Simutis 2015

Biotechnology Biotechnol. J. 2015, 10, 1115–1130 DOI 10.1002/biot.
201500016
Journal
www.biotechnology-journal.com
Review
Bioreactor control improves bioprocess performance
Rimvydas Simutis1 and Andreas Lübbert2

1 Department of Automation, Kaunas University of Technology, Kaunas, Lithuania
2 Department of Biochemistry/Biotechnology, Martin-Luther-University Halle-Wittenberg, Halle/Saale, Germany
The performance of bioreactors is not only determined by productivity but also by process quali-
Received 10 JAN 2015
ty, which is mainly determined by variances in the process variables. As fluctuations in these quan- Revised 09 MAR 2015
tities directly affect the variability in the product properties, combatting distortions is the main task Accepted 01 JUN 2015
of practical quality assurance. The straightforward way of reducing this variability is keeping the
product formation process tightly under control. Purpose of this keynote is to show that there is
enough evidence in literature showing that the performance of the fermentation processes can sig-
nificantly be improved by feedback control. Most of the currently used open loop control proce-
dures can be replaced by relatively simple feedback techniques. It is shown by practical examples
that such a retrofitting does not require significant changes in the well-established equipment.
Feedback techniques are best in assuring high reproducibility of the industrial cultivation process-
es and thus in assuring the quality of their products. Many developments in supervising and con-
trolling industrial fermentations can directly be taken over in manufacturing processes. Even sim-
ple feedback controllers can efficiently improve the product quality. It’s the time now that manu-
facturers follow the developments in most other industries and improve process quality by auto-
matic feedback control.
Keywords: Automation · Bioprocess Performance · Feedback Control · Fermentation · Process State Estimation
1 Introduction perspective, it is also necessary to keep the variations in

the product and biomass trajectories within the individ-
There are many possibilities to improve the performance ual production runs and from batch-to-batch within nar-
of a fermentation process. The most often used criterion row limits.
is maximizing the productivity. From a product quality While in the beginning of the development of a pro-
duction process for a new therapeutic protein the pro-
ductivity is industry’s number one priority, the focus in
the final phases of the process developments is product
Correspondence: Prof. Andreas Lübbert, Department of Biochemistry/ quality. Unfortunately, most molecular details of the prod-
Biotechnology, Martin-Luther-University Halle-Wittenberg, ucts that influence their efficacy cannot quantitatively be
D-06120 Halle/Saale, Germany, monitored online during product formation process. Thus,
E-mail: andreas.luebbert@biochemtech.uni-halle.de
the goal formulated in FDA’s PAT initiative [1] is online
Abbreviations: ANN, artificial neural network; BCR, base consumption rate; supervising the cultivation in order to timely detect devi-
cCPR, cumulative carbon dioxide production rate; CHO, Chinese hamster ations from the desired path that may lead to product vari-
ovary cells; cOUR, cumulative oxygen uptake rate; CPR, carbon dioxide pro- ations. By experience the best way to keep the product
duction rate; CTR, carbon dioxide transfer rate; EMA, European Medicine within its specs is making sure that the cultivation close-
Agency; EPO, erythropoietin; FDA, Federal Drug Administration; OUR, oxy- ly follows the path of the reference process from which the
gen uptake rate; p, pressure; PAT, process analytical technology; pCO2, dis-
desired product was obtained. In the manufacturing jar-
solved carbon dioxide partial pressure; PCV, packed cell volume; PF, parti-
cle filters; PID, proportional-integral-differential controller; pO2, dissolved
gon this led to the phrases: “The process is the product”
oxygen partial pressure; RVM, relevance vector machine; SNR, signal-to- [2], or with the words of the quality assurance icon
noise ratio; SVM, support vector machine; T, temperature; tCPR, total car- Taguchi: “Processes of good quality produce products
bon dioxide production rate; tOUR, total oxygen uptake rate without variability” [3, 4].
© 2015 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim 1115

Biotechnology Biotechnol. J. 2015, 10, 1115–1130
Journal
www.biotechnology-journal.com www.biotecvisions.com
A straightforward procedure of reducing the variabili- practical applications in industry and they write their
ty is referred to as parameter design in quality engineer- papers for control specialists and not for applicants. The
ing. This aims in setting the adjustable parameters in statement of Luyben [7] for the case of chemical engi-
such a way that the process becomes more robust with neering: “Complex elegant control systems look great on
respect to the usually appearing fluctuations in the paper but soon end up on ‘manual’ in an industrial envi-
process conditions. The dream of quality strategists in ronment” is still valid for biotechnology as well. Even
biotechnology is finding robust paths within the design mathematically well-educated scientists have their diffi-
space, on which the process is kept when its reference culties to catch at a reasonable expense what they need
variable, e.g. the feed rate, follows a predefined recipe. for their own work [8].
Throughout the manufacturing processes in the Here we try to translate some important control ideas
process industries, however, the most effective technolo- to the language of biotechnologist and biochemical engi-
gy employed to increase robustness is closed-loop control. neers and provide examples from fermentation technolo-
Here the deviations from the desired trajectories are gy. A main aim of this paper is to bridge the gap between
measured and when they exceed relevant sizes they are control theory and practical applications at fermentation
compensated for by correcting actions. This review sheds processes and so to encourage fermentation specialists to
some light on the current state of bioreactor control that improve their processes by feed-back control.
enables process performance. We constrain the discus-
sion to controlling cultivations for recombinant protein
production. 2 Bioprocess models relevant to process
A compilation of the literature discussing this topic supervision and control
was given by Schuler and Marison [5], who focused their
interest on controlling the specific biomass growth rate μ. 2.1 Dynamic models for biotechnological process
These authors pointed out, that μ is the key variable of
many bioprocesses, consequently monitoring and con- Control has many meanings which often vary between
trolling μ is very important in the development of reliable communities. We use it in the engineering sense. Engi-
and robust processes for industrial production of high-val- neers develop their controllers in a model-aided way.
ue products such as recombinant proteins. Unfortunately Since controllers must dynamically react on disturbances
the direct control of μ can be problematic in industrial it is necessary to consider the process dynamics in these
applications where the batch-to-batch variability must be models [9]. Process models can be used to derive efficient
kept low. Consequently it is necessary to develop alterna- control algorithms in different ways: e.g. they can be used
tive control methods that control μ indirectly. in model predictive controllers or they can be transformed
With an indirect control of the specific growth rate a into transfer-function forms and used for designing tradi-
high reproducibility and quality of cultivation processes tional PID controllers. Models also are necessary for deriv-
can be achieved. Unfortunately, with a few exceptions, ing algorithms for process supervision, for process fault
e.g. Sauer and Preuss [6], such control techniques are cur- detection, analysis, and correction. Since the computing
rently not used in industrial protein manufacturing power of today’s industrial automation systems is steadi-
processes. Substrate feeding is often controlled manually ly increasing, model-aided methods for bioprocess super-
by process operators or in an open-loop control mode. vision and control can be implemented without problems.
FDA, in its PAT initiative [1] already recognized the prob- The dynamic process models are mainly based on
lem but could, within the decade up to now, not signifi- mass balances around the culture which are formulated
cantly change the situation. for the key variables influencing the process performance.
Main obstacle for introducing advanced controls are To a first approximation, the process performance can be
missing accurate, precise, and timely measurements of quantified by the mass mp of product that has been pro-
the state of the process. Also introducing new develop- duced within a predefined production time of length tf,
ments in fermentation control is strongly retarded as com- which is a function of the biomass concentration X(t),
panies are too much focused on biologic aspects of their the culture volume V(t) and the performance of the
processes often ignoring the fact that improved cellular cells, expressed by the specific product formation rate
performance immediately demands for a better supply of 1 dP
the organisms. They also hesitate running into problems π(t) =
X dt
with the administrations although the latter claim to sup-
tf
port improvements in knowledge-based process supervi-
sion and control that allow reducing risks with respect to ∫0
m p = π XVdt (1)
process and thus product quality.
A further obstacle is the scientific literature on Hence, the trajectories of the product and biomass con-
process control. There many researchers consider control centrations, P(t) and X(t) are primarily required to quanti-
theory as a branch of practical mathematics far away from fy the process performance.
1116 © 2015 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim

