,AcR + CL,HAc
ethanol concentration surpasses 35 g/L, after which the concentra-
tions of cells and products continue to increase for a while but The remaining substrate that is consumed can then be assumed to be
eventually drop as a result of low substrate conversion. To express used in Equations (3) and (4), and the corresponding reaction rates are
this behavior, the standard noncompetitive enzyme inhibition model calculated from Equation (11), where νAcR, j is the reaction rate of AcR
used for ethanol inhibition is only activated after CL,EtOH reaches this (Equations (5) or (6)) using substrate j (i.e., CO or H2), for example, νAcR, CO
threshold concentration. in Equation (11) corresponds to ν5R as calculated from Equation (10). The
total consumption/production rates of other components then follow the
νmax, j ⋅ CL, j
νj = − ⋅ IE ⋅ IA ⋅ ICO, j , j = CO, H2 (7a) stoichiometry of Equations (3–6) as calculated with Equations (12–15).
KS, j + CL, j
DE MEDEIROS ET AL. | 2477

(νj + 2νAcR,j) gas inside the reactor; QG,in and QG,out are the gas volumetric flow rates
νkR = − , (j, k ) = (CO, 3) , (H2 , 4) (11)
4 (L/hr) in/out the vessel, with the latter calculated as described in
Calculation of special terms; kLaj are mass transfer coefficients
νCO2 = 2ν3R − 2ν4R + 2ν5R (12)
calculated as described in Calculation of special terms; QL is the liquid
volumetric flow rate (L/hr). The specific rates νj, μ, and rd were
νEtOH = ν5R + ν6R (13)
presented in Reaction rates and are calculated accordingly at each time
(14) point; subscript in refers to inlet gas and liquid concentrations; and XP is
νHAc = ν3R + ν4R − ν5R − ν6R
the cell purge fraction, that is, the fraction of cells that are not recycled
νH2 O = −2ν3R + 2ν4R − ν5R + ν6R (15) to the vessel.

Hj MML
2.2 | Mass balance equations mj ∈ NC = (21)
RTρL
The mass balance equations are presented in the following manner: the
ρL RT
concentration fields, excepting biomass concentration, are divided into mj ∈ C = (22)
MML γj Psat, j
four categories regarding their phase (gas, G or liquid, L) and species type
(noncondensable [NC] or condensable [C]). The governing differential
2.3 | Calculation of special terms
equations, Equations (16–20), assume isothermal and isobaric operation,
as well as homogeneity and constant liquid and gas volumes in the Certain terms that appear in the right‐hand side of the ordinary
reactor. differential equations (ODEs), but which are not state variables, are
For NC species j in the gas phase, j ∈ {CO, H2 , CO2} : calculated as explained in the following.

dCG, j
dt
=
1
VG( )
⋅ (QG,in CG, j,in − QG,out CG, j )
2.3.1 | Outlet volumetric gas flow rate QG,out

− kL aj
( CG , j
mj ∈ NC
− CL, j
)( ) VL
VG
(16) QG,out is calculated from a mole balance in the gas phase considering
isobaric conditions inside the vessel. Taking into account the mass
transfer of NC species (j ∈ NC) from gas to liquid and the mass transfer
For C species j in the gas phase, j ∈ {EtOH , HAc , H2 O} :
of C species (j ∈ C) from liquid to gas, the total gas mole flow rate leaving
the reactor is calculated at each time with Equation (23). QG,out is then
dCG, j
dt
= ( )
1
VG
⋅ (QG,in CG, j,in − QG,out CG, j ) + kL aj
CL, j
mj ∈ C
− CG , j
( )( )
VL
VG calculated with the assumption of ideal gas in Equation (24).

(17)
NG, out ⎛
mol ⎞
= QG,in ∑ CG, j,in − ∑ ⎛ ⎛ CG, j − C ⎟⎞ V ⎞
⎜kL aj ⎜ L, j L⎟
For NC species j in the liquid phase, j ∈ {CO, H2 , CO2} : ⎝ hr ⎠ j j ∈ NC ⎝ ⎝ mj ∈ NC ⎠ ⎠

+ ∑ ⎛ ⎛ CL, j − C ⎟⎞ V ⎞
⎜kL aj ⎜ (23)
( )
G, j L⎟
dCL, j
dt
= ( )
QL
VL
⋅ (CL, j,in − CL, j ) + kL aj
CG , j
mj ∈ NC
− CL, j + νj CX (18) j∈C ⎝ ⎝ mj ∈ C ⎠ ⎠

For C species j in the liquid phase j ∈ {EtOH , HAc , H2 O} :

QG,out
( )
m3
hr
=
NG,out RT
P
(24)

dCL, j
dt
=
( )
QL
VL
⋅ (CL, j, in − CL, j ) − kL aj
CL, j
mj ∈ C (
− CG, j + νj CX
) (19)
2.3.2 | Mass transfer coefficients
The mass transfer coefficient kL a for air in water at T = 36°C is
calculated via Equations (25)‐(27). It considers a weighted
For the biomass concentration (in the liquid phase):
average between the values of k La estimated at 20°C for
noncoalescing (kL a0(20) ) and coalescing (kL a1(20) ) broth according to
dCX
dt
=
QL
VL ( )
⋅ (−CX ⋅ XP ) + μCX − rd (20)
the correlations proposed by van’t Riet (1979) for air in water
(Equations (25c) and (25d)), where P g/V L is the impeller power per
unit volume, which is estimated from the impeller ungassed
The gas–liquid equilibrium factors mj ∈ NC , mj ∈ C in Equations (16–20)
power P ug (Equation (26)) and the correlation for the ratio P g/Pug
are described in Equations (21) and (22), where R = 8.314 Pa·m3/mol·K
in Equations (27) (Cui, Van der Lans, & Luyben, 1996). The
is the ideal gas constant; MML and ρL refer to liquid phase molar mass
weighting factor f 0 is an unknown parameter which is estimated
(kg/mol) and density (kg/m3) assumed pure water at 36°C; and the
in this study. In Equations (25)‐(27), all variables are in SI units,
respective physical parameters—Henry’s law constants Hj (Pa), satura-
except for the temperature which is in °C. The ungassed power
tion pressures Psat,j (Pa) and infinite‐dilution activity coefficients γj∞—
number is assumed to be N p = 12.4 for two impellers (cases C1
can be found in the Table S1. VL and VG are the volumes (L) of liquid and
and C2) or N p = 16.5 (case C3) for three impellers based on the
2478 | DE MEDEIROS ET AL.

