You are on page 1of 13



Bioreactors are one of the most important parts of technological production
system. The choosing proper reactor and setting optimal parameters of its work
are crucial for optimum results during production. The complexity of the
bioreactor lies in the number of input parameters that exist in these cases when
compared to their simpler chemical counterparts. Conventional reaction
engineering techniques that are basically designed
for chemical reactor design therefore cannot handle such enormous input
multiplicity both at the level of reactor designing as well as in the process control
designing. Therefore there exists a need to take help of mathematical modeling
in the design of bioreactors. With various software now available for
mathematical modeling it has become quite easy to apply these modeling
techniques with ease even though the designer has no much previous familiarity
with those modeling techniques.

MATLAB® is a high-performance language for technical computing. It integrates

computation, visualization, and programming in an easy-to-use environment
where problems and solutions are expressed in familiar mathematical notation.
Typical uses include Math and computation Algorithm development Data
acquisition Modeling, simulation, and prototyping Data analysis, exploration, and
visualization Scientific and engineering graphics Application development,
including graphical user interface building .MATLAB is an interactive system
whose basic data element is an array that does not require dimensioning. This
allows you to solve many technical computing problems, especially those with
matrix and vector formulations, in a fraction of the time it would take to write a
program in a scalar non-interactive language such as C or FORTRAN.
The name MATLAB stands for matrix laboratory. MATLAB was originally written
to provide easy access to matrix software developed by the LINPACK and
EISPACK projects. Today, MATLAB engines incorporate the LAPACK and BLAS
libraries, embedding the state of the art in software for matrix computation.
MATLAB has evolved over a period of years with input from many users. In
university environments, it is the standard instructional tool for introductory and
advanced courses in mathematics, engineering, and science. In industry,
MATLAB is the tool of choice for high-productivity research, development, and
MATLAB features a family of add-on application-specific solutions called
toolboxes. Very important to most users of MATLAB, toolboxes allow you to learn
and apply specialized technology. Toolboxes are comprehensive collections of
MATLAB functions (M-files) that extend the MATLAB environment to solve
particular classes of problems. Areas in which toolboxes are available include
signal processing, control systems, neural networks, fuzzy logic, wavelets,
simulation, and many others.


Neural networks

Neural networks are composed of simple elements operating in parallel. These

elements are inspired by biological nervous systems. As in nature, the network
function is determined largely by the connections between elements. We can
train a neural network to perform a particular function by adjusting the values of
the connections (weights) between elements.
Commonly neural networks are adjusted, or trained, so that a particular input
leads to a specific target output. Such a situation is shown below. There, the
network is adjusted, based on a comparison of the output and the target, until the
network output matches the target. Typically many such input/target pairs are
used, in this supervised learning, to train a network.
Batch training of a network proceeds by making weight and bias changes based
on an entire set (batch) of input vectors. Incremental training changes the
weights and biases of a network as needed after presentation of each individual
input vector. Incremental training is sometimes referred to as "on line" or
"adaptive" training.

Bench scale reactor control modeling using neural networks

The main advantage with neural networks as discussed earlier is the

development of the model using dry lab simulation data obtained from previous
models. One such example is as follows:

This is the basic design of the first neural network designed. It is the general
framework for long-range prediction for predictive control. (Z_ denotes the
signals z(_); z(_ + 1); z(_);Z 2 (U; ^ X; ^Y );u the manipulated input, y the output
measurements and x the system state. The predictive controller applies tentative
U-values and takes the ^XX and/or ^ Y values to optimize the process
performance. In this example, u _ vf (t); y _ (pH, conductivity), x _ ([A]; [B]; [C];

