Development of an Explainable Model for a Gluconic Acid Bioreactor and Profit Maximization through Grey Wolf Optimizer Trained Artificial Neural Network Technique

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 10 Issue: 03 | Mar 2023 www.irjet.net p-ISSN: 2395-0072
© 2023, IRJET | Impact Factor value: 8.226 | ISO 9001:2008 Certified Journal | Page 465
Development of an Explainable Model for a Gluconic Acid Bioreactor
and Profit Maximization through Grey Wolf Optimizer Trained
Artificial Neural Network Technique
Sucharita Pal1, Sandip Kumar Lahiri1*
1Chemical Engineering Department, National Institute of Technology Durgapur, Durgapur – 713209
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - The current study focuses on building a model of
a laboratory scale bioreactor using a grey wolf optimizer
trained ANN approach, then optimizing it for profit
maximization. The glucose to gluconic acid bioprocess was
employed as a case study. The fermenter is a multiphase
enzymatic bioreactor and developing a viable first principle-
based model is difficult due to its complexity; on the other
hand, data-driven models lack explanability. Asaresult, inthis
study, a general methodology was developed in which a data-
driven technique such as the grey wolf optimizer trained ANN
technique was used as a modelling tool, and the model was
then post-processed to increase themodel'sexplainability. The
model was chosen for its ability to accurately describe the
underlying physics of the system. It was subjected to
optimization after the establishment of an acceptable model.
The goal of this study was to maximize gluconic acid yield,
which has a substantial impact on the process' profitability.
Using an evolutionary algorithm on the created model, an
ideal solution was determined.
Key Words: Grey wolf optimizer trained ANN technique;
Bioprocess; Modelling; Optimization; Genetic algorithm
1.INTRODUCTION
The business environment has altered drastically in the
previous decade as a result of severe worldwide
competition. Because of globalisation, chemical process
industries are seeing their profit margins erodeasa resultof
more harsh competition in a volatilemarket.Theonlywayto
properly mitigate these challenges is to improve process
efficiency through yield maximisationthroughtechnological
innovation. Chemical companies all around the world are
seeking for new and innovative ways to save costsandboost
revenues. The application of artificial intelligence-based
techniques to extract value from huge amounts of
experimental data through data mining and knowledge
discovery is one of the most intriguing creative tactics to
investigate.
Chemical bioreactors are catching the interest of scientists
who are looking for new methods to make money. The
reactor is the only major piece of machinery that adds real
value to raw materials by turning them into finished goods.
In this aspect, reactor optimization has a tremendous
potential impact on overall profitability [1]. As a first step in
optimizing the bioreactor, modelling complex systems of
chemical processes is critical. In biochemical reactions,
complex reaction kinetics and thermodynamics were
involved. Building a credible phenomenological model for
bioreactors is a time-consuming and difficult task that
requires a detailed grasp of heterogeneous catalytic
behaviour such as mass diffusion and catalyst deactivation,
among other things. Furthermore, in laboratories, various
parameters such as agitation speed, temperature, negative
influence of toxins contained in incoming gas, diffusional
coefficients, and others affect reaction rate, yield, and
selectivity, the processes of which are unknown. As a result,
the reactor's optimization is hampered by a lack of
understanding of chemical reactiondynamics.Forsafety and
dependability concerns, most chemical bioreactorsremaina
black box, and scientists do not tamper with them. This
causes bioreactors to operate inefficiently, which has a
considerable influence on the profitability. As a result, inthe
chemical sector, reactors are viewed as an unknown
territory. A minor increase in catalyst selectivity and
reaction yield, on the other hand, has a major impact on raw
material consumption and overall profitabilityinlarge-scale
operations. [1].
To derive the kinetic equation of a given biochemical
reaction, laboratory scale kinetic analysis is typically
performed in a very ideal and controlled context. The
applicability of such an approach for kinetic modelling of
bioreactors is questionable due to the presence of poisons
and inert in input gases, and the varying heat and mass
transfer environment. Furthermore, chemical engineers are
unwilling to commit adequate time and effort to a full
examination of these complex reaction mechanisms due to
the restricted market window for chemical goods. As a
result, an alternative straightforward method is to develop
approximated reactor modelsforthesecomplicated reaction
systems using a data-driven effective computational
strategy, which can then be used to optimize the reactorand
increase profit.
The key challenge is figuring out how to leverage this
plethora of data to produce more money, because most
chemical laboratories collect and store vast volumes of
reactor input and output operational data. Data is the new
oil, and data-driven modelling tools such as artificial neural
networks (ANN) and support vector machines(SVM)arethe