equation available in the New Brunswick Bioflo manual; N is the 3.1 | Estimation of model parameters
−1
agitation rate in s ; us is the gas superficial velocity (volumetric
The unknown model parameters β were estimated as β̂ using the
gas flow at the inlet divided by the reactor cross sectional area).
maximum likelihood principle (MLP; Himmelblau, 1970), with the
In all cases, the reactor is assumed to have a height/diameter
experimental data from the case studies presented in Table 1, which
ratio of 2 and an impeller diameter of 40% the reactor diameter,
are structured into five categories of response (for C1 and C2, j = 1
as standard in New Brunswick Bioflo bioreactors.
… NR with NR = 5) or three categories of response (for C3, j = 1 … NR
kL a(20) with NR = 3): ethanol (CL,EtOH), acetic acid (CL,HAc), and biomass (CX)
= 1.024(20 − T ), T = 36 (25a)
kL a(T ) liquid phase concentrations (g/L), as well as CO and H2 conversions
(XCO and XH2 [%]), which indirectly provide information about the
kL a(20) (hr−1) = f0⋅kL a0(20) + (1 − f0)⋅kL a1(20) (25b)
concentrations of these species. For C3A,B,C the gas conversions
were not available, so only the liquid concentrations were used. The

( () )
0.7
Pg MLP is built with three assumptions: (A1) independency of NE
kL a0(20) (hr−1) = 3600⋅ 0.002 (us )0.2 (25c)
VL
experiments (i = 1 … NE); (A2) the model is correct; and (A3)
experimental responses (yj,i) are uncorrelated and follow normal

( () )
0.4
Pg
kL a1(20) (hr−1) = 3600⋅ 0.026 (us )0.5 (25d) probability density functions (PDFs) around unknown correct
VL
responses (ηj,i) according to the variance model in Equation (30),
where rj,i are known response‐experiment factors and σε2 is the
Pug = Np ρL N3di 5 (26)
unknown fundamental variance (Himmelblau, 1970).

QG,in⋅N 0.25
di 2
≤ 0.055,
( 1−
Pg
Pug ) (
= 9.9
QG,in⋅N 0.25
di 2 ) (27a) ( )
yj, i → N ηj, i , σ j2, i , σ j2, i = rj, i σε2 (30)

( )
With Equation (30) and assumptions (A1) and (A3), it can be
QG,in⋅N 0.25
di 2
≥ 0.055, 1−
Pg
Pug
= 0.52 + 0.62
( QG,in⋅N 0.25
di 2 ) shown that the identities in Equations (31) result for y j , the NE × 1
(27b) vector of experimental values of response j at all points, where Ε(.),
Cov (.), W j , and η j represent, respectively, the expectancy operator,
The individual k Laj for each species is then obtained from the the variance‐covariance matrix operator, the NE × NE diagonal weight
reference air‐water k La by applying the penetration theory as in matrix for response j and the NE × 1 vector of correct values for
Equation (28) (Talbot, Gortares, Lencki, & de la Nouë, 1991), response j.
where Dfj is the mass diffusivity of species j in water
(Table S1). Ε ( y j) = η j (31a)

1/2 Cov ( y j ) = σε2⋅ W j−1 (31b)

kL aj = kL a
( )
Df j
Dfair
(28)

(
W j−1 = Diag rj,1, rj,2,…,rj, NE ) (31c)
3 | NU MERIC AL M ETHODS
It can also be shown (Himmelblau, 1970) with assumptions (A1),
The dynamic fermentation model described by the ODEs, Equations (A2), and (A3), and Equations (30)–(31) that the application of the
(16)–(20), and its supplemental algebraic equations in Model MLP to this multi‐response (NR = 5 or 3) estimation problem results in
description represent a nonlinear algebraic‐differential system which the minimization of the weighted sums of squares of residuals
demands specialized numerical solvers for stiff problems. In the written in Equation (32), where β̂ is the NP x 1 vector of estimated
present case, the ode15s variable‐order method from MATLAB was ()
parameters and yˆ j βˆ is the corresponding NE x 1 vector of model
used for time integration from a feasible initial condition, given the predicted responses. Due to its high nonlinearity and likely multi-
appropriate value of the vector of model parameters in Equation (29). modal nature, the objective function was minimized using the
The β vector of parameters (NP x 1,NP = 15) comprises the 14 kinetic metaheuristic method genetic algorithm (ga MATLAB function), but
parameters explained in Reaction rates, as well as the kLa weighting a bounded simplex algorithm (fminsearch MATLAB function) was also
factor f0 applied to deepen a candidate optimum when a good estimate of

(29)
[ AcR
βT ≡ νmax,CO νmax,H2 KS,CO KS,H2 KIE KIA KI,CO YX ,CO YX ,H2 νmax,CO KSAcR AcR AcR
,CO νmax,H2 KS ,H2 k d f0 ]
DE MEDEIROS ET AL. | 2479

initial point was known. In both cases, sensible lower and upper (1 − α )⋅100% probability (α = .05) of the Fisher PDF with degrees of
bounds were stipulated for β̂ . These bounds are displayed in freedom (1,NR⋅NE − NP ).
Table S2, jointly with ad hoc variable transformations to convert
the original unrestricted simplex algorithm into a bounded simplex βˆk − t1 − α / 2 ⋅ σˆ βˆk ≤ βk ≤ βˆk + t1 − α / 2
algorithm. ⋅ σˆ βˆk , σˆ βˆk = ⎡C⌢
o v ( βˆ) ⎤ (33)
⎣ ⎦kk
NR
Min ∑ ( ),
ψj βˆ T (βˆk)
2

j=1 () (
ψj βˆ = y j − yˆ j ) ( )
W j y j − yˆ j , j = 1... NR R (βˆk ) = 2
> ϕ1 − α (1, NR ⋅ NE − NP ) ⇒ βk ≠ 0
{βˆ } (σˆ βˆ )
k

(32) (34)

