Utilization of Artificial Neural Network in optimizing square

cascades for separation of 130Te isotope

*
Morteza, Imani; Mohammademad Adelikhah; Mehdi, Aghaie
Engineering Department, Shahid Beheshti University, G.C., P.O. Box: 1983963113, Tehran, Iran
Corresponding Author’s E-mail: m_aghaie@sbu.ac.ir

Abstract
Separation of stable isotopes has been considered today due to their widespread use. The separation
of the isotope 130Te is important for medical applications, and the production of radioisotopes up to
high concentrations. In this paper, square cascade optimization to achieve 99.9% concentration of
this isotope by the gray wolf optimization algorithm is presented. In the optimization, instead of
solving nonlinear equations of concentration distribution in the cascade, a trained neural network is
used to predict the value of the objective function. To train the neural network, 400 randomly
generated data from the simulation results were used. Predicting the objective function using a
neural network leads to a 98% reduction in optimization execution time. Using this method, the
optimal cascade separates 3409 g of 130Te with 99.9% concentration from 10 kg of natural tellurium
during one year.
Keywords: Square cascade, Separation, 130Te, Gray wolf algorithm, Neural network.

1. INTRODUCTION

Today, stable isotopes are widely used in the industry, especially in medical sciences. So, the development of
methods for extracting them from natural resources is essential. One of the efficient methods for separation of
stable isotopes is the use of Gas Centrifuge (GC) cascades [1]. Tellurium consists of eight stable isotopes
with different applications. For example, the heavy component, 130Te, is used in 131I radioisotope production,
which can be applied in life science for healthcare and medical applications and pharmaceuticals
industries[2]. Square and taper cascades can be used to separate multicomponent isotopes by GCs. In taper
cascades, using model cascades such as R, Q and quasi ideal, only a cascade can be designed that is used to
separate one isotope[3]. This cascade cannot be used to separate other isotopes; to do this, the cascade
structure needs to be redesigned. Therefore, it is not economically viable to design a separate cascade capable
of separating the isotopes individually. Accordingly, the design and construction of cascades in commercial
applications is going to the direction that separation of a wide range of stable isotopes would be possible by a
single cascade. Nowadays, square cascades are used for this purpose[4]. Due to the recycle flows in the first
and last stages, square cascades can be used in low-feed flow rates and different cuts, which have a high
flexibility in operation and can be adjusted by the control valves between the stages of the cascade.
Separation of isotopes using a square cascade is associated with great complexity. Determination of process
parameters to reach the desired concentration is associated with great complexity. For example, for the square
cascade, the mentioned parameters are the cascade feed flow rate, feed location, cascade cut, feed flow of
GCs and cut of the first stage. By changing the parameters, different concentrations of the desired isotope can
be achieved in the light or heavy flow of the cascade. Due to this and the relationship between the cascade
parameters, it is not possible to achieve a definite theoretical solution method for determining the optimal
process parameter for separating each isotopes, but optimization algorithms can be used to achieve the
desired goal[5].
In optimization problems, the definition of objective function is important. In various researches, the
objective functions that have so far been studied, are separation capacity, total feed flow of stages, number of
GCs used in the cascade, D parameter along with total interstage flow rates[6], and the desired isotope
concentration along with total interstage flow in taper cascades. In this paper, the objective function is to find
the maximum amount of product in a given amount of natural feed of tellurium.