Journal
Process models are most often formulated assuming ing the state variables to the possible measurement vari-
ideal stirred tank behavior of the bioreactors, since then ables. This requires an additional model, the so-called
the models are systems of ordinary differential equations. soft-sensors or measurement equations
Otherwise, complex partial differential equations would
become necessary that cannot easily be solved for most y = h(c,F) (5)
realistic initial and boundary conditions.
In industrial practice, most cultivation processes are These are usually static functions mapping the state vari-
operated in the fed-batch mode, where the process ables c onto the measurement quantities y. This way of
dynamics is kept under control by the feed of the sub- formulation is chosen to directly allow comparing meas-
strate with rate F and concentration cF. When their state urement values with model predictions. It must be distin-
can accurately be described by c = [X,S,P], a vector con- guished from the inverse representation which used
taining the concentrations of biomass X, substrate S and when values of state variables are compared or for state
product P, and the biochemical formation rate by the vec- estimation.
tor R, the ordinary differential equations assume the form The choice of appropriate measurement variables y
can directly be derived from equation (4). We see that
dc F
= R + (c F − c) (2) product and biomass formation are very closely related to
dt W
the oxygen uptake, the carbon dioxide formation and the
The feed leads to an increase in the culture mass W, base consumption rates.
which is basically described by This is the first but not the only one argument to meas-
ure these quantities. The second one is that currently very
dW
=F (3) robust and reliable measurement devices are available for
dt
measuring OUR and CPR as well as the base consumption
The amounts of the various species within the culture do during pH control. From the point of view of process
not change independently of each other. They are bound supervision in industrial manufacturing this is a very
to the stoichiometry of the biochemical processes taking important issue.
place. As in chemical engineering it proved useful to bio- The third point speaking for measuring these quanti-
chemical engineering to approximate the stoichiometry ties is that all three variables are global variables that rep-
by means of a gross reaction equation that considers only resentatively reflect the culture conditions as a whole as
those species which change to considerable amounts compared to pO2 and pCO2 measurements which can
during the conversion process. only be made locally at a limited number of points in the
As pointed out in more detail in advanced textbooks reactor.
(e.g. [10]), this reaction equation is formulated in a form Alternatively the metabolic heat formation rate HFR
that shows how the quantities consumed are element- can be measured for monitoring the conversion process.
wise distributed on the metabolic products. As well-known this is proportional to the oxygen con-
sumption rate, where the proportionality constant is prac-
S + Yso O + Ysn N = Ysx X + Ysp P + Ysc C + Ysw H (4) tically constant even for different organisms. HFR can also
be measured using established and robust measurement
Here S is the amount of substrate consumed using the devices for temperatures and flow rates. Again this is an
amount O of oxygen and N, the nitrogen source (most integral measurement, however, it requires considering
often NH3), to form biomass X, product P, carbon dioxide additional heat formation (e.g. heat produced by agita-
C and water H. The Yij are the corresponding stoichio- tion) and heat loss mechanisms (e.g. evaporation heat). In
metric coefficients or yields. production-scale reactors with low culture viscosities, the
Important to our discussion is that the formation rates heat produced by the agitator as a function of the impeller
of biomass and product are simply related to the con- speed can separately be measured, stored and then used
sumption rates of oxygen, the nitrogen source, and pro- online by table-lookup. Thus, it is an attractive alternative
duction rate of CO2. These stoichiometric constraints to off-gas measurements [11].
must be considered in the equations (2) within the volu- Clearly, there is always some uncertainty with respect
metric conversion rate vector R that is otherwise domi- to the accuracy of the models in Eq. (2)–(4). No model
nated by kinetic expressions. Note that the stoichiomet- completely describes the dynamic behavior of the real
ric coefficients change from the biomass growth to the system. Deviations may be due to phenomena that have
product formation phase. been neglected, e.g. unideal mixing leading to internal
The state variables X, S, and P are not directly mea- time delays. However, for most control and supervision
sureable online, at least not to an acceptable accuracy, purposes the model formulated in Eq. (2)–(4) is good
rate, and with sufficiently small time lags. From the point enough, as feedback control automatically reduces the
of view of process control they are thus unsatisfactory. effects of small uncertainties. It is possible to design feed-
This problem, however, can be solved indirectly by relat- back systems on strongly simplified models [12]. It’s one

Journal
of the key uses and the most useful properties of feedback relationships already mentioned, do only work well
that it provides robustness to uncertainty. Consequently, enough in the biomass growth phase of the cultivation
feedback control is well-suited to increase process and processes. They most often fail during the transition into
product quality. the recombinant product formation phases.
This can intuitively be understood by recognizing that The alternatives discussed in literature are static
the feedback controller becomes part or component of the black-box relationships. In its PAT initiative FDA/EMA
processes which (e.g. by changing the flow rates of sub- already proposed to make use of the redundancies in the
strate feeding) change the dynamic behavior of the various possible measurement variables by means of mul-
process. One central aim is making the entire system tivariate approaches, where the first approximation is a
more robust against randomly appearing distortions by multivariate linear regression between OUR, CPR, and
appropriately designing the controller component. BCR and X.
The set point profiles, i.e. the concrete trajectories A simple, but very effective improvement is using
along which to run the cultivation processes, may they be instead the corresponding cumulative signals cOUR and
derived by trial and error or by model-aided optimization, cCPR. cBCR is already a cumulative signal. As already
should not only be chosen with respect to productivity but shown by Jenzsch et al. [16] much smaller estimation
also concerning robustness. errors compared to the Luedeking-Piret approach can be
Although productivity is traditionally the first per- achieved. A further upgrade can be obtained when quad-
formance variable considered, robustness becomes more ratic terms of the signals are included, e.g.:
and more important in therapeutic protein manufacturing
processes. Usually robustness is obtained at the cost of X = a1cOUR + a2cOUR² + a3cOURcCPR + a4cCPR +
productivity. One way to cope with these conflicting a5cCPR² + … (6)
demands is feedback control which allows keeping both
product quality and productivity at high levels. Such expressions are still linear in their parameters ai and
thus they can still be determined using standard library
2.2 Measurement models for process supervision routines for multiple linear regression.
Encouraged from the success of these estimates
Indispensable for online process supervision as well as which are nonlinear functions of the measurement vari-
feedback control are measurements of variables that time- ables, it is straightforward to try the step to measurement
ly deliver information about the current state of the process. equations, which are also nonlinear in the parameters.
As the state variables themselves can usually not be meas- A relatively simple first approach is using
ured accurately enough online, their values must be
1
derived indirectly from appropriate measurement quanti- X = w ( h) (7)
1 + exp(−w ( i) MV )
ties. As shown above, the most informative measurement
variables in aerobic fermentation processes are the oxygen where X is formulated as a function of the measurement
consumption rate (OUR), the carbon dioxide production values MV, e.g. MV = [cOUR, cCPR, cBase]. w(h) and w(i)
rate (CPR), and the base consumption rate (BCR). are parameter vectors corresponding to the ai in Eq. (6).
Unfortunately, due to changes in the metabolism Eq. (7) is referred to as a feedforward artificial neural net-
across the cultivation, the gross stoichiometry, discussed work (ANN). Hence, ANNs are by no means esoteric rep-
in equation (4) does not accurately relate the biomass to resentations based on black arts, they are simple static
OUR, CPR, or BCR. Pure linear relations as stoichiometry parametric black-box functions. Jenzsch et al. [18] com-
suggests, are insufficient. Even simple quasi-mechanistic pared the various biomass estimators discussed so far
models such as the Luedeking-Piret-type-relationships and showed that the ANN (7) led to the lowest estimation
[13] are not accurately enough for most purposes, since error.
the yields of oxygen consumption per biomass growth The parameters of ANNs must be determined by fit-
and the corresponding maintenance coefficient change ting (7) to available measurement data, which is usually
significantly, particularly upon the induction of heterolo- referred to as network training. It’s just nonlinear param-
gous product formation. Hence, more accurate mappings eter fitting and can be performed with numerical library
are required. routines. Most program libraries offer routines particular-
Various investigators (e.g. [14–15]) investigated alter- ly suited to neural networks.
native online estimation techniques. They can be divided Such training algorithms are also available within the
into static and dynamic relationships. process automation systems which are installed in all big
biotech companies (e.g. Siemens AG, NeuroSystems:
2.2.1 Static estimators for the values Add-on for PCS7 Process Control Systems, Siemens
of the key state variables Industrial Automation Systems, Erlangen, Germany 2014;
Unfortunately most static equations that are derived on Emerson, DeltaV Product Data Sheet DeltaV Neural,
mechanistic reasoning, e.g. the Luedeking–Piret-type Emerson Electric Co., Ferguson, Missouri, 2013). This pro-