The factors rj, i (j = 1…NR, i = 1…NE) of the variance model of 3.3 | Steady‐state sensitivity and multiobjective
experimental responses in Equations (30) and (31), were chosen optimization
considering plausible variances of experimental values—for example, After the estimation of kinetic parameters, the model was used to
(10% of value)2—as well as the interests of the modeling framework, study the effects of several process conditions on the steady‐state
which can privilege more adherence onto some experimental productivity of ethanol (i.e., CL,EtOH ∙ Drate). The steady states were
responses (e.g., ethanol concentration) in detriment of others (e.g., obtained by integrating the ODE system until all the state variables
acetic acid concentration). The underlying fact is seen in Equation showed absolute gradients smaller than 10−6. This procedure was
(31c): as rj, i decrease the respective elements of the weight matrix found to be faster than solving the system of nonlinear algebraic
W j rise, increasing the “pressure” for adherence of ŷ j onto y j . In this equations, as this required the initial guesses to be very close to the
regard, the following choices were made after multiple estimation actual solutions. It can also be shown that, for a wide range of initial
test runs (a) for ethanol liquid concentrations (j = 1) rj, i = (0.025⋅yj, i )2; conditions, the steady state was stable and independent of such
(b) for acetic acid liquid concentrations (j = 2) rj, i = (0.05⋅yj, i )2 ; (c) for specifications (phase‐portraits depicting the dynamic trajectories are
biomass concentrations (j = 3) rj, i = (0.05⋅yj, i )2 ; (d) for CO conversions presented in Figure S2), therefore an arbitrary set of initial
(j = 4) rj, i = (0.1⋅yj, i )2; and (e) for H2 conversions (j = 5) rj, i = (0.1⋅yj, i )2 . conditions equal to those of case study C1 was used. With this
The experimental response values were read from the dynamic framework, the sensitivity was analyzed with respect to the gas
profiles (C1 and C3) or steady‐state outcomes (C2) reported in the composition (varying the molar fractions of CO and H2), the GRT, the
case studies considered here (see Table 1). For C1 and C3, the dilution rate (Drate), and also to the kinetic parameters under
( )
predicted responses yˆ j ti, βˆ at each time (with i = 1…NE) were different conditions of GRT and Drate. On the basis of these results,
obtained via ode15s numerical integration starting from the initial the process was optimized using multiobjective genetic algorithm for
conditions of liquid composition reported in the respective papers; the maximization of two conflicting objectives: Ethanol productivity
( )
for C2, the predicted responses yˆ j i, βˆ were obtained with the and CO conversion. The decision variables were three operating
integration starting from arbitrary initial conditions until sufficient conditions (GRT, Drate, and XP—cell purge fraction) and nine kinetic
time to reach steady state (as explained further in Steady‐state parameters, which could possibly be tuned with the design of the
sensitivity and multi‐objective optimization the steady state was nutrient medium, the choice of strain and/or genetic engineering. In a
found to be nonsensitive to the initial conditions). In all cases, the last study, the H2:CO ratio was also included as a decision variable.
initial gas‐phase concentrations were considered equal to the inlet For this optimization routine, the bounds were specified based on the
gas concentrations and the liquid‐phase concentrations of NC species ranges of kinetic parameters estimated for the five case studies (see
were considered equal to gas–liquid equilibrium concentrations. Table S3).

3.2 | Significance of parameters 4 | R E S U L T S AN D D I S C U SS I O N

The confidence intervals of the estimated parameters βˆk were
calculated with Equation (33), where t1 – α/2 is the abscissa at 4.1 | Parameter estimation and confidence
(1 − α / 2)⋅100% probability (α = .05) of the t Student PDF with intervals
NR⋅NE − NP degrees of freedom (NP = 15), and the estimated standard The full parameter vector β̲ in Equation (29) was first estimated with
deviations σˆ βˆk are the kth diagonal elements of the estimated variance‐ the data from C1. Since C2 employed the same strain, the maximum
covariance matrix of the parameters. For a detailed explanation the uptake rates (νmax), saturation constants (Ks), and inhibition constants
reader is referred to the Supporting Information Materials provided (KI) were fixed and the remaining eight parameters were re‐
with this work. The F test to reject the null hypothesis (i.e., parameter βk estimated with the steady‐state data from C2. It should also be said
is significant) with 95% probability is given (Himmelblau, 1970) in that the ethanol inhibition term was excluded from this case study
Equation (34), where ϕ1 − α (1, NR⋅NE − NP ) is the abscissa at since the reported data did not achieve the threshold concentration
2480 | DE MEDEIROS ET AL.

T A B L E 2 Parameter estimates with their 95% confidence intervals

Parameter Unit C1 C2 C3A C3B C3C
−1 −1
νmax,CO mmol·g ·hr 46.3 ± 5.33 46.3 ± 6.79 37.5 ± 5.28 37.5 ± 6.99 37.5 ± 2.70
−1 −1
νmax,H2 mmol·g ·hr 31.6 ± 5.30 31.6 ± 8.47 29.5 ± 3.81 29.5 ± 4.02 29.5 ± 2.27

KS,CO mmol/L 0.0115 ± 0.000637 0.0115 ± 0.00631 0.0454 ± 0.00670 0.0454 ± 0.0112 0.0454 ± 0.00525

KS,H2 mmol/L 0.675 ± 0.0853 0.675 ± 0.235 0.718 ± 0.0732 0.718 ± 0.197 0.718 ± 0.0427

KI,EtOH mmol/L 217 ± 38.9 – – – –

KI,HAc mmol/L 962 ± 127 962 ± 594 869 ± 117 869 ± 85.2 869 ± 183

KI,CO mmol/L 0.136 ± 0.0224 0.136 ± 0.110 0.827 ± 0.110 0.827 ± 0.179 0.827 ± 0.110

YX,CO g/mol 0.754 ± 0.133 1.34 ± 0.226 1.69 ± 0.365 1.92 ± 0.463 2.41 ± 0.301

YX,H2 g/mol 0.201 ± 0.0233 0.156 ± 0.0623 0.248 ± 0.0399 0.248 ± 0.0369 0.248 ± 0.0121
−1 −1
AcR
νmax,CO mmol·g ·hr 24.2 ± 2.85 37.6 ± 17.6 13.0 ± 1.27 20.2 ± 1.98 8.581 ± 0.620

AcR
KS,CO mmol/L 388 ± 20.3 303 ± 163 223 ± 16.1 557 ± 169 483 ± 57.2
−1 −1
AcR
νmax,H mmol·g ·hr 1.76 ± 0.166 22.2 ± 7.62 15.9 ± 1.73 15.9 ± 3.39 15.9 ± 1.16
2