new IC engines of our day. In the last decade, ANN and SVR
have become very popular black box modelling
methodologies, with a number of applications developed for
biochemical reactors. Engineers, on the other hand, reject
ANN and SVM models because they are difficult to
understand and provide a black box model. The ANN model
offers no insight into the underlying physics of chemical
reactors.
Despite its high prediction capacity, the model suffers from
explainability limits because it produces a black box type
equation consisting of a complex sigmoidal function with
several tuning elements known as weights and biases. To
acquire better understanding and profit, process engineers
prefer intelligible equations in differential/ algebraic form
that relate output variables to input features. A closed form
equation that can describe the effect of key process
parameters on the output variableispreferredinthecreated
model. SVM and ANN both have the difficulty of providing
closed form explainable equations that are portable and
simple to implement in a DCS system. Gaining insights and
obtaining a closed form explainable model equation is
crucial for engineers to accept the model's use in real life
Despite the ANN's outstanding prediction skills, there are
few uses of this technology in chemical reaction engineering
work in biochemical reactors. An attempt was made in this
work to use ANN modelling on a gluconic acid bioreactor.
One of the main objectives of this studyistoconverttheANN
model into an explainable closed form equation, which will
give crucial information aboutthereactor'sphenomenology.
The second purpose of this project is to train the ANN model
using freshly found nature-inspired grey wolf optimization
methodologies. Learning is a vital part of every neural
network, and it has piqued the interest of many researchers.
In most applications, the traditional [7] or upgraded [8–10]
Back-Propagation (BP) algorithm is employed to train
feedforward neural networks (FNNs). The BP method is a
gradient-basedalgorithm withsomeflaws,includingdelayed
convergence [11] and a proclivity to stay stuck in local
minima [12]. During the learning processofFNNs,thegoal is
to identify the best combination of connection weights and
biases to achieve the least amount of error. FNNs, on the
other hand, frequently converge on points that are the best
answer locally but not worldwide. In other words, rather
than the global minimum, learning methods lead FNNs to
local minima. According to [13], the BP algorithm's
convergence is strongly dependent on the weights, biases,
and parameters' initial values.Learningrateandmomentum
are two of these characteristics. A popular approach in the
literature is to enhance the problems of BP-based learning
algorithms by using unique heuristic optimization methods
or evolutionary algorithms.
Some of the heuristic optimization approaches used to train
FNNs include Simulated Annealing (SA) [13,14], Genetic
Algorithms (GAs) [15], Particle Swarm Optimization (PSO)
algorithms[16–20],MagneticOptimizationAlgorithm(MOA)
[21], and Differential Evolution (DE) [22]. Some algorithms,
such as SA and GA, can reduce the chance of local minima
trapping, but they still have poor convergence rates,
according to [11].
Despite the widespread usage of meta-heuristics in ANN
learning, none of them has done well in all applications.
Existing metaheuristics also have a number offlaws[23-27],
such as slow convergence speed, trapping in local minima,
long computational time, tuning many parameters, and a
difficult encoding scheme. As a result, it appears that
boosting the efficiency of ANN learning in various domains
requires either improving the performanceofexistingmeta-
heuristics or proposing new ones. [18] presented GWO, a
new stochastic and metaheuristic optimization technique.
The efficacy of the GWO approach for training FNNs is
studied in this work.
Because of its financial importance in global markets, the
current study chose to model a gluconic acid bioreactor.
Gluconic acid is used as the metal supplement of calcium,
iron, etc. in pharmaceuticals and as an acidulent in the food
industry. It also finds applications as a biodegradable
chelating agent, filler, metal cleaner,dyestabilizer,andinthe
textile industry for removing instructions. A reliable first
principle-based model is rarely accessible in the literature
due to a lack of understanding and complexity of multiphase
enzymatic reactions occurring in gluconic acid bioreactors.
Data driven modelling is a potential alternative strategydue
to the vast amount of bioreactor operating parameter data
available after multiple runs. The current project aims to
make use of a huge amount of process data to create a
framework for converting the data's information into profit.
The study's next goal is to use the developed model to
increase the gluconic acid factory's profit. This is performed
by optimizing the input process parameters utilising a
model-based, nature-inspired metaheuristic optimization
method in order to enhance gluconic acid yield (i.e., reactor
performance). The GA is used to improve the input space of
the bioreactor’s ANN model in order to give paretooptimum
solutions that achieve the objective in the most efficient
manner possible.
2. Case study of gluconic acid bioreactor
2.1. Background
Because of the lack of understanding and complexity of
multiphase catalytic processes in lab scale batch reactors,
trustworthy first principle-based models are hard to come
by. Data driven modelling is a feasible alternative strategy
due to the vast amount of reactor operating parameter data
available from bioreactors. The current projectaimstomake
use of a huge amount of process data to create a framework
for converting the data's information into profit.