Journal
Figure 1. Typical example showing the

accuracy with which a simple artificial
neural network determines the biomass
in a 20k animal cell culture from online
measured signals of aeration rate of oxy-
gen, total gas supply, base consumption,
and culture volume [51] (Reprinted with
permission). PCV means packed cell vol-
ume, it is equivalently used to biomass
in the pharmaceutical industries using
cell cultures.
vides evidence for the fact that artificial neural networks is better known in form of Einstein’s famous sentence:
are already integral part of industrial process monitoring “Everything should be kept as simple as possible, but not
and control techniques. Figure 1 illustrates the quality of simpler!”
an ANN-based biomass estimation in a production-scale This problem led to a new generation of black-box rep-
animal cell culture. More complicated estimates should resentations which are sparse in the number p of free
only be used when such simple networks fail. parameters. They were developed in the machine-learn-
With all parameter fit procedures, so also with respect ing community, a branch of contemporary statistics. The
to ANN-training as well as for the training of the later currently most widely accepted sparse representations
mentioned estimation techniques, the question arises, are so-called kernel techniques. These are linear combi-
how many training datasets must be provided. The gen- nations of nonlinear basis functions, i.e. linear forms,
eral answer is that this depends on the quality of the train- where the basis-functions are data centered, i.e. fixed to
ing data as the following discussion shows. In the cited available process data. The most well-known variant is
references 20–30 data sets were often used. the “Support Vector Regression” where a specialized fit-
Statistics teaches that the estimation error, character- ting procedure leads to a very small number p of basis-
ized by the total variance σ ²t, is a sum of two terms, the functions (and hence coefficients) to accurately describe
modelling uncertainty σ ²bias that appears as a bias and the relationship between input and output data [19, 20].
the parameter uncertainty σ ²param. The bias is inversely Another approach, the “Relevance Vector Machine” of
proportional to the model complexity, most often charac- Tipping [21], takes the same kernel approach, i.e. the
terized by the number p of free parameters, while the same linear form, but determines the number and the val-
parameter uncertainty is proportional to p but inversely ues of the parameters using Bayesian statistics. This
proportional to the number N of measurement data automatically optimizes the number of free parameters.
employed. The practical advantage of using these sparse kernel
techniques is that repeated fits result in the same param-
σ ²t = σ ²bias + σ ²param eter set, which is not guaranteed with ANNs. Generally
σ ²bias ~ 1/p (8) the kernel techniques should give more reliable estimates
σ ²param ~ p/N with the same amount of data as compared with simple
ANNs.
The conflicting dependency of the two summands with
respect to the model complexity is referred to as the vari- 2.2.2 Dynamic state estimators for biomass
ance-bias tradeoff. It says that for a data set of size N, and its specific growth rate
increasing the model complexity p reduces the bias, but Throughout all industries, the most often used dynamic
increases the parameter uncertainty and vice versa. state estimation procedure is the Kalman filter. Conse-
The heuristic solution of the conflict is using as small quently Kalman filters were intensively analyzed and
p values as possible: Simpler models may capture the investigated by many authors involved in monitoring and
underlying structure better and may thus have a better control of bioprocesses starting with Stephanopoulos and
prediction or generalization performance. This statement San [22]. The work was extended by many others, e.g. by

Journal
Bogards and Van de Wouwer [23] and Venkateswarlu [24] tion that the uncertainties in the estimates and the meas-
to name only two of them. urements can accurately enough be described by Gauss-
The Kalman-Filter works recursively and provides ian statistics, i.e. by the first and the second moments
one-step-ahead estimates of the state variables. It is a pre- only. The UKF is much simpler to implement as compared
dictor-corrector procedure, where the dynamic model (2) to the EKF and usually leads to better estimates [26, 27].
is used to predict the next estimate, which is then cor- Chai et al. [28] also concluded that the UKF is simpler to
rected using the difference between the corresponding implement and that it provides in biological waste water
values of the measurement equations (5) and the current plants more accurate estimates than the EKF. Similar con-
measurements. The weight with which this difference is clusions were made by Zhu and Feng [29] who applied
used to correct the predicted state is referred to as the EKF and UKF algorithms in the glycerol fermentation
Kalman gain. The convincing advantage of this approach process.
is that it takes into account the uncertainties in the The second problem is the model accuracy. Both, the
dynamic model and the measurement data. EKF as well as the UKF, do only work perfectly if the mod-
As we are dealing with nonlinear processes, the orig- els are at all able to describe the process. The dynamic
inal Kalman filter, which assumes linear process models, models are flexible enough to following the dynamics of
does not work properly. State of the art across all indus- the cultivation processes. More problematic are the
tries [25] is the extended Kalman filter (EKF), which uses measurement equations. The simple Luedeking-Piret
the same philosophy coping with the nonlinearity of the approach does not lead to sufficiently accurate estimates.
process model by linearizing it at the current process time Dropping the linearization constraints opens the way
instant. for using all the measurement models discussed in the
Two problems arise with applications of the Kalman previous section on static state estimators, even the more
filter to cultivation processes. The first is that linearization complex nonlinear black-box models such as ANN, SVM,
is particularly difficult when step changes appear in the and RVM. As shown in Fig. 2, the combination of a black-
process. The second is the accuracy by which the process box model, here an ANN, with the dynamic process mod-
can be described with the available models. el (2) leads to a robust estimate of the biomass trajectory
In order to avoid linearization errors, the extended in a recombinant protein production process. The ANN
Kalman filter is more and more being replaced by can easily be replaced by an SVM or a RVM or any other
unscented Kalman filters (UKF) which allow using the up-to-date static biomass estimator.
full nonlinear process and measurement models, main- As the static biomass estimators (ANN, SVM, RVM)
taining the original philosophy of the Kalman filter. The already led to robust biomass estimates, as discussed
Kalman gain, however, is still determined on the assump- above, the question arises, why at all using the complex
Figure 2. UKF estimate of X, P, and μ,

where Eq. (2)–(3) were used as the
dynamic model and an ANN was used
as measurement Eq. (5) to determine
the measured quantities cOUR and cCPR
from the model predictions of X.

Journal
model-based Kalman filters? There are several points to putation time than EKFs, they lead to lower estimation
be mentioned: The Kalman filter is of another quality as errors both for biomass concentration and specific growth
compared to static black-box estimators in so far as it con- rate estimations. An important practical advantage is that
siders details of the current process online while static the tuning procedure for Particle filters is simpler.
estimators that were fitted offline to a series of data sets
describe the behavior of the process on the average. The
Kalman filter takes care of this by considering the model 3 Quality-by-design for cultivation processes
and measurement uncertainties.
When the dynamic model in Eq. (2) works well, the Biochemical engineers are not exclusively interested in
product concentration can also be estimated, i.e. the responding to process noise. As far as possible they try to
measurement equations used within the UKF do not nec- construct process equipment and operational procedures
essarily need estimates of the product mass. so that unavoidable disturbances do not significantly
A further point is that the dynamic approach can also influence the process performance. Wherever possible,
be used to estimate model parameters that are changing they aim in constructing robust processes.
during the cultivation process, e.g. the specific growth
and product formation rates. For that purpose, the state 3.1 Process robustness measures
vector can formally be augmented by these quantities.
When we allow for random variations of these quantities, Many industrial bioreactors depict, although equipped
the Kalman filter will adapt their values step by step. This with all conventionally used controllers, considerable defi-
leads to estimates of the trajectories not only of the state ciencies in their batch-to-batch reproducibility. Detailed
variables but also of these parameters. analyses of the process data often showed that the vari-
Figure 2 gives an example for an UKF estimating the ability in the inoculum size is the reason. Since the bio-
biomass growth rate profile for a cultivation experiment. mass is initially growing exponentially, small deviations
The measurement variables used were cOUR and cCPR. from the set points quickly lead to significant deviations
Hence, the UKF can recursively estimate not only the in its trajectories. The difficulty is that in the original data,
time development of the state variables but also parame- small random deviations from the target are usually not
ters in the dynamic process model. Again, the advantage detectable.
of UKF over EKF is that it makes use of the original dynam- During an initial batch phase, the specific biomass
ic model as is used for the process simulations as well. growth rate μ is nearly all time equal to its maximal value
μmax. Only when practically all substrate is consumed, the
2.2.3 Particle filters as advanced bioprocess state growth rate ceases. Then in most industrial fermentation
estimators processes, substrate feeding is switched on, which usu-
Although EKF and UKF algorithms have been providing ally increases the substrate concentrations to obtain max-
good results for state estimation of important process imal specific growth rates in order to keep the biomass
variables like biomass concentration and specific growth growth phase short. Consequently, as mentioned, initial
rate, they have not been widely used in industrial appli- biomass variations become drastically amplified with
cations in biotechnology for two main reasons: (i) for time.
Kalman filters it is well known that tuning of the algo- Another aspect that leads to drastic variations from
rithms to special plants, i.e. determining the particular batch to batch in unlimited growth situations is that the
covariance matrices, may be quite complicated; and (ii) cell properties change slightly from batch to batch.
they are restricted to either weakly nonlinear models that A change in μmax, e.g. a diminution from 0.7 h–1 to 0.675 h–1,
allow for local linearization (EKF) or practically Gaussian corresponding to 3.5% difference not uncommon in indus-
distribution densities for the distortions. Both restrictions trial practice, will propagate an initial biomass concen-
can be lifted with Particle filters PF which are recursive tration of 0.25 g/kg to 82.7 g/kg and 65.2 g/kg respective-
implementations of Monte-Carlo-based signal processing ly when the substrate does not limit biomass growth. This
techniques [30, 31]. makes a difference of 21%. Such simple rough estimates
With the increasing computing powers in the process show that a high reproducibility in principle cannot be
automation systems, such algorithms become attractive achieved in such processes when they are operated in a
for state estimation in biotechnological processes. mode of unlimited growth.
The algorithms approximate the Bayesian posterior This process instability can be converted into a stable
probability density functions of state variables with a set process mode, by running the growth phase with an
of randomly chosen weighted samples, each sample exponential feeding corresponding to a specific growth
being referred to as a particle. Recently we applied PF rate μset which is slightly smaller than μmax and the
algorithms in virtual experiments of Escherichia coli cul- desired initial biomass concentration X0 set. The question
tivations [32]. Though the simulation results showed that how to eliminate small deviations in the initial biomass
Particle filter algorithms require significantly more com- concentrations and how to make this process repro-