AcR
KS,H mmol/L 464 ± 37.1 586 ± 287 72.7 ± 6.68 72.7 ± 12.6 72.7 ± 7.87
2
−1
kd h 0.00697 ± 0.000297 0.00862 ± 0.00453 0.0119 ± 0.00135 0.0112 ± 0.00271 0.00959 ± 0.00163

f0 – 0.988 ± 0.0464 0.958 ± 0.339 0.700 ± 0.0699 0.973 ± 0.243 0.988 ± 0.123

considered in the model. Similarly, for the three case studies C3A,B,C the other case studies this percentage is less than 20%. In fact, for
it was considered that the difference in nutritional supplement would most of the estimated parameters, C2 presents the highest
affect the cell yield, product selectivity, death/maintenance and the uncertainties among the cases, which is also due to the large variety
degree of coalescence of the liquid; therefore β̲ was estimated for of experimental conditions adopted in this case and relatively small
C3A (without ethanol inhibition) and for C3B and C3C all parameters number of samples. It should be noted that as more experiments are
CO CO
were fixed except for YCO, X , νmax,AcR , K S,AcR , kd , and f0, which were re‐ performed, new data can be incorporated into the modeling frame-
estimated. The results are presented in Table 2 along with their 95% work presented here and the parameters can be re‐estimated with
confidence intervals. Results of F test score for significance of higher accuracy.
parameters can be found in Table S4. The cell yields, specifically YX,CO, showed a wide variation among the
The estimated values of the maximum CO and H2 uptake rates five case studies, being the highest for C3C (the experiment with high
νˆmax,CO and νˆmax,H2 are comparable with CO uptake rates reported by concentration of corn steep liquor) at 2.41 g/mol (nominal value). As
other authors: Chen et al. (2018) obtained CO uptake rates from 41 expected the value of YX,CO increases from C3A to C3C as a result of
to 43 mmol·g−1·hr−1 in continuous cultures of C. autoethanogenum in a increasing concentrations of nutritional supplement. Clearly this
bubble column; Mohammadi et al. (2014) estimated a maximum rate parameter is specific to the culture conditions and microbial strain,
of 34 mmol·g−1·hr−1 for C. ljungdahlii; and Gaddy et al. (2007) and this can be verified by looking at the diversified results of cell yield
−1 −1
reported a large range of 14–100 mmol·g ·hr for different in syngas fermentation reported by different authors, some of which are
operating conditions with C. ljungdahlii. The saturation constants Ks, in good agreement with this study: 0.25 g/mol for clostridial bacteria P7
which are inversely related to the microbe’s affinity to the substrate, (Rajagopalan, Datar, & Lewis, 2002); 1.4 g/mol for C. ljungdahlii (Phillips,
reflect the disparity observed between the conversions of CO and H2 Clausen, & Gaddy, 1994); 2.1–3.2 g/mol for C. ljungdahlii (Mohammadi,
in cases C1 and C2: KS,CO is around 2% the value of KS,H2 in C1 and Mohamed, Najafpour, Younesi, & Uzir, 2016); 2 g/mol for Rhodospirillum
C2, and 6% in C3, while the pure component solubility of CO is only rubrum (Kerby, Ludden, & Roberts, 1995); 7.2 g/mol for E. limosum
approximately 13% higher that of H2 at the culture temperature. KIST612 (Chang et al., 2001).
With regard to the inhibition constants, it can be concluded that the The results generated by the model with the different parameter
effect of CO on H2 uptake was higher in cases C1 and C2 given the vectors are shown in Figures 2–4 along with the respective
lower value of KI,CO for these cases, although it should also be noted experimental points. The model showed overall reasonable predictive
that cases C3 use a small percentage of only 5% H2 in the feed gas. power for ethanol, acetic acid, and biomass concentrations, although
Acetic acid inhibition was also found to be statistically significant certain dynamic features were only roughly captured. For example, in
although its large value in all cases (>850 mmol/L) indicate small cases C3 the acetic acid peak around 75 hr was flattened and slightly
inhibitory effects under the conditions of the experiments considered displaced to the right (this was also the tendency of the experimental
here. For Case C2, the uncertainty associated with this parameter is data going from case C3A to C3C). In Case C1, the model was able to
also notably high, reaching around 60% of its nominal value while in predict the conversion decrease after 500 hr, but the experimental
DE MEDEIROS ET AL. | 2481

(a) (a)

(b) (b)

F I G U R E 2 Predicted dynamic profiles (solid lines) and F I G U R E 3 Predicted steady‐state responses (solid lines) and
experimental points (circles) for case study C1. (a) Concentration of experimental points for case study C2. (a) Concentration of products
products and cells in the liquid; (b) conversions of CO and H2 [Color and cells in the liquid; (B) conversions of CO and H2 [Color figure can
figure can be viewed at wileyonlinelibrary.com] be viewed at wileyonlinelibrary.com]

data also suggest a recovery which could not be predicted by the were more favorable to ethanol production, that is, with higher
AcR AcR
model. In the modeling framework, this decrease is a consequence of νmax, j (j = CO, H2 ). The saturation constants KS , j (j = CO, H2 ) were

joint inhibitory effects of ethanol and CO, the latter which similar if we consider the confidence intervals.
accumulates in the liquid phase due to impaired uptake as a The last parameter, f0, indicates the level of coalescence in the
consequence of the former, and acetic acid to a smaller extent. liquid, with higher f0 (as in cases C1 and C2) meaning the liquid is
Since this is the only experiment with such high concentrations of highly noncoalescing and thus enables higher gas–liquid mass
ethanol, it is unclear whether this behavior is due to product transfer coefficients. It is worth noting that f0 increased from Case
inhibition or if other external factors could be the cause of this C3A to C3C and specially from C3A to C3B (when 1 g/L yeast extract
perturbation. Although it is likely that ethanol exhibits inhibitory was replaced with 10 g/L corn steep liquor).
effects, as demonstrated by Ramió‐Pujol et al. (2018) and Férnandez‐
Naveira, Abubackar, Veiga, and Kennes (2016), a final conclusion
4.2 | Sensitivity of process conditions and kinetic
cannot be drawn from the current set of experiments with regard to
parameters
this matter, and further experimental investigation is needed to
evaluate critical product concentration and inhibition constants. With the fitted models, the performance of the bioreactor was
Case C2 (steady state) showed the highest deviations from the evaluated for different conditions of gas composition, dilution rate
experimental data as well as parameter uncertainties, which is and gas flow rates. For these sensitivity analyses, the parameter
probably due to the large range of process conditions encompassed vector estimated in C2 was used as basis. The effects of syngas
by the data. It should also be noted that it is unclear whether the composition are depicted in Figure 5 for ethanol productivity and CO
medium composition was kept fixed or not during these experiments. conversion. It can be seen that both responses are enhanced with the
Nonetheless, the model was still good at capturing the tendency of CO content, but there is a maximum outcome at H2:CO close to 1
the data, especially the concentrations of products and cells. In and the peak is slightly dislocated to the left (higher H2:CO) for CO
comparison with C1, which used the same strain, the AcR parameters conversion. This result suggests that the syngas composition can be
2482 | DE MEDEIROS ET AL.