Neural-Network Estimator-Predictor
All networks use a sigmoidal activation function for the hidden layer, a linear
activation function for the input and the output layers, and an additional bias node
in the input and hidden layers. All neural-network blocks contain five sigmoidal
hidden nodes. This network complexity was chosen based on extensive
experience gained from simulations. Trialand- error design of the network
complexity, based on actual process runs, could lead to better networks, but
would require prohibitive experimental expense.
The prediction horizon was chosen to be 30 sampling instants (one minute) for
training all predictor networks. This choice seems to be a good compromise
between accurate
long-range prediction and reasonable training times. Further increase of the
training prediction horizon did not significantly improve the prediction accuracy,
but increased the training effort considerably. B. Feedforward Networks A
feedforward network is used to estimate the process state and output at sampling
instant (this is the “Estimator” block of Fig. 2). It takes as inputs the sum of the
moles of B fed up to the current sampling instant, the sampling instant, the
volume, and the current measurements of the pH and electrical conductivity. As
outputs, it yields the estimates at of the concentrations of A, B, C, and D. This
network has thus six input nodes (including the bias node), four output nodes,
and 54 weights.
Another feed forward network is then used to predict the future concentrations.
The concentration estimates of the first network are passed to this network,
which takes as further inputs the sampling instant for which the estimates were
computed, the sampling instant for which the long-range prediction has to be
made, , the sum of the feed up to , and the volume at time . As output, it yields
the estimates for of the concentrations of A, B, C, and D. This network has nine
input nodes, four output nodes, and 69 weights. The use of the current sampling
instant as an input into the first network, and into the second, is necessary in
order to provide the network with some temporal context within the batch.
Without this input, the networks can not learn a reasonable mapping. The
sampling instant is not directly related to the state, but facilitates proper
interpretation of the other input information.
External-Feedback Networks
Input Representation: Equations used to develop this model are the algebraic
and differential equations of the known part of the system. A straightforward
realization of these equations as a network requires the concentrations of A, B, C
and D,4 the current feed rate, and the volume as network inputs. Due to the
relatively high sampling rate, which is 2 s compared to the reaction time of 1800
s, the system can be simplified in the time-discrete description. Due to the
marginal dilution during a sampling interval, the dilution effect on the reaction can
be neglected, and the reaction and the dilution can be decoupled. Thus the
concentrations and the current feed rate per volume of reaction mixture are used
as inputs to the network, and the dilution due to the feed during the sampling
interval is computed only at the end of the sampling interval. This approach,
which is applied for all external-feedback networks, with and without other prior
information, avoids the complexity of an additional network input. The errors
introduced by not taking the dilution during each sampling interval into account
are negligible, and the model of the
volume itself (9) is accurate since no mixing nonidealities are present and the
amount of feed added is known accurately.

Design of control systems for CSTR using neural networks in MATLAB

Consider the case of a simple continuous stirred tank reactor as shown in the
figure below.

For the above reactor a neural network has been designed in MATLAB using the
neural network toolbox. The highlighted region is the neural network (NN)
predictive controller which receives inputs from the reference data as well as
from the plant output in order to create an online/ adaptive training for control of
the CSTR.
The results of such a neural network can be simulated using Simulink software in
the MATLAB itself. The resultant Simulink of the above network is shown below.

The simulink results are generally considered 99% accurate if the data given is
accurate enough. And hence these results are considered more or less
equivalent to experimental results. The simulink results can be generated by
giving different input data so as to find out the best possible input for optimal


Narmax model
Non-linear black box models are difficult to handle in general because the
spectrum of possible model descriptions is very wide. The area is quite diverse
and covers topics from mathematically approximation theory via narmax
representations of many non-linear systems require only a few terms.

Wavelet model
A wavelet is a waveform of effectively limited duration that has an average value
of zero. Wavelet analysis is a new and promising set of tools and techniques for
analyzing these signals. Wavelets have scale aspects and time aspects,
consequently every application has scale and time aspects. To clarify them we
try to untangle the aspects somewhat arbitrarily. For scale aspects, we present
one idea around the notion of local regularity. For time aspects, we present a list
of domains. When the decomposition is taken as a whole, the de-noising and
compression processes are center points. Wavelets concepts can be applied for,
One-Dimensional Continuous Wavelet Analysis
One-Dimensional Complex Continuous Wavelet Analysis
One-Dimensional Discrete Wavelet Analysis
Two-Dimensional Discrete Wavelet Analysis
One-Dimensional Discrete Stationary Wavelet Analysis
Two-Dimensional Discrete Stationary Wavelet Analysis
One-Dimensional Wavelet Regression Estimation
One-Dimensional Wavelet Density Estimation