1
In this study, the enrichment of 130Te (heaviest isotope of the tellurium element) has been investigated by
using a square cascade of GCs. To determine the best square cascade, the process parameters, which include
cascade feed flow rate, feed location, cascade cut, feed flow of GCs and cut of the first stage are optimized by
the Gray Wolf Algorithm. The optimization process with the aim of increasing product from a defined input
has been done by gray wolf algorithm based on artificial neural networks (ANN). Artificial Neural Networks
(ANNs) are the computational models which copy the human brain’s learning and decision making abilities.
Before they are used, they are made to go through a learning process[7]. Their decision-making ability and
correctness of the decisions depend on several key factors such as the structure of network, type of activation
functions in the neurons, period of learning, and number of neurons used during the learning process. The
ANN is a well-established tool for different forecasting problems in different areas, with over 40 years of
application [8]. Because of their universal approximate functional form, ANN seems to be a good choice for
modeling non-linear dependency. In this research, the ANN provides the fast access to the fitness function
values. Finally, the optimized parameters are reported and results are discussed.

1. Square cascades

In a square cascade, GCs are connected in parallel and in series. The number of GCs is equal in all stages; i.e.
the input flow rate at all stages is the same. Figure 1 illustrates a square cascade how the control valves are
placed among the stages. Another feature of square cascades is the use of recycle flows in the first and last
stages of the cascade. Square cascade simulations can, generally, be divided into two stages: a) calculation of
interstage flow rates, and b) calculation of concentration distribution. In the first step, using the equations
obtained from the total mass balance at each stage and at mixing points, all flow rates between the steps will
be calculated. In the second step, by solving the mass conservation equations for each component in stages
and mixing points, the concentration distribution along the cascade will be calculated [9].

Figure 1. Schematic diagram of a square cascade

2.1. Calculation of interstage flows

The external hydraulic parameters of cascade are introduced as follows: cascade feed flow ( ), waste flow
( ) and product flow ( ). The internal hydraulic parameters of cascade are product flow of stages ( ), waste
flow of stages ( ) and interstage cut ( )[10].

Figure 2. Schematic of a square cascade

Another internal hydraulic parameter is input feed to the stages ( . By placing the following equations in a
linear equation system, the hydraulic parameters are calculated as follows ( )[11]:

2
 (1)
(2)

(3)

|(4)

(5)
(6)
2.2. Concentration distribution in the cascade
The particle mass conservation equation for isotope i in step n is written as follows [12]:
(7)

Similarly, for the input feed node (mixing point), the mass conservation equation for isotope i is obtained as
follows:
(8)

In Equations 7 and 8, is the isotope i concentration in the product stream, and is the isotope i
concentration in the waste stream. Another relation used is the step separation factor is as follows[13]:

(9)

In addition to the relationships mentioned above, the following conditions should be calculated in the input
and output flows for each stage[14]:
(10)

The number of flows in the cascade is fewer than the number of the equations. Therefore, it is necessary to
specify two parameters in the input, so that the number of the unknowns and the equations are equal to each
other. In this paper, these two parameters are the cascade cut and the cut of the first stage.
Using Equations (7) and (8), the following equation is obtained:
(11)

In Equation (11), the known values are of the right side of the equation, and the values of the unknown
concentrations are of the left side of the equation. The coefficients can be named and the relation can be
written in the following simplified form:

(12)

To solve the above non-liner equation, the q iteration method is used. In this method, a parameter called q is
defined, which is equal to the ratio of the concentration of one isotope in the enriched stream to the
concentration of the same isotope in the prepared stream[15]:

(13)

By applying this definition to the separation factor relation, the following relation will be obtained:

3
 (14)

(15)

Using Equation (15), the value of concentration can be determined as follows:

(16)

By placing the above relation to the left of equation (12), the following relation is obtained [13]:
(17)

The above equation must be solved at each stage for each isotope to calculate the concentration of all
components.

2. Optimization

3.1. Objective function

In square cascade design, there are five parameters, including feed location, feed flow rate of the cascade, the
cut of the first stage, cascade cut and feed flow of GC for optimization. Square cascade can be used in low
feed flow rate and the cascade cut can change in a wide range, so the maneuverability of the square cascade is
very high such that it has a very high flexibility in separating the isotopes. After determining the structural
parameters of the cascade, in order to achieve the target isotope concentration, at this stage, the cascade must
be optimized to achieve the set goal. The objective function specified in this article is as follows:

(18)

In the objective function, the value of the final concentration and the recovery factor of the cascade, R, are
applied as constraints. The index De represents the desired value of the parameter and the calc index
represents the values calculated from the simulation. In the objective function, the fixed values of are
equal to 10, and is equal to 10000.