(a) (a)

(b)
(b)

(c)

F I G U R E 5 Steady‐state ethanol productivity achieved with

different gas compositions (yCO2 = 1 − yCO − yH2 ) with fixed
conditions: GRT = 20 min, Drate = 0.025 hr−1, N = 500 rpm, cell
recycle = 90% [Color figure can be viewed at wileyonlinelibrary.com]

purge) and without. Both cases demonstrate how lower values of

GRT (i.e., higher gas flow rates) enhance the productivity due to
higher supply of substrate as well as higher gas–liquid mass
transfer coefficient, although at the expense of the CO conver-
sion. Moreover, as typically observed in chemostat cultures, the
F I G U R E 4 Predicted dynamic profiles (solid lines) of products productivity is a concave function of the dilution rate with a clear
and cells, and experimental points (circles) for case studies C3. (a)
maximum—in this case, also dependent on the gas flow rate.
Case C3A; (b) Case C3B; (c) Case C3C [Color figure can be viewed at
wileyonlinelibrary.com] From Figure 6, it is also clear that cell recycle enhances the
ethanol productivity (the maximum increases from around
tuned to improve the performance of the bioreactor, but the optimal 1.13–1.44 g·L−1 ·hr−1 ) and broadens the region of operation
composition would, of course, depend on the balance between extra without cell wash‐out. Moreover, the maximum cell mass
productivity/conversion in the bioreactor and extra energy costs in concentration increases from 4.5 to around 16 g/L when cell
upstream operations (gasification and gas conditioning). It was also recycle is used (response surfaces shown in Figure S6). The
observed that cell mass concentration always increased with the effects of agitation are not shown here, but response surfaces
fraction of CO, going from near 1 g/L at low values up to 11 g/L with with this variable can be found in Figure S3. Evidently, increasing
pure CO (figure shown in Figure S5). the agitation rate also enhances the mass transfer of CO and H 2
Assuming fixed gas composition of a CO‐rich gas, the between the gas and the liquid, which allows for higher
response surfaces shown in Figure 6 were generated to illustrate conversions and ethanol productivity; however, at the price of
the effects of dilution rate (Drate ) and GRT with cell recycle (10% higher energy consumption.
DE MEDEIROS ET AL. | 2483

(a) (b)

(c) (d)

F I G U R E 6 Steady‐state ethanol productivity and CO conversion as a function of gas residence time (GRT) and liquid dilution rate (Drate) (a,b)
with 90% cell recycle; (c,d) without cell recycle. Fixed conditions yCO = 0.65, yH2 = 0.2, yCO2 = 0.15, N = 500 rpm. The axes are rotated in (a) and (c)
[Color figure can be viewed at wileyonlinelibrary.com]

The same response surfaces were constructed for H2‐rich gas (see Parameter estimation and confidence intervals, indicating that
Figure S4), which had overall the same shape and tendencies as Figure 6. such microbial properties can be customized with the selection of
In accordance with Figure 5, it was also observed that increasing the strain and medium composition, besides of course genetic
content of H2 by adopting a gas composition of [H2:CO:CO2] = [50:45:5] engineering which would be a natural extrapolation of this
increased the maximum productivity to around 1.6 g·L−1·hr−1 when cell conclusion. The performance of the bioreactor can therefore be
recycle was used. However the maximum productivity under no cell improved with integrated design considering the simultaneous
−1 −1
recycle actually decreased from 1.13 to 0.93 g·L ·hr . effects of biokinetic parameters and process conditions.
With regard to the effects of kinetic parameters, it was
observed that the operating conditions contributed significantly
4.3 | Optimization of ethanol productivity and CO
to the sensitivity of this type of variable. Two illustrative
conversion
examples are given in Figure 7, where the parameters vary from
0.05 to 2 times their nominal value (as obtained in C2), and 6 The solutions to three optimization runs are shown in Figure 8.
combinations of low/high gas flow rate and dilution rate are The multiobjective optimization was first solved for three
employed. It can be seen that not only do lower values of GRT operating conditions (GRT, D rate, and XP) and nine kinetic
improve the productivity, but they also enhance the effects of parameters (all excepting the saturation and inhibition constants,
changing the kinetic parameters (see the inclination of solid lines which were fixed at the values obtained for C2). The lower and
in Figure 7 in comparison with the dashed lines). Figure 7a also upper bounds were chosen based on the intervals of the
suggests that increasing YX,CO will eventually lead to the same parameters estimated in Table 2 (see Table S3); for the operating
outcome of productivity for different values of Drate under the conditions, the GRT was free to vary in the range 5–50 min, D rate
same GRT (see green and blue lines). Very low values of Drate , in the range 0.005–0.2 hr −1 and XP in the range 0.1–1 (this
however, showcase the opposite trend: The two orange lines meaning no cell recycle). In the first run, the gas composition was
(D rate = 0.005 hr −1 ) are coincident until a bifurcation occurs at fixed for a CO‐rich gas, that is, [H2 :CO:CO 2] = [20:65:15]. For this
β / βˆ ≈ 0.75. These and other kinetic parameters showed con-
k k case, the Pareto‐optimal points reflected the classical problem of
siderable variation between the five estimations as presented in two conflicting objectives (Figure 8a): Higher gas flow rates
2484 | DE MEDEIROS ET AL.