Journal
Figure 3. (A) Result of the Monte-Carlo-Simulations showing the impact of the distribution of the initial biomass concentrations (lower left) on the bio-
mass trajectories for the cases μ = μmax (red) and μ < μmax (blue). The characteristic asymmetry for the case μ = μmax can clearly be seen. (B) Experimental
validation (Data from [51], reprinted with permission).
ducible was answered by Jenzsch et al. [33]. What is eas- production bioreactors. Some new techniques, like auto-
ily confirmed with simple calculations is that when associative artificial neural networks are more powerful in
X0 > X0 set the cells see less substrate than they need and detecting cases where values of measurement variables
thus their growth rate becomes smaller. Correspondingly, do not fit into the concert of the others.
when X0 < X0 set, the cells see more substrate than expect- In advanced distributed control systems, dynamic
ed and their specific growth rate rises. This obviously bioprocess models can be implemented to detect devia-
leads to a self-adjusting behavior of the culture. A rise in tions from the target. Such software extensions are now
μ, however, is only possible when the cells are not already referred to as virtual plants. They act as process simula-
growing at their maximal pace. The consequence is that tors that differ from traditional ones used for process
an initial batch phase or substrate feeding with μmax must design and optimization in that they run synchronously
be avoided when the objective is keeping the batch-to- with the running cultivation process while using all the
batch reproducibility high. original software modules for measurement devices, actu-
A straightforward way of determining the sensitivity ators and controllers that are also used during real process
of state or control variables on the variability of other [37]. When a process failure has been appearing and the
process variables is the Monte-Carlo-Simulation. A result corresponding responsible variables have been identified,
showing the influence of an initial biomass variability is automated fail-safe techniques were proposed to finish
shown in Fig. 3. the process properly without violation of its specification
limits [38].
3.2 Supervision and fault analysis In everyday practice the acceptance of more elaborate
of cultivation processes or complex techniques has been motivated by the failure
of simpler approaches. If open-loop control fails in assur-
As process failures can be very expensive in production ing the desired process reproducibility, considering feed-
reactors, supervision tools should supplement the control back control is the straightforward next step.
automation systems. As such, special algorithms are
straightforward that are able to detect errors in cultivation
processes online. Simple soft-sensors based on mass and 4 Quality-by-control
carbon balances around the culture are very useful to
check for the consistency of real-time data while the Expensive process automation systems are used at prac-
bioreactor is operating [34–36]. Unfortunately, even these tically all fermentation processes in the protein manufac-
simple techniques are not common-place in industrial turing industries. These coordinate a large amount of

Journal
measurement and control modules. On the first glance, ing intensity which changes with the course of the culti-
these processes seem to be well controlled. However, the vation. Thus, despite of some attempts of controlling the
controllers installed in practice are most often very simple substrate concentration that appeared in literature, the
controllers that independently keep under control envi- substrate concentration is no alternative to the specific
ronmental quantities, like pressure, temperature, pH, pO2, growth rate as a control variable when the batch-to-batch
etc., in microbial and cell cultures. Controlling the state or reproducibility is an objective.
physiological variables is the exception. With respect to The first serious alternative to the specific growth rate
state variables these systems only perform some open μ as the control variable is obviously the biomass concen-
loop control, i.e. they act as simple programmable units tration X itself. The decision for a concrete operational
that route the values of action variables, e.g. the substrate procedure is not only the decision for an appropriate μset(t)
feed rate in fed-batch cultures, according to a predefined profile but also for a profile of the corresponding biomass
recipe. trajectory Xset(t). When the biomass trajectory is tightly
Before going into details of advanced control, some kept on its desired path, the specific biomass growth rate
basic problems that often constraint these controllers profile will be followed as well. Biomass control works
with respect to enhancing the performance of fermenta- well, as the prerequisite, the estimation of the biomass
tion processes need to be discussed. concentration, can be performed to a high accuracy as
was pointed out above.
4.1 What to control during cultivation runs However, an even better choice for an indirect μ-con-
trol is keeping the process on the desired trajectory for the
As mentioned in the introduction, most investigators in mass of carbon dioxide mCO2 produced during the culti-
advanced bioprocess control took the specific biomass vation [42]. mCO2 (and mO2) are tightly related to the bio-
growth rate μ as their control variable since this is con- mass profile as is shown by the stoichiometry (Eq. (4)).
sidered the basic physiological variable [5, 39, 40, 41]. As The advantage of mCO2 is that it can directly be comput-
mentioned, the specific growth rates should be μ < μmax ed as the total cumulative of the measured carbon dioxide
to obtain a high process reproducibility. However, direct production rate CPR, without going back to any special
μ-control leads to two problems: estimator.
(i) μ control requires online-determined μ-values. Esti- The set point profile for mCO2 can be taken from the ref-
mates of μ from biomass trajectories usually do not erence cultivation, the “golden batch”. It can also be
yield noiseless results even not after sophisticated fil- determined using initial biomass concentration X0, the
tering techniques. Generally determining time deriv- specific growth rate set point profile and estimates
atives from noisy data is not recommended. Even the between growth rate, biomass concentration and carbon
advanced indirect estimations techniques such as dioxide production rate (e.g. Luedeking–Piret-relation-
Kalman filters (EKF or UKF) as well as Particle filters ships, ANN, SVR or RVR models).
(PF) result in noisy μ-estimates and are thus not well In order to support process reproducibility, the set
suited for robust feedback control. point profile should be lowered by 2.5–5%. When the
(ii) When short-time disturbances occur in the process or process is tightly kept on the mCO2 profile, the process will
in the instrumentation, for instance when the sub- also follow the desired specific growth rate profile and this
strate feeding is shortly interrupted, that lead to a leads to high biomass and product concentration repro-
slightly changed biomass concentration, such an off- ducibility during the cultivation runs. Here it is very
set cannot be eliminated later on when the controller important to recognize that because of the cumulative
exactly follows a predefined μ-profile. This was clear- nature of the set point variable mCO2, the random process
ly pointed out in [33]. distortions will not significantly influence the process
The conclusion is, that when the objective is batch-to- reproducibility.
batch reproducibility, the specific growth rate μ cannot be Now, where we know which variable is best suited for
recommended as the direct control variable. process supervision and control we must look for the main
It is important to stress that the regulation of the sub- reasons that restrict the information content of its signals
strate concentration is no alternative for μ-control. The in real cultivation runs. Their information content is best
main reasons are: (i) as for a high batch-to-batch repro- characterized by the signal-to-noise ratio SNR. In prac-
ducibility the specific growth rates must be lower than tice, the biggest impact on SNR has reducing the noise.
μmax as shown above, the substrate concentrations S in
this case should be considerably below 0.1 g/kg; this is 4.2 Improving SNR for the key measurement
lower than the normal working range of the currently variables
available online measurement devices; and (ii) substrate
measurements are local measurements that, particularly The higher the signal-to-noise ratio of the measurements
in larger reactors, do not reflect a representative value. used by the controller to determine the deviations from its
Hence a substantial bias appears that depends of the mix- set point, the more precisely the controller can react.