2.2. Reactions:
Commercially, gluconic acid is produced primarily using
two biological methods, but more expensive chemical and
electrochemical ways are also available. The most prevalent
biochemical approaches are freecell fermentation and
immobilised enzyme-based glucose bioconversion (glucose
oxidase, GOD, of Aspergillus niger and Gluconobacter). The
GOD converts glucose to glucono-d-lactone, which is then
hydrolyzed by lactonase to gluconic acid.Producinggluconic
acid with immobilised enzymes is a costly and time-
consuming technique due to obstacles in the immobilisation
and separation phases; extra difficulties develop as a result
of denaturization of the enzymes. During free-cell
fermentation, mycelia are exposed to a variety of mass and
heat-transfer stresses. Mechanical agitation aids in the
removal of these restrictions, but it creates a turbulent flow
that can result in cell disintegration, cell fracture or surface
erosion, and pellet breaking. As a result, there may be a
sudden or gradual decline in cellular activity. On the other
hand, fermentation of gluconicacidbycellsimmobilisedona
support matrix under submerged conditions is a cost-
effective and efficient method.
Gluconic Acid Yield= (Moles of gluconic
acid produced)/(Moles of glucose consumed) (1)
2.3. Process flow diagram
A new batch fermentation procedure for producing
gluconic acid from glucose has recently been devised, withA.
niger immobilised on a cellulosic fabric support matrix
generating higher yields. The enhanced overall productivity
of this technique is mostly due to the increased interaction
between dissolved oxygen and the fungal mycelia. To
optimisetheaforementionedreaction,acontinuoussubstrate
dripping mechanism (see Figure 1) is used instead of the
mechanical agitation used in free-cell fermentation. As a
result,thebioreactor'syielddeterminesthetotalprofitability.
FIGURE 1 Experimental setup of the bioprocess taken from
[33]
2.4. Production Objectives
The purpose of this research is to develop a mathematical
model of the new glucose to gluconic acid batch
fermentation process and to discover the best process
conditions for higher gluconic acid yield. The fermenter
model was built using experimental data that took into
account the impacts of substrate (glucose), biomass, and
dissolved oxygen levels. Figure 1 depicts the complex
reaction and mass transfer pathways involvedintheglucose
to gluconic acid bioconversion usingA.nigerimmobilisedon
cellulose microfibrils. Because the physicochemical events
that drive bioconversion, as well as the kineticandtransport
mechanisms that follow them, are poorly understood.
3. Background materials
This part discusses the fundamental background materials
required for a complete understanding of the proposed
method. The Multi-layer perceptron (MLP) as a Feed-
forward neural network (FNN) is introduced first, followed
by GWO approaches, which are then compared to the
suggested technique for MLP learning in bioreactors
modelling. MLP network is followed by a brief discussion of
the hybrid ANN and GWO training method.
The feed-forward neural network (FFNN) is one of the most
common ANNs and receives a lot of academic attention
because of its ability to map any function to an unlimited
degree of precision. The multi-layer perceptron has been
employed in a variety of sectors, includingfinance,medicine,
engineering,geology,physics,andbiology.Nonlinearprocess
modelling, fault diagnostics, and process control are all
common applications in the field of chemical engineering.
As shown in Figure 2, MLP contains one input layer that
receives external inputs, one or more hidden layers,andone
output layer that displays the results. All levels, with the
exception of the input layer, are made up of processing
nodes and activation functions. The input layer provides
data, and the network nodes perform calculations in
successive layers until each output node receives an output
value.
3.1. Training of ANN by GWO
There are three techniques to train FNNs with a heuristic
algorithm in general. The ideal mix of weights and biases for
a FNN with the least amount of error isfirstdiscoveredusing
heuristic techniques. Second, inordertofindtheappropriate
structure for a FNN in a specific situation, heuristic
approaches are applied. Finally, an evolutionary algorithm
can be used to tune the parameters of a gradient-based
learning system, such as the learning rate and momentum.

In the first scenario, the structure is fixed before training
FNNs. The goal of a training method is to find a good value
for all connection weights and biases in order to lower the
FNNs' total error. In the second instance, the FNN structures
are different. A training technique is applied to a FNN to
determine the right structure for a given problem. Change
the structure of the FNN by manipulating the connections
between neurons, the number of hidden layers, and the
number of hidden nodes in each layer.
The first technique in this study is used to apply GWO to a
FNN; these operations are referred to as hybrid FNN-GWO.
The FNN's structure is fixed, and the GWO method choosesa
set of weights and biases that givestheFNN theleastamount
of error. In order to design FNN-GWO, the following
fundamental features must be defined. In FNN-GWO, build a
fitness function based on the FNN's mistake before
evaluating agents' fitness. Second, an encoding strategy for
the FNN-GWO agents should be devised to encode theFNN's
weights and biases. These elements are described in further
depth below.
3.1.1. Fitness function
The fitness function used in thisarticle[10]hasthefollowing
formula:
In Figure 2, which has two layers, the number of input nodes
is equal to n, the number of hidden nodes is equal to h, and
the number of output nodes is equal to m. (one input, one
hidden, and one output layer). At the end of each learning
period, the output of each hidden node is calculated as
follows:
, j =
1, 2, …, h (2)
In , n is the number of the input nodes,
wij is the connection weight from the ith node in the input
layer to the jth node in the hidden layer, Ɵj is the bias
(threshold) of the jth hidden node, and xi is the ith input.
After calculating outputs of the hidden nodes, the final
output can be defined as follows:
, k = 1,2, …, m,
(3)
where wkj is the connection weight from the jth hidden node
to the kth output node and Ɵk is the bias (threshold) of the
kth output node.
Finally, the learning error E (fitnessfunction)iscalculatedas
follows:
(4)
(5)
where q is the number of training samples, is the desired
output of the ith input unit when the kth training sample is
used, and is the actual output of the ith input unit when
the kth training sample is used. Therefore, the fitness
function of the ith training sample can be defined as follows:
Fitness (Xi) = E(Xi) (6)
FIGURE 2: FNN with a 2-3-1 structure
3.1.2 Encoding strategy
Following the specification of the FNN-GWOfitnessfunction,
the next step is to select an encoding technique for each
FNN-GWO agent to encode the FNN's weights and biases.
There are three approaches to encode and express the
weights and biases of FNNs for each agent in evolutionary
algorithms, according to [10]. These are the methods of
vector, matrix, and binary encoding. Each agent is encoded
as a vector in vector encoding. During training, each agent
reflects all of the FNN's weights and biases. Each agent is
encoded as a matrix in matrix encoding. Because we're
interested in training FNNs, we've employed the matrix
encoding technique in this piece. An example of this
encoding strategy for the FNN of Figure 2 is provided as
follows:
,
(7)
, , ,
(8)