(a) (a)

Normalized Parameter
GRT = 49.8 min
Drate = 0.17 h-1

GRT = 5.5 min

Drate = 0.044 h -1

Ethanol Productivity [g.L-1.h-1]

(b)
(b) GRT = 32.7 min
Drate = 0.052 h-1

Normalized Parameter
GRT = 8.2 min
Drate = 0.058 h -1

Ethanol Productivity [g.L-1.h-1]

(c)
GRT = 29.9 min
Drate = 0.047 h -1
F I G U R E 7 Sensitivity of steady‐state ethanol productivity to
H2:CO = 0.82
kinetic parameters under different conditions of gas residence time
Normalized Parameter

(GRT) and liquid dilution rate (Drate). (a) Cell yield on CO (YX,CO); (b)
cell death rate constant (kd). Fixed conditions yCO = 0.65, yH2 = 0.2,
yCO2 = 0.15, N = 500 rpm. The corresponding profiles of cell mass
concentration are shown in Figure S7 of the Supporting Information
Material [Color figure can be viewed at wileyonlinelibrary.com]
GRT = 8.6 min
Drate = 0.06 h-1
H2:CO = 0.85
(lower GRT) can enhance the productivity at the expense of CO
conversion, as a higher fraction of the gas is also wasted. The
Pareto front in Figure 8a is projected on the x–y axis, with the Ethanol Productivity [g.L-1.h-1]
−1 −1
highest productivity (1.5 g·L ·hr ) corresponding to 55% CO
conversion. These points are the so‐called nondominated solu- F I G U R E 8 Pareto‐optimal solutions for maximization of
steady‐state ethanol productivity and CO conversion (a) Fixed gas
tions, at which none of the objective functions can be improved
composition at 65% CO, 20% H2, and 15% CO2 with normalized
without harming the other. The decision variables GRT and Drate decision variables GRT and Drate plotted in the z‐axis; (b) fixed gas
tied to these points are plotted along the Pareto curve with their composition at 45% CO, 50% H2, and 5% CO2, all points correspond
normalized values on the z‐axis. The other decision variables to 100% CO conversion; (c) H2:CO ratio free to vary between 0
(including kinetic parameters) are not shown because their and 3, all points correspond to 100% CO conversion. In all cases the
decision variables are normalized with respect to their lower and
variation along the Pareto curve was considerably smaller, but
upper bounds, that is, GRT between 5 and 50 min and Drate between
maximum and minimum values of all decision variables from the
0.005 and 0.2 hr−1. The values of the most relevant decision variables
set of Pareto‐optimal points are presented in Table S5. In a at the solutions are shown in Tables S6–S8 [Color figure can be
second run (Figure 8b), the gas composition was changed to a H2 ‐ viewed at wileyonlinelibrary.com]
rich gas, that is, [H 2:CO:CO2 ] = [50:45:5]. Unexpectedly, for this
case the CO conversion could be maximized to 100% over a wide normalized GRT and Drate plotted on the y‐axis. It’s also worth
range of productivities, therefore instead of depicting a Pareto noting that, with a high content of H2 , even a relatively small GRT
front, Figure 8b presents the solutions of ethanol productivity of 8.2 min enabled full conversion of CO whereas also achieving a
obtained under 100% CO conversion, with the corresponding high productivity of 1.92 g·L −1 ·hr −1 —this point can thus be
DE MEDEIROS ET AL. | 2485

selected as the optimal solution in terms of both productivity and conditions the gains in productivity and conversion might compen-
conversion. sate for any extra spending with electricity. Other reactor designs
By analyzing the results of the kinetic parameters in both runs, it is should also be evaluated, such as bubble column, gas‐lift, and
clear that employing a H2‐rich gas also boosts the sensitivity of H2‐ membrane reactors. Ultimately, however, the bioreactor should be
AcR
related parameters, in special the parameter νmax,H2
which is related to optimized simultaneously with other unit operations, such as
the reduction of acetic acid—the average value of this parameter in the gasification and distillation, since optimal conditions in one unit
optimal solutions increased from 1.4 to nearly 20 mmol·g−1·hr−1 from the might lead to worse outcomes in other units with respect to
CO‐rich to the H2‐rich optimization run. Interestingly, the parameter economic and/or environmental issues.
AcR
νmax,CO was on average 20% smaller in the second case, while the other
parameters remained more or less constant. It was also noted that the
parameters of cell yield on CO and death constant were close to their 5 | CO NCL USIONS
specified bounds in both cases, with YX,CO being close to 2.4 g/mol (the
value estimated with the data from C3C) and kd being close to 0.007 hr−1 A dynamic model was presented for the production of ethanol via syngas
(the value estimated with the data from C1). While this result fermentation in a CSTR, and unknown kinetic parameters were estimated
demonstrates the efficacy of manipulating kinetic parameters to improve with literature data employing different conditions of gas flow rate,
bioreactor performance, it also raises the question as to whether it would dilution rate, syngas composition, and medium composition. The modeling
be feasible in real operation to use medium formulation and genetic framework was then used to evaluate the effects of different input
engineering to change the parameters independently from each other. variables on the outcomes of ethanol productivity and gas conversion,
Finally, a third optimization was conducted adopting the H2:CO ratio and it was observed that cell recycle rate, gas flow rate, and H2 content
in the feed gas as a new decision variable, which was allowed to vary had clear positive effects on the productivity, while the dilution rate gives
between 0 (pure CO) and 3 (25% CO and 75% H2). Also for this case, a different maximum depending on the other variables. Moreover, the
shown in Figure 8c, the solutions did not form a Pareto front, as 100% kinetic parameters were found to have different sensitivity patterns
CO conversion was attainable for a wide range of productivities. Figure depending on the process conditions, for example some of them having
8c is hence analogous to Figure 8b but H2:CO is included, although it is larger effects on the productivity when higher gas flow rates are used.
basically constant for all solutions at around 0.78–0.86. The maximum Since these parameters are specific to the type of strain and composition
productivity that can be obtained with 100% CO conversion is just over of the liquid medium, we conducted an optimization of productivity and
−1 −1 −1
2.0 g·L ·hr , which is achieved with GRT = 8.6 min, Drate = 0.06 hr , conversion using operating conditions and kinetic parameters as decision
XP = 0.11, and H2:CO = 0.85 (i.e., 54% CO and 46% H2). The kinetic variables, thereby showing the possibility of attaining higher values of
parameters at this solution are: νmax,CO = 40.3 mmol·g−1·hr−1, both responses at the same time. Implementation of the results predicted
νmax,H2 = 34.8 mmol·g−1·hr−1, YX,CO = 2.37 g/mol, YX,H2 = 0.223 g/mol, in this work would require further studies connecting the kinetic
AcR
νmax,CO = 34.2 mmol·g−1·hr−1, AcR
νmax,H2
−1
= 16.6 mmol·g ·hr ,−1 AcR
KS,CO = 398 parameters to the exact aspects of the liquid medium and strain
AcR
mmol/L, KS,H 2
= 396 mmol/L, and kd = 0.00546 hr−1. It is noteworthy that capabilities, as well as more experiments investigating the inhibitory
all of these parameters, with the exception of the yield coefficients, effects of products and CO. Therefore, as more experimental data
remain relatively close to the nominal parameters estimated for C2, in become available, the modeling framework presented here can be used
fact inside their confidence intervals, which suggests that efforts should to re‐estimate parameters, generate more accurate results and provide
be concentrated on enhancing the cell yields and not, for example, the new insights for integrated process optimization.
maximum uptake rates (at least for the conditions of gas–liquid mass
transfer encompassed by this study). Even though νmax is directly
AC KNO WL EDG M EN TS
associated with the cell’s capacity to take up substrate, the uptake rate
might just be limited by the mass transfer, such that after a certain point The authors thank DSM and the BE‐Basic Foundation for the
there would be no actual gain with increasing νmax. financial support provided in the form of a Ph.D. scholarship for E. M.
Another result from the optimization studies is that high M. This work is part of a dual degree Ph.D. project under the
productivities would be attained with very large cell concentra- agreement between UNICAMP and TU‐DELFT.
tions reaching up to 30 g/L (results shown in Figure S8), although
this depends on the gas composition. For example, 1 g·L−1·hr −1 of
CON F LI CT OF IN TE RES T S
ethanol productivity would be attainable with H 2‐rich gas with
operating conditions and kinetic parameters that result in 10 g/L The authors declare that there are no conflict of interests.
of cell mass, while for CO‐rich gas the cell concentration would
be a little over 20 g/L for the same ethanol productivity.
NOME NCLATURE
Agitation rate and gas recycle rate are also important operating
variables which were not included in this study, but should be Greek Symb ols
evaluated in the future with the inclusion of power consumption as a
β , βˆ vector of parameters: correct, estimated
third objective function. It is possible, for example, that under certain
2486 | DE MEDEIROS ET AL.