Journal
Improving SNR is a prominent engineering task. The pre- the controlled variables. A typical example for pH is given
ciseness of the carbon dioxide volume fraction which is in Fig. 4.
directly measured in the vent-line of bioreactors and from In order to cope with this problem, the controller
which the CPR signal is derived, can considerably be parameters need to be adapted to changes in the dynam-
influenced by the quality of the pH control in the culture. ics. One possibility is to automatically tune the most
As the dependency of the specific growth rate from important PID-controller parameter only, the controller
pH is usually quite broad with a rather flat maximum (e.g. gain KP, using a variable, called scheduling variable, that
[43]), pH control was traditionally given not much atten- correlates well with changes in the process dynamics.
tion. However, the carbon dioxide transfer rate CTR quite This technique is referred to as gain scheduling [9]. It is
sensitively depends on the pH. Also, many cell properties not necessary to adjust KP to its optimal value all time, a
in particular the metabolism can change profoundly over simple shift following the gross changes of the process is
a narrow pH range (e.g. [43]). Moreover, the pH drastical- sufficient for this purpose.
ly changes some physicochemical properties. When we look for the envelope of the pH-fluctuations
We take pH control as an example to illustrate the resulting from a conventional PID controller, a high simi-
measures necessary to reduce process variances. The larity to the CPR or even the substrate feed rate patterns
same applies for temperature and pressure control as well become apparent. In a large number of validation runs
as particularly for pO2-control which is even more compli- with many different cultivations it was shown, that
cated to handle as there are more sources of disturbances increasing the controller gain proportional to the current
influencing its controllers. pH-controllers are installed at CPR or substrate feed rate improves the quality of the
practically all fermentation processes. Usually the fer- feedback control system significantly. Evidence for this
menter manufacturers deliver the controllers together finding can be found in [44] and [45], which was based on
with their fermenter control systems. By far most indus- data sets from more than 50 microbial and animal cell cul-
trially used controllers are PID controllers [25]. This also tivations. This is also shown in Fig. 4. Hence, while it is
applies to biotechnology. In currently used process often claimed that it might be difficult to find a proper
automation systems most controllers work in such a way. scheduling variable in industrial nonlinear gain schedul-
ing controllers, this point does not make problems in fed-
4.2.1 Gain scheduling a powerful concept batch fermentations.
for feedback control in bioreactors The most important result with respect to process
Although it is well known that the controller parameters quality control in protein formation processes, however,
depend on the dynamic characteristics of the process to resulting from this improved pH control (pO2 control
be controlled, in practice their parameters are kept con- works similar (cf. [44, 45]) is that the variability in the main
stant. This leads to variations in the noise on the values of supervision and control variables is substantially reduced
Figure 4. (A) Comparison of typical pH-records across E. coli cultures where the pH control was performed with a conventional PID-controller (black) and
a gain scheduled controller (red). (B) Comparison of the OUR and CPR signals obtained with both controller versions (Data from [45] reprinted with per-
mission).

Journal
with such improved pH and pO2 controllers as can imme- cell properties themselves, primarily the product and the
diately be seen in Fig. 4. Hence, although it might seem biomass profiles, which determine the process perform-
to be esoteric to keep the pH fluctuation below 0.02, its ance. In literature many much different controller variants
effect on the variability of the key process variables is con- have been discussed. Here a serious advice is to start
vincing. Note that fluctuations of CPR and OUR are prac- with simple versions and to increase the complexity of the
tically leading to proportional fluctuations in the estimat- controllers only when the simpler ones fail.
ed specific growth rate and biomass profiles.
The variances in the measurement signals can be 5.1 PID-Variants of feedback controllers
dropped even more, when cumulative signals of OUR and
CPR are taken during online process supervision. The 5.1.1 Gain scheduling for feedback control
cumulation process itself can be considered a powerful As already shown for pH control, standard PID controllers
noise filter. For estimations of biomass and product con- are insufficient for controlling μ, X, or mCO2 with high
centrations such cumulation is straightforward as these quality since the dynamics of the system is changing with
quantities are cumulative quantities as such. time and with it the optimal PID controller parameters.
Hence, PID controllers can only be used properly when
4.2.2 Constructive aspects of the bioreactor mechanisms are employed that adapt their parameters to
influencing process control quality the changing process dynamics.
With respect to controlling the environmental conditions The first question, whether or not one can employ the
of the cells in a bioreactor, the basic controllers for T, p, same techniques, as already discussed during pH control
pH, pO2, are not the only points of attack. The bioreactor’s can be answered positively. Indeed, gain scheduling con-
transport properties have a non-negligible influence on trol can be used for controlling the mass mCO2 of CO2 pro-
the controller performances. The mixing performance, as duced and its set point profile already contains the infor-
primarily characterized by the mixing time, has a large mation about the changes in the dynamics of the process.
influence on the response time on controller actions. Figure 5 presents typical experimental results of feed-
Every delay in the response times to controller actions back-controlled E. coli cultivations, where mCO2 was tak-
constrain the controller gains [12]. The consequence is en as the controlled variable. Here the trajectories of four
that where these time delays restrict the controller per- cultivations controlled along the same set point profile
formances too much, the mixing time must be reduced by showed a perfect batch-to-batch reproducibility. The cor-
increasing the stirrer speed or by impeller retrofitting. responding profiles of the action variable, the feed rate F
As is well-known to biochemical engineers, mixing shows that the controller does not need to intervene into
times are measured in stimulus-response experiments. the cultivations with heavy actions.
The measurement value critically depends on the places In animal cell systems it proved to be better to control
where the stimulus was made and where the responses the cultures along the mass of oxygen consumed, while in
are measured. Transferred to pH control that means that microbial cultures control along the mass of CO2-pro-
the response times on controller actions depends on the duced leads to better results.
places where the reactant was added and where the pH Since the cultivations are rather deterministic sys-
is measured. Experimental investigations showed that tems, the important point to recognize is that when one
the differences in the mixing or response time can be key variable is kept tightly under control, i.e. develop at a
more than 100% when different places in the reactor are rather low variability, the other follow.
chosen [46]. The shortest times are obtained when the
base is added close to an impeller in the middle of the 5.1.2 Feedforward-feedback control increases
reactor. Even better is using more than a single place. the control quality
It should be emphasized that it often pays to spend An additional basic idea of process control is to augment
more efforts on equipment and not too much on biological a feedback controller with a feedforward control compo-
aspects of the cultivation processes in order to reduce the nent. The intention is not stressing the controller with
process variability. A good biochemical engineer avoids foreseeable changes in the set points so that it only needs
one-dimensional thinking and tries to combine the many to compensate randomly appearing deviations from the
aspects of engineering involved in a well-balanced way. desired path. This is particularly important for changes in
the control profiles of fed-batch processes where the
action variable is changing across nearly the entire culti-
5 Advanced control of fermentation processes vation time.
Positive effects of the augmentation to a feedback con-
Advanced control aims in keeping the key state variables troller with a feedforward element are particularly large
exactly on the reference trajectories of what is considered when distortions appear that can be recorded before they
the golden batch. Here the key variables are no longer the can be recognized by the controller. An important exam-
variables that control the environment of the cells but the ple for cultivation experiments is the addition of antifoam