where W1 is the hidden layer weight matrix, B1 is the hidden
layer bias matrix, W2 is the outputlayerweight matrix, is
the transpose of W2, and B2 is the hidden layer bias matrix.
3.2. Grey Wolf Optimization (GWO): at a glance
GWO, a new stochastic and metaheuristic optimization
technique, was introduced by [28]. GWO's purpose is to
mimic the cooperative hunting behaviour of grey wolves in
the wild. The grey wolf pack's hierarchical organisation and
predation behaviour are promoted by this bionic
optimization approach, in which the wolves take prey by
surrounding, haunting, and attacking it under the command
of the top grey wolf [31]. The top three grey wolves werethe
target of this large-scale search methodology, but there was
no way to eliminate them.Theoptimizationtechniquediffers
from others in terms of modelling. As shown in Figure 3, it
produces a strict hierarchical pyramid.
FIGURE 3: Hierarchy of grey wolf (dominance decreases
from top down)
The typical size of a group is 5-12 persons. The layer, which
consists of a male and female leader, is the most powerful
and capable member of the team when it comes to making
decisions on predation and other activities. The second and
third layers of the hierarchy, respectively, are in charge of
aiding group organisations in their behaviour. The bottomof
the pyramid, often known as, is home to the majority of the
world's population. They are largely responsible for
satisfying the entire pack by maintaining the dominance
hierarchy, regulating the population's internal connections,
and caring for the young [28].GWOmathematical modelling:
The main key point of the GWO model is thesocial hierarchy,
encircling, hunting, attacking and searching prey.
A. Social hierarchy
In the model of GWO, α is considered as fittest solution. β
and δ are considered as second and third best solution,
respectively. The rest of the solutions are assumed to be ω.
B. Encircling
At first, the location of the prey is determined andduring the
hunting process grey wolves encircled the prey. The
following equations are proposed for mathematically
modelling.
(9)
(10)
Where is the number of current iterations, is the
position vector of one grey wolf, is the next
position vector it arrives, is the position vector of the
prey, and are coefficient vectors which are evaluated as
follows:
(11)
(12)
and are random vectors in [0,1], ais decreasing value
during the iteration in [0,2], typically (I is
the maximum number of iterations).
In this concept, grey wolves move around the best solution
in hyper-cubes within an n dimensional space and able to
detect the position of the prey and encircle it.
C. Hunting
Grey wolves have the ability to hunt prey with the guidance
of alpha after encircling the prey. The beta and delta also
take part in the hunting procedure on occasion. The first
three best solutions in the mathematical stimulation of
hunting behaviourupdatethepositionofothersearchagents
(including the omegas). In this approach, the following
equations are proposed.
(13)
(14)
(15)

D. Attacking prey (exploitation)
Grey wolves complete the hunting phase and ready to
capture the prey. For the purpose of mathematically
modelling, the value of gradually decrease. Therefore, the
fluctuation rate of is also decreased by .
where a is decremented from 2 to 0 over the course of
iterations. When the random values of , the next
position of search agent can be in any position between its
current position and the position of the prey. When
, the grey wolves would diverge from the prey to
achieve global search and when , the grey wolves
would converge towards the prey and complete it.
E. Searching prey (exploration)
GWO algorithm usesanefficient explorationmethodology by
allowing its search agents to update their position on the
basis of alpha, beta, delta and attacks towards the prey. This
mechanism creates a good diversity in the problem search
space. Grey wolves are stay away from each other for global
search of prey and close to each other for attacking the prey.
For , values between 1 and -1 are taken. vector is also
favoured the explorationtechniqueandrandomvalues
are used. After generating the random population, alpha,
beta and delta determined the position of the best prey. For
selection of exploration and exploitation, the value of is
decreased from 2 to 0 respectively.
The GWO algorithm terminates when the criterion is
satisfied. This metaheuristic approach is applied in various
real-world problems because of its efficient and simple
performance ability by tuning the fewest operators [28-32].
3.3. Hybrid learning of GWO and FNN network
In this section, GWO is used to learn FNN. A hybrid learning
of GWO and FNN networks (GWO-FNN) is utilised to
increase the network's accuracy. The algorithm will
simultaneously determine the set of weights and their
related accuracybytrainingthenetwork.TheFNN network's
network weights and biases can be expressed as a D-
dimensional vector. The vector for FNN is defined by
Equation 16. To optimize the FNN weights using GWO
methods, each particle's dimensionisregardedasa vectorD.
D = (Input × Hidden) + (Hidden × Output) +Hidden bias +
Output bias, (16)
where Input, Hidden and Output are referred the number of
inputs, hidden and output neurons of FNN network
respectively. The number of biases in the hiddenandoutput
layers is also known as Hidden bias and Output bias. A
dataset is collected, normalized, and read to begin the GWO-
FNN. Following that, the appropriate number of inputs,
output, and hidden neurons are specified to establish the
particle dimension as Equation (5). The population is
initialized, and the training error is determined as a fitness
function following FNN training. Every particle (wolf)
modifies its velocity and position based on training error.
The new places represent the FNN network's new weights,
which are supposed to minimize the fitness function. The
fitness function is computed based on test set error. These
steps will go on until meeting stop conditions.
4. MATHEMATICAL MODELLING OF BIOPROCESS
4.1. Selection of input and output variables for
modelling
Because reactor yield has such a large impact on overall
profitability, it is kept as an output variable. All reactor
operational factors that could affect yield are stored as a
"wish list" of input variables. Initially, all bioreactor
experimental data was acquired. Following that, all of the
input factors that could affect the output variable were
recorded after speaking with a technical specialist. After
that, a cross-correlation analysis was carried out. This
method was used to determinethecorrelationcoefficientsof
each input variable with the output variable, as well as the
inter input cross-correlation coefficients.
The following criteria are used to shortlist the input
variables.
(i) For a particular input variable, there should behigh
cross-correlation coefficient with output
variable.
(ii) The values of cross-correlation coefficients of inter
input variables should be low.
(iii) The input set of variables were kept as minimum as
possible to avoid complexity of the model.
Based on the above criteria 3 input variables are finally
shortlisted and tabulated in Table 1.
TABLE 1 Input Output variables for model building and
their range
Variables used in modeling Data Range
Input Variables
Glucose concentration, g/L (x1) 100.0–180.0
Biomass concentration, g/L (x2) 1.00–3.00
Dissolved oxygen concentration, mg/L
(x3)
10.0–60.0
Output Variables
Gluconic Acid Yield, % (Y1) 5.9–94.58