γj∞ activity coefficient at infinite dilution of component j in water ORCI D

at 36°C
Elisa M. de Medeiros http://orcid.org/0000-0002-3774-1721
ηj vector of correct responses j
μ biomass specific growth rate (h−1)
νj specific production/consumption rate of component j
R E F E R E N CE S
−1 −1
(mmol·g ·hr )
νkR specific rate of reaction number k (mmol·g−1·hr−1) Bertsch, J., & Müller, V. (2015). Bioenergetic constraints for conversion of
ρL mass density of water at 36°C (kg/m3) syngas to biofuels in acetogenic bacteria. Biotechnology for Biofuels, 8,
1–12. https://doi.org/10.1186/s13068‐015‐0393‐x
σε2 unknown fundamental variance
Buehler, E. A., & Mesbah, A. (2016). Kinetic study of acetone‐butanol‐
σ̂ estimated standard deviation ethanol fermentation in continuous culture. PLOS One, 11(8), 1–21.
https://doi.org/10.1371/journal.pone.0158243
Chang, I. S., Kim, B. H., Lovitt, R. W., & Bang, J. S. (2001). Effect of CO
Roman S ymbols partial pressure on cell‐recycled continuous CO fermentation by
Eubacterium limosum KIST612. Process Biochemistry, 37(4), 411–421.
CL,j concentration of component j in the liquid phase https://doi.org/10.1016/S0032‐9592(01)00227‐8
(mmol/L or g/L) Chen, J., Daniell, J., Griffin, D., Li, X., & Henson, M. A. (2018).
CG,j concentration of component j in the gas phase Experimental testing of a spatiotemporal metabolic model for
carbon monoxide fermentation with Clostridium autoethanogenum.
(mmol/L)
Biochemical Engineering Journal, 129, 64–73. https://doi.org/10.
CX concentration of biomass in the liquid (g/L) 1016/j.bej.2017.10.018
Cov , C⌢
ov variance‐covariance matrix: correct, estimated Chen, J., Gomez, J. A., Höffner, K., Barton, P. I., & Henson, M. A. (2015).
Df,j mass diffusivity of component j in water at 25°C Metabolic modeling of synthesis gas fermentation in bubble column
reactors. Biotechnology for Biofuels, 8, 1–12. https://doi.org/10.1186/
(cm2/s)
s13068‐015‐0272‐5
Drate dilution rate = QL/VL (hr−1) China News Service. (2018). Beijing iron maker to turn waste gas into
di impeller diameter (m) biofuels. Retrieved from http://www.ecns.cn/business/2018/02‐28/
f0 broth coalescence weighting factor 293914.shtml
Cui, Y. Q., Van der Lans, R. G. J. M., & Luyben, K. C. A. M. (1996). Local
GRT gas residence time = VL/QG,in (min)
power uptake in gas‐liquid systems with single and multiple rushton
Hj Henry’s law constant of component j dissolved in turbines. Chemical Engineering Science, 51(1), 2631–2636. https://doi.
water at 36°C (Pa) org/10.1016/0009‐2509(96)00128‐5
kL a0(20) , kL a1(20) mass transfer coefficient of air in water at 20°C: Férnandez‐Naveira, Á., Abubackar, H. N., Veiga, M. C., & Kennes, C.
noncoalescing and coalescing broth (hr−1) (2016). Carbon monoxide bioconversion to butanol‐ethanol by
Clostridium carboxidivorans: Kinetics and toxicity of alcohols. Applied
kl aj mass transfer coefficient of component j in water at
Microbiology and Biotechnology, 100, 4321–4240. https://doi.org/10.
36°C (hr−1) 1007/s00253‐016‐7389‐8
MML molar mass of water (kg/mol) Gaddy, J. L., Arora, D. K., Ko, C.‐W., Phillips, J. R., Basu, R., Wikstrom, C. V.,
N agitation rate (rpm or s−1) & Clausen, E. C. (2007). US Patent No. 7,285,402 B2.
Himmelblau, D. M. (1970). Process analysis by statistical methods. New
NE number of experimental points
York, NY: John Wiley & Sons, Inc.
NP number of unknown parameters Jang, N., Yasin, M., Park, S., Lovitt, R. W., & Chang, I. S. (2017).
Np ungassed power number Determination of volumetric gas–liquid mass transfer coefficient of
NR number of response variables used for parameter carbon monoxide in a batch cultivation system using kinetic
simulations. Bioresource Technology, 239, 387–393. https://doi.org/
estimation
10.1016/j.biortech.2017.05.023
P pressure inside reactor (Pa) Kerby, R. L., Ludden, P. W., & Roberts, G. P. (1995). Carbon monoxide‐
Pg, Pug gassed power and ungassed power (W) dependent growth of Rhodospirillum rubrum. Journal of Bacteriology,
Psat,j vapor pressure of component j at 36°C (Pa) 177(8), 2241–2244.
LanzaTech. (2018). LanzaTech commercial scale facilities. Retrieved April
QG, QL gas and liquid volumetric flow rates (m3/s, m3/hr, ml/
26, 2018, from http://www.lanzatech.com/facilities/
hr, or ml/min) Liew, F. M., Martin, M. E., Tappel, R. C., Heijstra, B. D., Mihalcea, C., &
R universal gas constant (J/mol·K) Köpke, M. (2016). Gas fermentation—A flexible platform for com-
rj,i stipulated response‐experiment factors mercial scale production of low‐carbon‐fuels and chemicals from
waste and renewable feedstocks. Frontiers in Microbiology, 7, 1–28.
T temperature (°C or K)
https://doi.org/10.3389/fmicb.2016.00694
us superficial gas velocity (m/s) Maddipati, P., Atiyeh, H. K., Bellmer, D. D., & Huhnke, R. L. (2011). Ethanol
VG,VL volume of gas and volume of liquid inside the reactor production from syngas by Clostridium strain P11 using corn steep
(ml, L, or m3) liquor as a nutrient replacement to yeast extract. Bioresource
Wj Technology, 102, 6494–6501. https://doi.org/10.1016/j.biortech.
weight matrix of response j
2011.03.047
XCO, XH2 gas conversions of CO and H2
de Medeiros, E. M., Posada, J. A., Noorman, H., Osseweijer, P., & Filho,
XP cell purge fraction R. M. (2017). Hydrous bioethanol production from sugarcane bagasse
y j, yˆ j vector of responses (g/L): experimental, predicted via energy self‐sufficient gasification‐fermentation hybrid route:
DE MEDEIROS ET AL. | 2487