Wavelet – naramax hybrid model

Local function expansion based model structures including the wavelet
decomposition techniques provide a powerful tool for representing non-linear
signals, even severely non-linear signals with discontinuities. Wavelet
decompositions outperform many approximation schemes and offer flexible
capability for approximating arbitrary functions. As a special form of the narmax
model, this hybrid model structure is referred to as the WANARMAX model.
This model has been applied in the case of ‘stochastic identification of bioreactor
process exhibiting input multiplicity’ [Rishi Amrit and Prabirkumar Saha]. The aim
of their work was to address the stochastic modeling issues related to bioreactor
process. They had used all the possible modeling techniques viz., block oriented
narmax model, bootstrap structure detection for narmax model and wavelet –
narmax model of which they found wavelet – narmax model to be the best of the
pick. Their work provides a lead to understanding input multiplicity in bioreactors
and the ways and means to solve the entire problem of simulation designs for
such complex reactors.


In Figure given above is the basic layout for the considered activated sludge
processis shown: from the secondary settler, the sludge is partially recirculated to
the bioreactor (Returned Activated Sludge, RAS) and partially wasted as excess
sludge (Waste Activated Sludge, WAS). The ASP is a biological process in which
microorganisms oxidize and mineralize organic matter. The microorganisms in
the activated sludge are mainly bacteria, which can be found also in the raw
wastewater incoming into the plant. The composition and the species depend not
only on the influent wastewater but also on the design and operation of the
wastewater treatment plant.
Given below is the basic mechanism along with the necessary equations that
have been used in the design of the activated sludge process.

The software currently in market that can be used for the simulation of the above
mentioned process are :
• WEST (Wastewater treatment plant Engines for Simulation and Training):
an interactive dynamic simulator. It is developed mainly at the
University of Gent, Belgium and current information about the software
can be found on
• SIMBA (Simulation programms f¨ur die Biologische Abwasserreinigung):
developed at the Institut f¨ur Automation und Kommonikation
(IFAK) in Germany. It can be considered a custom made version
of Simulink for wastewater treatment applications. A more extensively
compend about the simulator can be found on http://simba.
• EFOR is a stand-alone software package for the simulation of complete
wastewater treatment plant. It is developed mainly at the Danish
Technical University. The present progress can be found on http:

Using MATLAB for ASP modeling

the above case was modeled in MATLAB as shown below:

The simulation using simulink is as shown above. We can observe input
multiplicity in this case also. Earlier we had seen it while discussing the wavelet –
narmax model.
The results of simulink simulation are shown below:


The various modeling techniques discussed till now come to prove how easy it is
to model and simulate various bioreactors and their control systems using
various modeling & simulation techniques and also by using various modeling
software with MATLAB being the best suited in most of the cases. The wider
understanding of these mathematical models and modeling software would make
biotech engineers better equipped to handle the complex tasks of bioreactor
design and control of those bioreactors.


• B. Schenker and M. Agarwal, “Robust predictive control using

models,” in AIChE Annu. Meet. St. Louis, MO, Nov. 1993,
Paper 147j.

• D. E. Seborg, T. F. Edgar, and S. L. Shah, “Adaptive control strategies

for process control: A survey,” AIChE J, vol. 32, no. 6, pp. 881–913,

• Predictive Control of a Bench-Scale Chemical Reactor Based on Neural-

Network Models
Benedikt Schenker and Mukul Agarwal

• P. Terwiesch, D. Ravemark, B. Schenker, and D. W. T. Rippin,

process optimization under uncertainty: Theory and experiments,”
Computers Chem. Eng., vol. 22, pp. 201–213, 1998.

• K. Singh and J. Hahn. On the use of empirical Gramians for controllability

and observability analysis. In Proceedings of the American
Control Conference, June 8-10 2005. Portland, USA.

• M. Mulas and S. Skogestad. Control structure analysis for an activated

sludge process. In Proceedings of ICheap7, Chemical Engineer-
ing Transaction, pages 173–178, 2005.

• Process control and monitoring:

control and monitering.html

• Nonlinear adaptive control for bioreactors with unknown kinetics-

Ludovic Mailleret;, Olivier Bernard, Jean-Philippe Steyer

• Predictive Control of a Bench-Scale Chemical Reactor Based on Neural-

Network Models: Benedict Schenker and Mukul Agarwal

• Neural network based modeling and control of batch reactor:

M. Mujtaba, N. Aziz, and M. A. Hussain.

• Dynamic simulator for yeast fermentation bioreactor: Zoltan K. Nagy.