4.2. Data collection, Data cleaning and removal of
outliers
The GWO-FNN based model for the glucose to gluconic acid
bioprocess was created using experimental input-output
data from the fermenter. In this study, the gluconic acid-
producing strain Aspergillus niger NCIM 545 wasemployed.
The spore germination media, growth medium, and
cellulosic fibre support have all been well reported
previously.
The quality of data used to build a data-driven model is
widely accepted as determining the model'squality.Because
noisy and erroneous data can have a major impact on model
performance, data quality is an important consideration
when employing data driven modelling. Due to the high
number of process data, this research developed an
automated data cleaning technique that eliminates the need
for human intervention. In this work, the data was pre-
processed using multivariate Principal Component Analysis
(PCA). An automated MATLAB-based system was built to
build a multivariate statistical vector called t-squared from
the operating dataset. The corresponding rows in the t-
squared vector with values over the 95th percentile was
therefore considered outliers and were eliminated from the
dataset.
4.3. Modelling through FNN-GWO algorithm
An ANN-based model was built using the cleansed data. The
dataset, which included eight inputvariablesandtwooutput
variables (Table 1), was divided into a training set (80% of
the total data) and a test set (the remaining 20%). (20
percent of the total data). The training data set was used to
build the model by maximising the fitness value, and thetest
data set was used to cross-validate the result. Cross-main
validation's purpose is to improve the model's
generalizability.
The FNN-GWO-based model was developed using MATLAB
2019a code. The mean-squared error (MSE) between actual
and projected outputs was employed as the fitness function
in this study, and the programme was run in such a way that
the MSE value was minimized. Due to the stochastic nature
of the ANN, the software was run 100 times to create the
model.
4.4. Optimization through Genetic Algorithm
The model is then utilized to optimizethebioreactorprocess
parameters once a trustworthy and accurate bioreactor
model has been built. The goal is to find the best process
conditions for maximal bioreactor profitability. In other
words, the optimization algorithmshouldaimforthehighest
possible reactor yield.
A multi-objective genetic algorithmisutilisedinthisstudyto
establish a balance between two opposing goals.Thegenetic
algorithm (GA) has shown to be a powerful optimizationtool
that has been applied to a wide range of technical and
medical applications.
For implementation of GA algorithm, an objective function
was developed which is as follows (Equation 17):
F1(x) = 1/Y1(x)
(17)
where Y1(x) is the function of the model corresponding to
gluconic acid yield. Therefore, in GA F1(x) has to be
minimized in order to maximize the yield of the bioprocess.
5. Results and discussions
5.1. Performance of FNN-GWO model
The major purpose of this study is to create an accurate,
simple, portable, and easy-to-understand closed model
equation for a bioreactor.
The values of ANN parameters required for modelling were
determined using a trial-and-errorapproachanda literature
review. In the current work, the number of nodes in the
hidden layer fluctuates consistently from 5 to 25, and the
FNN GWO approach is used to find the appropriate weights
and bias that produce the lowest MSE between actual and
anticipated output each time.
Shortlisting the models: To choose a valid model from a
pool of likely candidates or representative model equations
with varying degrees of complexity and accuracy, the
following criteria were used:
(i) Simplicity: The model should be as straightforward
as possible. Model complexity was determined
by the number of nodes in the hidden layer.
(ii) Prediction accuracy: The gapbetweenexpectedand
actual yield.
(iii) The fundamental physics of the process should be
captured in the model equation. To put it
another way, model equations should includea
physical understanding of the system under
investigation rather than just a prediction
relationship. This isanimportantconsideration
for building realistic reactor models.
TABLE 2 Rules to select best model from experimental
observations.
Sl. no Parameters changed keeping all other parameters constant What happen to gluconic acid
yield?
1 If glucose concentration increase Increases
2 If biomass concentration increase Slightly decreases
3 If dissolved oxygen (DO) concentration increase Increases