3.2. Gray wolf algorithm (GWO)

To optimize square cascades, the Gray Wolf meta-heuristic algorithm is used. The Gray Wolf algorithm was
proposed by Mirjalili et al. in 2014 based on their group hunting. The method of hunting gray wolves is
mathematically modeled and the best solution is called alpha, the second optimal solution is beta, and the
third optimal solution is called delta; the rest of the answers that do not matter are called omega. The hunting
process is monitored by the alpha, beta and delta wolves. Omega wolves are looking for alpha, beta, and delta
wolves. Gray wolves basically follow three stages of hunting. The first stage is tracking and chasing the prey,
the second stage is immobilizing the prey by encircling it, and the third stage to attack the prey. The second
stage is introduced by Equations 19 and 20[16]:
(19)

(20)

Where and are the coefficient vectors, and are the prey and the gray wolf’s position, respectively,
and the determines the new position of the wolves. In addition, the parameter t represents the number of
iterative loops, and indicates the position of the wolves in the next iteration. The vectors and
are defined as follows[17]:

4
 (21)

(22)
Where, the vector changes linearly from zero to 2 along the iteration loop. In the first iteration, it is equal
to 2, and in the last iteration, it is equal to zero. The vectors and are random vectors in the range of zero
to 1. In modeling the hunting behavior of gray wolves, alpha, beta and delta wolves are used assuming that
they have more information about the hunting position; therefore, the position of omega wolves based on
these three wolves (optimal answers) is updated according to the following relationships[18]:
(23)

(24)

(25)

The vectors , and are the three optimal answers of the algorithm in the iteration of t. The vector
has a random value between -2a and 2a. When the wolves attack and when the wolves have
to change their position to find the prey (see for more details [16]).
To find the optimal cascade, the GWO algorithm first generates a random population of wolves (answers).
Square cascade will have five parameters for optimization, and five random parameters will be generated.
The values of these defined parameters are between the defined maximum and minimum values. For
example, the separation has random parameters as follows:
(26)
In this vector, NF is the feed location, Feed is the input feed flow rate, θ1 is the cut of the first stage, SF is the
feed flow rate of the centrifuges and θ_cas is the cascade cut. The value of fitness function is calculated by a
trained neural network.
3.3. Prediction of fitness function by an artificial neural network (ANN)
Artificial neural networks are systems of acting machine learning, in which a computer learns to perform
some task by considering training examples and data. They automatically produce identifying characteristics
from the learning data that they process. In fitting problems, neural network plans between a data set of
numeric inputs and a set of numeric outputs. A trained neural network can fit multi-dimensional mapping
problems well, given enough data and neurons in its hidden layer[19]. The networks can train with
Levenberg-Marquardt back propagation, and Scaled Conjugate Gradient, and Bayesian Regularization
algorithms. The artificial neural network Toolbox in Matlab has many applications in engineering projects
and it has been used in this research[20].
To generate a quick and rough prediction for fitness function in optimization algorithm, a feed-forward multi-
layer neural network with a 15 hidden layer has been used to predict the square cascade parameters. The
square cascade was simulated to apply the input data for the neural network. The amount and concentration
of product have been computed for 400 randomly generated data for square cascades. The network randomly
divides all samples into 3 groups including training, validation, and test data. 280 of the data have been
considered for training, 60 for network testing, and 60 for validation. Training data are presented to the
network during training, and the network is adjusted according to its error. Validation is used to measure
network generalization, and to halt training when generalization stops improving. Testing data are
independent from training and provides an independent measure of network performance. The Levenberg-
Marquardt algorithm has been chosen for the network. Training automatically stops when generalization
stops improving, as indicated by an increase in the mean square error of the validation samples. The main
steps of GWO-ANN optimization method have been illustrated in Figure 3.