Journal
Figure 5. (A)–(D) Data from four cultivation runs along the same mCO2 profile allowing to judge the batch-to-batch reproducibility of feedback-controlled
E. coli fermentations. (E)–(F) Six mO2-controlled CHO cultivations, (E) measured mO2-profiles, (F) measured viable cell count profiles.
agents which mainly affects the changes in oxygen trans- 5.2 Model-supported control uses predictions
fer rates. When the foam detector gives a signal to the of the process behavior
pump for antifoam agents, it can at the same time signal
the pO2 controller that a big distortion is to be expected in More sophisticated controllers perform even better than
the immediate future. A pO2 controller with feedforward adaptive PID controllers. In control engineering a favorite
extension can then reduce its actions for a short period in controller is the model predictive controller which does not
order to avoid unnecessary heavy responses. only react on deviations from the desired profile of the con-
In feedforward-feedback versions of mCO2-controllers, trol variable that already happened but makes predictions
the feedforward part can use the feed rate profile Fff(t) across some finite time horizon into the future, using a
from the reference cultivation while the feedback compo- dynamic process model which is solved across the time
nent is scheduled with the mCO2-profile. Fff(t) can also be horizon with the current state as initial conditions and with
estimated online from a biomass estimate and a desired some feasible choices for the manipulatable variables.
μset(t) profile. With the feedforward concept one tries to avoid
unnecessary reactions of a feedback controller, i.e. fore-
5.1.3 Cascading control responds faster on disturbances seeable or already planned changes in its actions. Model
Where there is a concern about the speed with which a predictive controllers go a step further and make use of a
mCO2- or a mO2-controller reacts upon deviations from its dynamic process model in order to already detect devel-
set point, additional information can be used taking the opments of the process’ state towards significant devia-
corresponding total rates tCPR or the tOUR signals as tions and try to avoid them by timely reactions [48].
additional sources of information. This can be done by a On the other hand, the development of model predic-
sequence of two PID controllers where the first takes the tive control algorithms requires deeper understanding of
mass pattern, here mO2, and the second its derivative, the the bioprocesses. Their implementation needs some
total oxygen uptake rate tOUR as the controlled variable. additional efforts to achieve sufficient robustness [49].
Such a construction was successfully applied to a CHO- Typical results of a model predictive control applica-
cell culture producing EPO [47]. tion for CHO cells producing EPO is shown in Fig. 6 [50].
Here, again, the culture is controlled along a mO2-profile.

Journal
Figure 6. Results of a model predictive control application: Product and

cell records for a mO2-controlled cell culture of CHO cells producing EPO
(reprinted with permission [50]) for four cultivation runs. (A) Controlled
variable mO2. (B) Viable cell count. (C) Product (EPO) mass.
Although the variability of the process variables drops open-loop-controlled fed-batch phase, where an expo-
with such model-aided controllers, they are unlikely to be nential feed is used that corresponds to a specific growth
applied at manufacturing cultivations in the near future, rate μset which is slightly smaller than μmax as explained
since the expense of designing, installing and maintain- in the open loop section above. Then, when more biomass
ing them is much larger as for the adaptive PID variants has been accumulated the signal-to-noise ratio of the cor-
discussed in the previous chapter. The improvement in responding measurement variables is usually good
performance is not so big that the expense will pay. It enough to switch over to feedback control.
would be fine if the simple and elegant PID variants would
be installed in production fermenters to a significant
amount. 6 Concluding remarks
An important point from the industrial manufacturing
perspective is that the monitoring and control procedures It turned out that the performance of fermentation
must be implementable into the currently used profes- processes in terms of the productivity and the batch-to-
sional process automation systems offered by companies batch reproducibility can largely be improved by relative
such as, Emerson, Yokogava, Honeywell, Siemens etc. For simple and easily understandable control procedures.
most of the tools we mentioned in this paper appropriate These techniques only make use of basic knowledge
modules were already designed, implemented and tested. about fermentation processes, expressed in simple
Changes in the process hardware are seldom required, in numerically exploitable process models. We strongly rec-
most production environments for recombinant proteins ommend indirectly controlling the specific biomass
the off-gas analyzers are already available since many growth rate of microbial as well as animal cell cultures.
years. Best experiences were made with controlling directly
A final remark with respect to feedback control at fer- measurable variables such as the mass of oxygen con-
mentation processes is necessary with respect to open sumed or the mass of carbon dioxide produced, both are
loop control. Feedback control needs timely, accurate and closely related to the biomass and specific growth rate
precise data about the current process state. In the first profiles.
few hours (about 0–4 h) of the cultivation time, the indi- Essential not only to a successful process control but
rect measurements of the state variables are usually quite also to its supervision is the precision and the accuracy by
noisy. This restricts the performance of all feedback con- which the state of the cultivation process can be meas-
trollers. Thus it is advisable to start all cultivations with an ured online. The most direct and informative way is using

Journal
off-gas measurements for this purpose. For measuring the

oxygen uptake and the carbon dioxide production rates, Andreas Lübbert is professor emeritus
several industrially well-established devices are available. of bioprocess engineering at the
With artificial neural networks as well as with later devel- Department of Biochemistry and
opments from the subject of machine-learning, e.g. the Biotechnology at the Martin-Luther-
support vector regression, well-performing estimates of University Halle-Wittenberg, Germany.
the state variables are possible. He obtained his PhD in Industrial
Quality by design is an important issue for bioprocess Chemistry at the Leibnitz University of
engineers. The design of process operational procedures Hannover, where he later received the
that are robust to the usually appearing random distor- venia legendi for the subject “Techni-
tions is an important concept. We presented an example sche Chemie”. His research interests include bioprocess optimization,
where the variety in the inoculum size is compensated for supervision, and control as well as bioreactor optimization with
by running the fermentation right from the beginning in respect to fluid dynamics and construction. On these subjects he
the fed-batch mode. The price that must be paid is a cooperated with numerous industrial companies in many projects and
slightly reduced productivity. as supervisor. For nearly two decades he served as chairman of the
In most practical cases, the straightforward way of EFB/ESBES working group of Bioreactor Performance.
reducing the variability is feedback control. Here it is usu-
ally not necessary to accept considerable reductions in
the productivity. When the well-performing measurement
techniques discussed are available it is quite easy to
install simple feedback control procedures that allow to Rimvydas Simutis is professor of
compensate for deviations of the process’ state from its automation and control system engi-
target. They are well-suited for keeping both productivity neering at Department of Automation,
and process reproducibility at high levels. Several exam- Kaunas University of Technology,
ples were given for microbial as well as animal cell culti- Lithuania. He obtained his PhD in Con-
vations. trol Engineering at Kaunas University of
It is often better to put more emphasis on reactor Technology and worked several years
design and less on sophisticated controller design. E.g. as Alexander von Humboldt grant hold-
the addition ports for the feed or correcting agents to the er and post-doctoral researcher at Leib-
fermenters should be designed carefully in order to avoid niz Universität Hannover. His research interest includes advanced
excessive time delays which directly affect the controller modeling, optimization and control system techniques with applica-
parameters. The same applies for the overall mixing per- tion in Bioengineering and Technical systems. He participated in vari-
formance of the reactors. ous research projects with biopharmaceutical and automation compa-
The advantage of feedback-feedforward control tech- nies as senior researcher or supervisor.
niques were not adequately recognized in the past. Sup-
porting a feedback controller by a feedforward component
essentially means telling the controller our knowledge reference feeding profile turned out to be more reliable
about changes in the set points to come. Then the con- strategy. However, usually after some time (about 4 h in
troller has to compensate only for the randomly apparent microbial cultures) the measurement data provide
deviations. enough information to switch over to feedback control
In control engineering science a large amount of what then leads to a much better performance. The
sophisticated control techniques have been developed. biggest mistake made in the past with respect to biore-
The most sophisticated controllers use dynamic models actor control was to start feedback controllers too early
that are able to predict the process behavior across a time when the measurements did not provide values accurate
horizon of several hours. These techniques were shown to enough to decide on controller actions.
also be applicable to fermentation processes, however, at
a quite high price of requiring control engineers that are
able to construct these controllers and to maintain them Rimvydas Simutis thanks the Research Council of Lithua-
for the process under consideration. In a time where even nia for financial supportthrough No. MIP-056/2013.
comparably simple PID-control procedures are scarcely
apparent in industrial practice, their application current- The authors declare no financial or commercial conflicts of
ly seems quite unrealistic. interest.
In the very beginning of cultivation processes the
accuracy of indirect measurements are usually not accu-
rate enough to smoothly follow the deviations from the set
point profiles. In this period open loop control along the