FIGURE 4: Actual vs. predicted plots of (a) gluconic acid
yield with training data (b) gluconic acid yield with testing
data
The ANN-acquired equations were subjected to the
aforementioned assessment to see if the produced equation
agrees with the experimental data. The models that were
created were thoroughly tested. Assume that a ten-test data
set was created with all variables set to their 50th percentile
value, except for oxygen inlet concentration, which was
changed at 10-point intervals from its minimum to
maximum value. After that, a graph of glucose concentration
vs. gluconic acid yield was made. Using these charts, Table 2
observation number 1 was confirmed. Only one model
equation for gluconic acid yield was chosen from the
shortlisted model equations because they are exceedingly
precise, obey the Table 2 observations, and capture the
internal physics of the reactor, such equation is considered
the representative model equation for gluconic acid yield.
The corresponding Coefficient of Determination (R2) and
Average Percentage Error (APE) of the above model for
training and test data have been mentioned in Table 3.
With high R2 and low APE for the gluconic acid yield (Table
3), it can be concluded that the predicted output values are
comparable to the actual output values and that the models
developed are reliable, fairly accurate, and capture the
inherent physics of the bioreactor. The model's
generalizability and accurate learning on nonlinear input
and output relationships are further indicated by the strong
R2 value on unseen test data and low APE.
Figure 4 depicts the models' prediction performance on
training and testing data. The fact that the real and forecast
curves almost coincide implies that the model has strong
prediction accuracy.
From Table 3 and Figure 4 it is concluded that developed
model is highly accurate and reliable as it also performswell
with unseen test data.
Generation of explainablemodel equations: ThoughANN
generates a closed form of equation which has very high
predictive capability, the developed equation is large and
complex and sometimes difficult to directly interpret. In
present study, a methodology is developed to enhance the
interpretability of the developed equations. Figure 5
summarizes the developed methodology. Those figures are
generated by changing one variable at a time from its
minimum to maximum value (10 steps) while keeping all
other 2 input variables at their 50-percentile value. Yield
equations developed by ANN is used to predict the yield
value in each case of these simulated test data. After plotting
was done, a trendline was drawn through each data whose
equation and R2 value is shown in figure. Based on visual
inspection and R2 value trend line curve was selected (like
straight line, or polynomial with degree 2 or 3 or more) so
that generated trendline almost matches with the data. As
seen from the figure 10, the developed trend lines are very
decisive and monotonically increasing and decreasing. As
mentioned earlier, they all match the actual observations
and obey the Table 2. In short, developed models captures
the nonlinear relationship between yield and reactor
operating parameters. These trend lines can be used by
scientists to get the insights on how a particular input
parameter affects the gluconic acid yield. For example, from
figure 5, it is quantitively clear that increasing glucose and
DO concentration, actually enhance the yield whereas
increasing biomass reduces the yield.
The yield decreases linearly with biomass concentration
with negative slope of 12.172, whereas the relation of yield
with glucose and DO concentration are nonlinear and
represented by second and third order polynomial,
respectively. Now these trend line equations are used to
develop the following explainable equation (Equation 18).
(18)
where x1, x2, x3 are the actual value of the 3 input variables
and x1, avg, x2, avg, x3, avg are the average (50 percentile) value of
input variables, respectively.
Each term in the equation 18 represents the change in yield
if a particular parameter deviates from itsaveragevalue.For
example, the term, represents the deviation
of biomass concentration from its average value and whenit
multiplied by co-efficient -12.172, represents the yield
penalty (or gain) due to biomass. In this way, all 3
parameters contribution is calculatedinequation18and itis
added with 72.71% (average yield) to get the actual yield.
Main advantage of this equation (equation 18) over ANN
model is that this equation is interpretable and easily
APE
Gluconic acid yield model 0.99 0.42 0.99 0.42
2
APE R
2
TABLE 3. Performance of ANN model
Training Testing
Model R

explainable to engineers. Equation is simple and contains
terms or parametric co-efficient which throw lights relative
importance of each parameter on the overall yield if they
deviate from this base value. Also, it indicates whether the
effect of each parameter is linear or non-linear.
Equation 18 is then used to predict yield of experimental
data and predicted, and actual yield is compared. Prediction
error is 0.42% and R2 is 0.99. This low value of prediction
error and high value of R2 signifies that developed equation
(equation 18) is highly accurate and reliable.
a) b)
c)
FIGURE 5: Influence of each variable on gluconic acid
yield
5.2. Optimization
5.2.1. Optimization through GA
Once the reliable models were successfully developed, the
models were subjected to optimization. Purpose of
optimization is to find the optimumvalueofrectoroperating
parameters to achieve maximum gluconic acid yield. One of
the critical tasks for optimization of any process is fixing of
search space at which the optimal process conditions are to
be found out. Therefore, before running the optimization a
lower bound and upper bound of the process variableswere
fixed in consultation with scientists. The lower bounds (LB)
and upper bounds (UB) considered in this case have been
depicted in Table 4.
Table 4. Lower bounds and upper bounds for
optimization
x1 x2 x3
LB 100 1 10
UB 180 3 60
With the help of GA tool in MATLAB, the optimum
experimental conditions were found out which gives the
gluconic acid yield of 99.59% (Table 5). The main advantage
of such a study is that it gives the experimental engineers a
strategy to run the reactor in optimum condition in real-
time. Due to unavailability of an explainable model, since
scientists have no idea about optimal solution,
experimentalists try to optimize the process heuristically
based on their experience and knowledge. The only action
scientist must do is to run the GA with proper bound in real-
time, and GA will provide a set of optimum operating
conditions that the scientist needs to set in the experiment.
Table 5. GA optimal solution
x1 x2 x3 Yield
161.80 1.00 58.64 99.59
6. Conclusion
From the existing operating data, this study uses Artificial
Neural Networking to construct an accurate model of a
gluconic acid bioreactor. An ANN creates a closed model
equation that is portable and may be used in a control
system. The true value of this research is that it has
produced an explainable model equation that is very
accurate and provides insights into the process. The
produced model equations are based on the underlying
physics of the process and are in line with the observations
and experiences of the experimentalist. After that, the
developed model equations are used to construct optimum
solution that optimize gluconic acid yield and thus ensure
profit maximization.
References
1. Lahiri, Sandip K.
Profit MaximizationTechniquesforOperatingChemical
Plants. John Wiley & Sons, 2020.
2. Lakshminarayan
an, S., Fujii, H., Grosman, B., Dassau, E., Lewin, D.R.,
2000. New product design via analysis of historical
databases. Comput. Chem. Eng. 24, 671–676.
https://doi.org/10.1016/S0098-1354(00)00406-3
3. Grosman, B.,
Lewin, D.R., 2002. Automated nonlinear model
predictive control usinggeneticprogramming.Comput.
Chem. Eng. 26, 631–640.
4. Searson, D.,
Willis, M., Montague, G., 2007. Co-evolution of non-
linear PLS model components.J.Chemom.21,592–603.
https://doi.org/10.1002/cem.1084
5. Barati, R.,
Neyshabouri, S.A.A.S.,Ahmadi,G.,2014.Developmentof
empirical models with high accuracy for estimation of
drag coefficient of flow around a smooth sphere: An
evolutionary approach. Powder Technol. 257, 11–19.
https://doi.org/10.1016/j.powtec.2014.02.045
6. Floares,
Alexandru G., and Irina Luludachi. "Inferring
transcription networks from data." SpringerHandbook