Simulation and financial analysis. Journal of Cleaner Production, 168, Roy, P., Dutta, A., & Deen, B. (2015). Greenhouse gas emissions and
1625–1635. https://doi.org/10.1016/j.jclepro.2017.01.165 production cost of ethanol produced from biosyngas fermentation
Mohammadi, M., Mohamed, A. R., Najafpour, G., Younesi, H., & Uzir, M. H. process. Bioresource Technology, 192, 185–191. https://doi.org/10.
(2016). Clostridium ljungdahlii for production of biofuel from synthesis 1016/j.biortech.2015.05.056
gas. Energy Sources, Part A: Recovery, Utilization, and Environmental Effects, van ’t Riet, K. (1979). Review of Measuring methods and results in mass
38(3), 427–434. https://doi.org/10.1080/15567036.2012.729254 transfer in stirred vessels nonviscous gas‐liquid. Industrial and
Mohammadi, M., Mohamed, A. R., Najafpour, G. D., Younesi, H., & Engineering Chemistry, Process Design and Development, 18(3),
Uzir, M. H. (2014). Kinetic studies on fermentative production of 357–364. https://doi.org/10.1021/i260071a001
biofuel from synthesis gas using Clostridium ljungdahlii. The Scientific Talbot, P., Gortares, M. P., Lencki, R. W., & de la Nouë, J. (1991).
World Journal, 2014(1), 1–8. https://doi.org/10.1155/2014/910590 Absorption of CO2 in algal mass culture systems: A different
Pardo‐Planas, O., Atiyeh, H. K., Phillips, J. R., Aichele, C. P., & characterization approach. Biotechnology and Bioengineering, 37,
Mohammad, S. (2017). Process simulation of ethanol production from 834–842. https://doi.org/10.1002/bit.260370907
biomass gasification and syngas fermentation. Bioresource Technology, Younesi, H., Najafpour, G., & Mohamed, A. R. (2005). Ethanol and acetate
245, 925–932. https://doi.org/10.1016/j.biortech.2017.08.193 production from synthesis gas via fermentation processes using
Phillips, J. R., Clausen, E. C., & Gaddy, J. L. (1994). Synthesis gas as anaerobic bacterium, Clostridium ljungdahlii. Biochemical Engineering
substrate for the biological production of fuels and chemicals. Applied Journal, 27(2), 110–119. https://doi.org/10.1016/j.bej.2005.08.015
Biochemistry and Biotechnology, 45/46, 145–157. https://doi.org/10.
1007/BF02941794
Phillips, J. R., Klasson, K. T., Clausen, E. C., & Gaddy, J. L. (1993). Biological SUPPO RTING IN F ORMATION
production of ethanol from coal synthesis gas. Applied Biochemistry and
Additional supporting information may be found online in the
Biotechnology, 39–40(1), 559–571. https://doi.org/10.1007/BF02919018.
Rajagopalan, S., Datar, R. P., & Lewis, R. S. (2002). Formation of ethanol from Supporting Information section.
carbon monoxide via a new microbial catalyst. Biomass and Bioenergy, 23,
487–493. https://doi.org/10.1016/S0961‐9534(02)00071‐5
Ramió‐Pujol, S., Ganigué, R., Bañeras, L., & Colprim, J. (2018). Effect of
How to cite this article: de Medeiros EM, Posada JA,
ethanol and butanol on autotrophic growth of model homoacetogens.
FEMS Microbiology Letters, 365(10), 1–4. https://doi.org/10.1093/ Noorman H, Filho RM. Dynamic modeling of syngas
femsle/fny084 fermentation in a continuous stirred tank reactor:
Richter, H., Molitor, B., Wei, H., Chen, W., Aristilde, L., & Angenent, L. T. Multi‐response parameter estimation and process
(2016). Ethanol production in syngas‐fermenting Clostridium ljungdahlii
optimization. Biotechnology and Bioengineering. 2019;116:
is controlled by thermodynamics rather than by enzyme expression.
Energy and Environmental Science, 9(7), 2392–2399. https://doi.org/10. 2473–2487. https://doi.org/10.1002/bit.27108
1039/c6ee01108j