Journal
[25] Auger, F., Hilairet, M., Guerrero, J. M., Monmasson, E. et al., Indus-
7 References trial applications of the Kalman filter. IEEE Trans. Ind. Electron. 2013,
60, 5458–5471.
[1] US Food and Drug Administration, Guidance for industry PAT – [26] Wan, E., Van der Merwe R., The unscented Kalman filter, Chap. 7, in:
a framework for innovative pharmaceutical development, manufac- Haykin, S. (Ed.), Kalman Filtering and Neural Networks, Wiley 2001.
turing and quality assurance, 2004. [27] Gustafsson, F., Hendeby, G., Some relations between extended and
[2] Dingermann, T., Zündorf, I., Biosimilars: Eine neue Gruppe rekom- unscented Kalman filters, IEEE Trans. Signal Process. 2012, 60,
binanter Arzneimittel. Pharm. unserer Zeit 2012, 41, 446–452. 545–555.
[3] Taguchi, G., Chowdhury, S., Wu, Y., Taguchi’s Quality Engineering [28] Chai, Q., Furenes, B., Lie, B., Comparison of state estimation tech-
Handbook, Wiley, New York 2005. niques, applied to a biological wastewater treatment process. 10th
[4] Rao, R. S., Kumar, C. G., Prakasham, R. S., Hobbs, P. J., The Taguchi International IFAC Symposium on Computer Applications in
methodology as a statistical tool for biotechnological applications: Biotechnology 2007, 1, 353–358.
A critical appraisal. Biotechnol. J. 2008, 3, 510–523. [29] Zhu, X., Feng, E., Joint estimation in batch culture by using
[5] Schuler, M. M., Marison, I. W., Real-time monitoring and control of unscented Kalman filter. Biotechnol. Bioprocess Eng. 2012, 6, 1238–
microbial bioprocesses with focus on the specific growth rate: Cur- 1243.
rent state and perspectives. Appl. Microbiol. Biotechnol. 2012, 94, [30] Gustafsson, F., Gunnarsson, F., Bergman, N., Forssell, J. et al., Parti-
1469–1482. cle filters for positioning, navigation and tracking. IEEE Trans. Sig-
[6] Sauer, T., Preuss, K. H., Verbesserte Prozessführung, Process Phar- nal Process. 2002, 50, 425–437.
maTEC 2005, 5, 50–52. [31] Rawlings, J. B., Bakshi, B. R., Particle filtering and moving horizon
[7] Luyben, W. L., Process modeling, simulation and control for chemi- estimation. Comput. Chem. Eng. 2006, 30, 1529–1541.
cal engineers. McGraw-Hill, New York 1990. [32] Simutis, R., Galvanauskas, V., Levišauskas, D., Repšyte, J., Grincas,
[8] Bechhoefer, J., Feedback for physicists: A tutorial essay on control. V., State estimation of a biotechnological process using extended
Rev. Mod. Phys. 2005, 77, 783–834. Kalman filter and Particle filter. Int. J. Biol.. Food Vet. Agric. Eng.
[9] Åström K. J., Hägglund, T., Advanced PID Control, ISA: The Instru- 2014, 8, 851–855.
mentation, Systems and Automation Society, Research Triangle [33] Jenzsch, M., Gnoth, S., Beck, M., Kleinschmidt, M. et al., Open-loop
Park 2005, NC 27709. control of the biomass concentration within the growth phase of
[10] Villadsen, J., Nielsen, J., Lidén, G., Bioreaction Engineering Princi- recombinant protein production processes. J. Biotechnol. 2006, 127,
ples, 3rd Ed., Springer, Berlin 2011. 84–94.
[11] Schaepe, S., Kuprijanov, A., Aehle, M., Simutis, R., Lübbert, A., Sim- [34] Dochain D., Bioprocess Control, John Wiley&Sons, London 2008.
ple control of fed-batch processes for recombinant protein produc- [35] Herwig, C., Marison, I., von Stockar, U., On-line stoichiometry and
tion with E. coli. Biotechnol. Lett. 2011, 33, 1781–1788. identification of metabolic state under dynamic process conditions.
[12] Åström, K. J., Murray, R. M., Feedback systems, an introduction for Biotechnol. Bioeng. 2001, 75, 345–354.
scientists and engineers, Princeton University Press, 2009. [36] Stephanopoulos, G., Aristidou, A., Nielsen, J., Material balances and
[13] Luedeking, R., Piret, E. L., A kinetic study of the lactic acid fermen- data consistency, Metabolic Engineering, Academic Press, New
tation batch process at controlled pH, J. Biochem. Microbiol. Tech- York 1998.
nol. Eng. 1959, 1, 393–412. [37] McMillan, G. K., Benton, T., Zhang, Y., Boudreau, M. A., PAT tools for
[14] Mandenius, C. F., Recent developments in the monitoring, modeling accelerated process development and improvement. Bioprocess Int.
and control of biological production systems. Bioprocess Biosyst. 2008, 6, 34–42.
Eng. 2004, 26, 347–351. [38] Gnoth, S., Simutis, R., Lübbert, A., Fermentation process supervi-
[15] Sundström, H., Enfors, S. O., Software sensors for fermentation sion and strategies for fail-safe operation – a practical approach.
processes. Bioproc. Biosyst. Eng. 2008, 31, 145–152. Eng. Life Sci. 2011, 11, 94–106.
[16] Hocalar, A., Türker, M., Karakuzu, C., Yuezgec, U., Comparison of dif- [39] Lee, J., Youn-Hee, C., Shin-Kwon, K., Hyung-Hwan, P., Ik-Boo, K.,
ferent estimation techniques for biomass concentration in large Production of human leukocyte interferon in Escherichia coli by con-
scale yeast fermentation. ISA Trans. 2011, 50, 303–314. trol of growth rate in fed-batch fermentation. Biotechnol. Lett. 1989,
[17] Luttmann, R., Bracewell, D. G., Cornelissen, G., Gernaey, K. et al., 11, 695–698.
Soft sensors in bioprocessing: A status report and recommenda- [40] Yoon, S. K., Kang, W. K., Park, T. H., Fed-batch operation of recom-
tions. Biotechnol. J. 2012, 7, 1040–1048. binant E. coli containing Trp promoter with controlled specific
[18] Jenzsch, M., Simutis, R., Eisbrenner, G., Stückrath, I., Lübbert, A., growth rate. Biotechnol. Bioeng. 1994, 43, 995–999.
Estimation of biomass concentrations in fermentation processes for [41] Soons, Z. I. T. A., Voogt, J. A., van Straten, G., van Boxtel, A. J. B.,
recombinant protein production. Bioproc. Biosyst. Eng. 2006, 29, Constant specific growth rate in fed-batch cultivation of Bordetella
19–27. pertussis using adaptive control. J. Biotechnol. 2006, 125, 252–268.
[19] Schölkopf, B., Smola, A. J., Learning with Kernels, MIT Press, Cam- [42] Jenzsch, M., Gnoth, S., Kleinschmidt, M., Simutis, R., Lübbert, A.,
bridge, MA, 2002. Improving the batch-to-batch reproducibility of microbial cultures
[20] Suykens, J. A., Gestel, T, Brabanter, J., Moor, B., Vanderwalle, J., during recombinant protein production by regulation of the total car-
Least Squares Support Vector Machines. World Scientific, Singapore bon dioxide production. J. Biotechnol. 2007, 128, 858–867.
2002. [43] Pirt, S. J., Principles of Microbe and Cell Cultivation, Blackwell Sci-
[21] Tipping, M. E., Sparse Bayesian learning and the relevance vector entific, Oxford 1975.
machine. J. Mach. Learn. Res. 2001, 1, 211–244. [44] Gnoth, S, Kuprijanov, A., Simutis, R., Lübbert, A., Simple adaptive
[22] Stephanopoulos, G., San, K. Y., Studies on on-line bioreactor identi- pH control in bioreactors using gain-scheduling methods. Appl.
fication: Theory. Biotechnol. Bioeng. 1984, 26, 1176–1188. Microbiol. Biotechnol. 2010, 85, 955–964.
[23] Bogaerts, P., Van de Wouwer, A., Software sensors for bioprocesses. [45] Kuprijanov, A., Gnoth, S., Simutis, R., Lübbert, A., Advanced control
ISA Trans. 2003, 42, 547–558. of dissolved oxygen concentration in fed batch cultures during
[24] Venkateswarlu, C., Advances monitoring and state estimation of recombinant protein production. Appl. Microbiol. Biotechnol 2009,
bioreactors. J. Sci. Ind. Res. 2004, 63, 491–498. 82, 221–229.

Journal
[46] Sieblist, C., Jenzsch, M., Pohlscheidt, M., Lübbert, A., Insights into [50] Aehle, M., Bork, K., Schaepe, S., Kuprijanov, A. et al., Increasing
large scale cell culture reactors: I. Liquid mixing and oxygen supply. batch-to-batch reproducibility of CHO-cell cultures using a model
Biotechnol. J. 2011, 6, 1532–1546. predictive control approach. Cytotechnology 2012, 64, 623–634.
[47] Kuprijanov, A., Schaepe, S., Aehle, M., Simutis, R., Lübbert, A., [51] Schaepe S, Jenzsch M, Kuprijanov A, Simutis R, Lübbert A., Batch-
Improving cultivation processes for recombinant protein produc- to-batch reproducibility of fermentation processes by robust opera-
tion. Bioprocess Biosyst. Eng. 2012, 35, 333–340. tional design and control. Pharm. Bioprocess. 2013, 1, 297–307.
[48] Camacho, E. F., Bordons, C., Model Predictive Control in the Process
Industry, Springer, London 1995.
[49] Santos, L. O., Dewasme, L., Coutinho, D., Van de Wouwer, A., Non-
linear model predictive control of fed-batch cultures of micro-organ-
isms exhibiting overflow metabolism: Assessment and robustness,
Comput. Chem. Eng. 2012, 39, 143–151.