of Bio-/Neuroinformatics. Springer, Berlin,Heidelberg,
2014. 311-326.
7. Horne, B.G.,
1993. Progress in supervised neural networks. Signal
Process. Mag. IEEE 10, 8–39.
8. Hagan, M.T.,
Menhaj, M.B., 1994. Training Feedforward Networks
with the Marquardt Algorithm. IEEE Trans. Neural
Networks 5, 989–993.
https://doi.org/10.1109/72.329697
9. Adeli, H., Hung,
S.L., 1994. An Adaptive Conjugate Gradient Learning
Algorithm. Appl. Math. Comput. 62, 81–102.
10. Zhang,N.,2009.
An online gradient method with momentum for two-
layer feedforward neural networks. Appl. Math.
Comput. 212, 488–498.
https://doi.org/10.1016/j.amc.2009.02.038
11. Zhang, J.R.,
Zhang, J., Lok, T.M., Lyu, M.R., 2007. A hybrid particle
swarm optimization-back-propagation algorithm for
feedforward neural network training. Appl. Math.
Comput. 185, 1026–1037.
https://doi.org/10.1016/j.amc.2006.07.025
12. Gori,M.,Tesi,A.,
1992. On the problem of local minima in
backpropagation.IEEETrans.PatternAnal.Mach.Intell.
13. Shaw, D.,
Kinsner, W., 1996. Chaotic simulated annealing in
multilayer feedforward networks. Can. Conf. Electr.
Comput. Eng. 1, 265–269.
https://doi.org/10.1109/ccece.1996.548088
14. Koh, C.S., Hahn,
S.Y., 1994. Detection of Magnetic Body using Artificial
Neural Network with Modified Simulated Annealing.
IEEE Trans. Magn. 30, 3644–3647.
15. Montana, D.J.,
Davis, L., 1989. Training FeedforwardNeural Networks
Using Genetic Algorithms. Proc. 11th Int. Jt. Conf. Artif.
Intell. - Vol. 1 89, 762–767.
16. Kiranyaz, S.,
Ince, T., Yildirim, A., Gabbouj, M., 2009. Evolutionary
artificial neural networksby multi-dimensional particle
swarm optimization. Neural Networks 22, 1448–1462.
https://doi.org/10.1016/j.neunet.2009.05.013
17. Settles, M.,
Rylander, B., 2002. Neural network learning using
particle swarm optimizers. Adv. Inf. Sci. Soft Comput.
224–226.
18. Zhang, C., Li, Y.,
Shao, H., 2000. A new evolved artificial neural network
and its application. Proc. World Congr. Intell. Control
Autom. 2, 1065–1068.
https://doi.org/10.1109/wcica.2000.863401
19. van den Bergh,
F., Engelbrecht, A.P., 2000. Cooperative Learning in
Neural Networks using Particle Swarm Optimizers.
South African Comput. J. 26, 84–90.
20. Zhang, C., Shao,
H., Li, Y., 2000. Particleswarmoptimizationforevolving
artificial neural network. Proc. IEEE Int. Conf.Syst.Man
Cybern. 4, 2487–2490.
https://doi.org/10.1109/icsmc.2000.884366
21. Mirjalili, S.,
Sadiq, A.S., 2011. Magnetic Optimization Algorithm for
training Multi Layer Perceptron. 2011 IEEE 3rd Int.
Conf. Commun. Softw. Networks, ICCSN 2011 42–46.
https://doi.org/10.1109/ICCSN.2011.6014845
22. Si, T., Hazra, S.,
Jana, N.D., 2012. Artificial neural network trainingusing
differential evolutionary Algorithm for classification.
Adv. Intell. Soft Comput. 132 AISC, 769–778.
https://doi.org/10.1007/978-3-642-27443-5_88
23. Leung,Y.,Gao,Y.,
Xu, Z. Ben, 1997. Degree of population diversity - A
perspective on premature convergence in genetic
algorithms and its Markov chain analysis. IEEE Trans.
Neural Networks 8, 1165–1176.
24. Hrstka, O.,
Kučerová, A., 2004. Improvementsofreal codedgenetic
algorithms based on differential operators preventing
premature convergence. Adv. Eng. Softw. 35, 237–246.
25. Liang, J.J., Qin,
A.K., Suganthan, P.N., Baskar, S., 2006. Comprehensive
learning particle swarm optimizer for global
optimization of multimodal functions.IEEETrans.Evol.
Comput. 10, 281–295.
https://doi.org/10.1109/TEVC.2005.857610
26. Gao,W.feng,Liu,
S. yang, Huang, L. ling, 2012. Particle swarm
optimization with chaotic opposition-basedpopulation
initialization andstochasticsearchtechnique.Commun.
Nonlinear Sci. Numer. Simul. 17, 4316–4327.
https://doi.org/10.1016/j.cnsns.2012.03.015
27. Moslemipour,G.,
Lee, T.S., Rilling, D., 2012. A review of intelligent
approaches for designing dynamic and robust layouts
in flexible manufacturing systems. Int. J. Adv. Manuf.
Technol. 60, 11–27. https://doi.org/10.1007/s00170-
011-3614-x
28. Mirjalili, S.,
Mirjalili, S.M., Lewis,A.,2014.GreyWolfOptimizer.Adv.
Eng. Softw. 69, 46–61.
https://doi.org/10.1016/j.advengsoft.2013.12.007
29. Faris,H.,Aljarah,
I., Al-Betar, M.A., Mirjalili, S., 2018. Grey wolfoptimizer:
a review of recent variants and applications. Neural
Comput. Appl. 30, 413–435.
https://doi.org/10.1007/s00521-017-3272-5