5
 Figure 3. GWO-ANN optimization algorithm

3. Results and Discussion

To separate the tellurium isotope 130 from 10 kg of natural feed in order to achieve 99.9% concentration, a
square cascade with 30 stages and 6 centrifuges in each stage was considered. Table 1 shows the natural feed
concentrations of tellurium. Because the 130Te is the heaviest isotope of tellurium, separation is done from the
heavy isotopes withdrawal of the cascade. In Figure 4, the separation factor of centrifuge machines in terms
of input feed has been illustrated.

Table 1: Natural feed concentration of tellurium

Te 120 Te 122 Te 123 Te 124 Te 125 Te 126 Te 128 Te 130
0.0009 0.0255 0.0089 0.0474 0.0707 0.1880 0.3170 0.3416

Figure 4. Variation of separation factor in terms of inlet gas flow to centrifuge

The network training process is repeated several times by modeling data to obtain a network with the least
error. The network error regression for training data, validation data, and test data are shown in Figure 5.
According to the figure, values close to 1 confirm the proper operation of the network.

6
 Figure 5. The neural network training regression

Figure 6: The neural network training mean square error

Figure 6 shows the mean square error for training process, and as can be seen this parameter confirms that the
training process performs well and the best validation performance is 0.00019858 at epoch 20.
After confirming network performance, to continue the optimization process, the initial search agents and the
number of iterations were considered 30 and 1000 respectively. Table 2 displays all parameters of the 10 best
runs and the top fitness values for the GWO-ANN. The optimization operation by GWO takes about 1 h on a
personal computer (Intel (R) Core (TM) i7-9750H CPU @ 4.40 GHz). It takes about 50 s by replacing the
network with the fitness function.

Table 2. The GWO-ANN results for 130Te separation

Run Feed SF
NF θ1 θ_cas Concentration Produc (gr) Duration (day)
No (mgr/s) (mgr/s)
1 18 0.317 0.467 1.06 0.668 0.99903 3311.118 365
2 18 0.317 0.439 1.02 0.667 0.99903 3323.181 365
3 21 0.317 0.417 1.00 0.676 0.99918 3235.296 365
4 18 0.317 0.464 1.01 0.667 0.99902 3327.625 365
5 22 0.317 0.463 1.05 0.677 0.99901 3390.071 365
6 18 0.317 0.463 1.05 0.667 0.99908 3300.673 365
7 21 0.317 0.463 1.05 0.677 0.99923 3253.067 365
8 22 0.317 0.413 1.01 0.677 0.99932 3228.066 365
9 14 0.317 0.473 1.00 0.659 0.99900 3409.262 365
10 22 0.317 0.441 1.00 0.659 0.99902 3401.222 365

7
Figure 7 is shown the concentration of isotopes along the optimal cascade (Run No.89). As can be seen, 130Te
reach to highest concentration in first stage.

Figure 7. The concentration distribution of isotopes in the optimum cascade

Figure 8 shows the shear factor between the steps for the optimal cascade obtained. As can be seen, there are
three different values of the shear coefficient for the stages, and by setting these values between the stages,
the cascade can be operational.

Figure 8. The cut of stages in the optimum cascade

4. Conclusion

In this analysis, optimization of the square cascade for 130Te separation to high concentration is presented.
GWO algorithm has been applied to optimize the square cascade and a trained neural network has been used
for predicting the fitness function values. The results showed that the neural network can operate with
acceptable accuracy to find the fitness performance parameters and reduces the executive time of the
optimization by 98 percent. The optimal parameters were surveyed by GWO to find the minimum value of
the fitness function. Finally, the amount of product with the desired concentration (99/9%) for the 30-stage
square cascade was determined. In this optimal cascade, by consuming 10 kg natural tellurium, 3409 gr 130Te
isotope with 99.9% concentration obtained.