ISSN 1860-6768 · BJIOAM 10 (8) 1091–1326 (2015) · Vol. 10 · August 2015
Systems & Synthetic Biology ·

Nanobiotech · Medicine
8/2015
Bioengineering
Industrial biotechnology
Bioprocess development
Cover illustration
Special Issue: ESBES. This issue of BTJ highlights a selection of articles presented at the
10th European Symposium on Biochemical Engineering Sciences (ESBES), which was
held in Lille, France, September 7–10, 2014. The issue is edited by Guilherme Ferreira
and Philippe Jacques and includes articles on metabolic engineering, protein expression
www.biotechnology-journal.com
and bioprocess development. The cover shows the Grande Place in Lille. © bbsferrari –
Fotolia.com
Biotechnology Journal – list of articles published in the August 2015 issue.
Editorial: Bioengineering for a better quality of life Review

Guilherme Ferreira and Philippe Jacques Towards plant protein refinery: Review on protein extraction
http://dx.doi.org/10.1002/biot.201500419 using alkali and potential enzymatic assistance
Yessie W. Sari, Wilhelmus J. Mulder, Johan P. M. Sanders
Forum and Marieke E. Bruins
Process analytical technologies in food industry – challenges http://dx.doi.org/10.1002/biot.201400569
and benefits: A status report and recommendations
Bernd Hitzmann, Ralph Hauselmann, Andreas Niemoeller, Review
Daryoush Sangi, Jens Traenkle and Jarka Glassey Partitioning in aqueous two-phase systems:
http://dx.doi.org/10.1002/biot.201400773 Analysis of strengths, weaknesses, opportunities and threats
Ruben R. G. Soares, Ana M. Azevedo, James M. Van Alstine
Commentary and M. Raquel Aires-Barros
A new use for existing technology – continuous precipitation http://dx.doi.org/10.1002/biot.201400532
for purification of recombination proteins
Veena Warikoo and Rahul Godawat Review
http://dx.doi.org/10.1002/biot.201400840 Engineering microbial cell factories: Metabolic engineering
of Corynebacterium glutamicum with a focus on non-natural
Review products
Clinical-scale purification of pluripotent stem cell derivatives Sabine A. E. Heider and Volker F. Wendisch
for cell-based therapies http://dx.doi.org/10.1002/biot.201400590
Gonçalo M. C. Rodrigues, Carlos A. V. Rodrigues,
Tiago G. Fernandes, Maria Margarida Diogo Review
and Joaquim M. S. Cabral Current progress in high cell density yeast bioprocesses
http://dx.doi.org/10.1002/biot.201400535 for bioethanol production
Johan O. Westman and Carl Johan Franzén
Review http://dx.doi.org/10.1002/biot.201400581
Bioreactor control improves bioprocess performance
Rimvydas Simutis and Andreas Lübbert Research article
http://dx.doi.org/10.1002/biot.201500016 Continuous precipitation of IgG from CHO cell culture
supernatant in a tubular reactor
Mini-Review Nikolaus Hammerschmidt, Beate Hintersteiner, Nico Lingg
Photobioreactors with internal illumination – and Alois Jungbauer
A survey and comparison http://dx.doi.org/10.1002/biot.201400608
Martin Heining and Rainer Buchholz
http://dx.doi.org/10.1002/biot.201400572
© 2015 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim www.biotechnology-journal.com

Research Article Research Article
Fermentation broth components influence droplet coalescence Maximizing the utilization of Laminaria japonica as biomass
and hinder advanced biofuel recovery during fermentation via improvement of alginate lyase activity in a two-phase
Arjan S. Heeres, Karin Schroën, Joseph J. Heijnen, fermentation system
Luuk A. M. van der Wielen and Maria C. Cuellar Yuri Oh, Xu Xu, Ji Young Kim and Jong Moon Park
http://dx.doi.org/10.1002/biot.201400570 http://dx.doi.org/10.1002/biot.201400860

Modeling leucine’s metabolic pathway and knockout prediction Improving recombinant protein production in the
improving the production of surfactin, a biosurfactant from Chlamydomonas reinhardtii chloroplast using vivid Verde
Bacillus subtilis Fluorescent Protein as a reporter
François Coutte, Joachim Niehren, Debarun Dhali, Stephanie Braun-Galleani, Frank Baganz and Saul Purton
Mathias John, Cristian Versari and Philippe Jacques http://dx.doi.org/10.1002/biot.201400566
Research Article
Research Article Regulation of the NADH pool and NADH/NADPH ratio
A xeno-free microcarrier-based stirred culture system for the redistributes acetoin and 2,3-butanediol proportion in Bacillus
scalable expansion of human mesenchymal stem/stromal cells subtilis
isolated from bone marrow and adipose tissue Teng Bao, Xian Zhang, Xiaojing Zhao, Zhiming Rao,
Joana G. Carmelo, Ana Fernandes-Platzgummer, Taowei Yang and Shangtian Yang
Maria Margarida Diogo, Cláudia Lobato da Silva http://dx.doi.org/10.1002/biot.201400577
and Joaquim M. S. Cabral
http://dx.doi.org/10.1002/biot.201400586 Research Article
Engineering a branch of the UDP-precursor biosynthesis
Research Article pathway enhances the production of capsular polysaccharide
Dynamic flux balancing elucidates NAD(P)H production in Escherichia coli O5:K4:H4
as limiting response to furfural inhibition in Saccharomyces Donatella Cimini, Elisabetta Carlino, Alfonso Giovane,
cerevisiae Ottavia Argenzio, Ileana Dello Iacono, Mario De Rosa
Pornkamol Unrean and Carl J. Franzen and Chiara Schiraldi
http://dx.doi.org/10.1002/biot.201400833 http://dx.doi.org/10.1002/biot.201400602

Newly designed and validated impedance spectroscopy setup Phenotypic variability in bioprocessing conditions can be
in microtiter plates successfully monitors viable biomass online tracked on the basis of on-line flow cytometry and fits to a
Bettina Luchterhand, Jannis Nolten, Sadik Hafizovic, scaling law
Tino Schlepütz, Sandra Janine Wewetzer, Elke Pach, Jonathan Baert, Romain Kinet, Alison Brognaux,
Kristina Meier, Georg Wandrey and Jochen Büchs Anissa Delepierre, Samuel Telek, Søren J. Sørensen,
http://dx.doi.org/10.1002/biot.201400534 Leise Riber, Patrick Fickers and Frank Delvigne
Research Article
A versatile, non genetically modified organism (GMO)-based
strategy for controlling low-producer mutants in Bordetella
pertussis cultures using antigenic modulation
Philippe Goffin, Thomas Slock, Vincent Smessaert,
Philippe De Rop and Philippe Dehottay
© 2015 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim www.biotechnology-journal.com

Simutis 2015

Uploaded by

Copyright:

Available Formats

Simutis 2015

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Simutis 2015

Uploaded by

Copyright:

Available Formats

Biotechnology Biotechnol. J. 2015, 10, 1115–1130 DOI 10.1002/biot.

Bioreactor control improves bioprocess performance

Rimvydas Simutis1 and Andreas Lübbert2

1 Introduction perspective, it is also necessary to keep the variations in

© 2015 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim 1115

1116 © 2015 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim

© 2015 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim 1117

1118 © 2015 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim

Figure 1. Typical example showing the

© 2015 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim 1119

Figure 2. UKF estimate of X, P, and μ,

1120 © 2015 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim

© 2015 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim 1121

1122 © 2015 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim

© 2015 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim 1123

1124 © 2015 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim

© 2015 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim 1125

1126 © 2015 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim

Figure 6. Results of a model predictive control application: Product and

© 2015 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim 1127

off-gas measurements for this purpose. For measuring the

1128 © 2015 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim

© 2015 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim 1129

1130 © 2015 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim

Systems & Synthetic Biology ·

Biotechnology Journal – list of articles published in the August 2015 issue.

Editorial: Bioengineering for a better quality of life Review

© 2015 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim www.biotechnology-journal.com

Research Article Research Article

Research Article Research Article

© 2015 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim www.biotechnology-journal.com

You might also like