30. Saremi, S.,
Mirjalili, S.Z., Mirjalili, S.M., 2015. Evolutionary
population dynamics and grey wolf optimizer. Neural
Comput. Appl. 26, 1257–1263.
https://doi.org/10.1007/s00521-014-1806-7
31. Nadimi-
Shahraki, M.H., Taghian, S., Mirjalili, S., 2021. An
improved grey wolf optimizer for solving engineering
problems. Expert Syst. Appl. 166, 113917.
https://doi.org/10.1016/j.eswa.2020.113917
32. Mirjalili, S.,
Saremi, S., Mirjalili, S.M., Coelho, L.D.S., 2016. Multi-
objective grey wolf optimizer: A novel algorithm for
multi-criterionoptimization. ExpertSyst.Appl.47,106–
119. https://doi.org/10.1016/j.eswa.2015.10.039
33. Singh Cheema,
J.J., Sankpal, N. V., Tambe, S.S., Kulkarni, B.D., 2002.
Genetic programming assisted stochastic optimization
strategies for optimization of glucose to gluconic acid
fermentation. Biotechnol. Prog. 18, 1356–1365.
https://doi.org/10.1021/bp015509s
34. Pal, S., & Lahiri,
S. K. (2022). Grey wolf optimizer trained ANN
technique for development of explainable model of
commercial ethylene oxide reactor and multi-objective
optimization to maximize profit.
35. Pal, S.,
Chowdhury, S., Hens, A., & Lahiri, S. K. (2022). Artificial
intelligence based modelling and multi-objective
optimization of vinyl chloride monomer (VCM)plantto
strike a balance between profit, energy utilization and
environmental degradation. Journal of the Indian
Chemical Society, 99(1), 100287.
36.
APPENDIX:
Experimental Data Utilized for Building GP Based Model
taken from [33]
batch
no
glucose
concn
(x1)
(g/L)
biomass
concn
(x2)
(g/L)
DO
(x3)
(mg/L)
gluconic
acid yield
(Y1) (%)
1 100 1 10 5.9
2 150 2 10 29.42
3 120 2 15 20.76
4 150 2.5 15 35.51
5 150 3 15 35.16
6 120 2 25 27.77
7 120 2 30 34.48
8 150 2 30 57.86
9 150 3 25 49.32
10 150 2 40 78.99
11 150 2 45 89.48
12 150 2 50 94.5
13 180 2 50 89.63
14 150 3 40 79.05
15 150 2.5 50 94.58
16 150 2.5 55 93.41
17 150 2.5 60 91.26
18 160 2.5 60 93.67
19 175 3 55 92.69
20 160 3 60 93.3
21 180 3 60 88.13
22 150 3 60 92.7
23 100 3 60 20.04
24 100 2 10 6.13

Development of an Explainable Model for a Gluconic Acid Bioreactor and Profit Maximization through Grey Wolf Optimizer Trained Artificial Neural Network Technique

More Related Content

Similar to Development of an Explainable Model for a Gluconic Acid Bioreactor and Profit Maximization through Grey Wolf Optimizer Trained Artificial Neural Network Technique (20)

More from IRJET Journal (20)

Recently uploaded (20)

Development of an Explainable Model for a Gluconic Acid Bioreactor and Profit Maximization through Grey Wolf Optimizer Trained Artificial Neural Network Technique