5. REFERENCES
[1] P. D. Manson, Benedict, P. D. Thomas H. Pigford, and P. D. H. W. Levi, “Nuclear Chemical
Engineering, Second Edition,” 1981.
[2] L. Y. Sosnin, A. N. Tcheltsov, A. P. Kuchelev, G. V. Remin, and A. A. Hobotov, “Centrifugal
extraction of highly enriched 120Te and 122Te using the non-steady state method of separation,” in
Nuclear Instruments and Methods in Physics Research, Section A: Accelerators, Spectrometers,
Detectors and Associated Equipment, 2002, vol. 480, no. 1, pp. 36–39, doi: 10.1016/S0168-

8
 9002(01)02044-7.
[3] S. Zeng, L. Cheng, D. Jiang, V. D. Borisevich, and G. A. Sulaberidze, “A numerical method of
cascade analysis and design for multi-component isotope separation,” Chem. Eng. Res. Des., vol. 92,
no. 11, pp. 2649–2658, Nov. 2014, doi: 10.1016/j.cherd.2013.12.016.
[4] F. Ezazi, M. H. Mallah, J. K. Sabet, and A. Norouzi, “A new method for multicomponent mixture
separation cascade optimization using artificial bee colony algorithm,” Progress in Nuclear Energy,
vol. 124. Elsevier Ltd, p. 103371, 01-Jun-2020, doi: 10.1016/j.pnucene.2020.103371.
[5] F. Mansourzadeh, A. A. G. Khamseh, J. Safdari, and A. Norouzi, “Utilization of harmony search
algorithm to optimize a cascade for separating multicomponent mixtures,” Prog. Nucl. Energy, vol.
111, pp. 165–173, Mar. 2019, doi: 10.1016/j.pnucene.2018.11.005.
[6] F. Mansourzadeh, J. Safdari, A. G. H. Khamseh, A. Norouzi, and M. Khajenouri, “Comparison of
optimum tapered cascade and optimal square cascade for separation of xenon isotopes using
enhanced TLBO algorithm,” Sep. Sci. Technol., vol. 53, no. 13, pp. 2074–2087, Sep. 2018, doi:
10.1080/01496395.2018.1443135.
[7] C. Feng et al., “Artificial neural networks combined multi-wavelength transmission spectrum feature
extraction for sensitive identification of waterborne bacteria,” Spectrochim. Acta - Part A Mol.
Biomol. Spectrosc., vol. 251, p. 119423, Apr. 2021, doi: 10.1016/j.saa.2020.119423.
[8] I. Sahin and I. Koyuncu, “Design and implementation of neural networks neurons with radbas,
logsig, and tansig activation functions on FPGA,” Elektron. ir Elektrotechnika, no. 4, pp. 51–54,
2012, doi: 10.5755/j01.eee.120.4.1452.
[9] M. Imani, A. R. Keshtkar, A. Rashidi, J. K. Sabet, and A. Noroozi, “Investigation on the effect of
holdup and cascade shape in NFSW cascades,” Prog. Nucl. Energy, vol. 119, p. 103182, Jan. 2020,
doi: 10.1016/j.pnucene.2019.103182.
[10] S. Zeng, M. Zhou, and C. Ying, “Theoretical and Experimental Study of a Non-stationary Isotope
Separation Process in a Gas Centrifuge Cascade,” Sep. Sci. Technol., vol. 38, no. 11, pp. 2375–2394,
Jan. 2003, doi: 10.1081/SS-120022278.
[11] S. Zeng and C. Ying, “Transient process in gas centrifuge cascades for separation of multicomponent
isotope mixtures,” Sep. Sci. Technol., vol. 36, no. 15, pp. 3439–3457, Nov. 2001, doi: 10.1081/SS-
100107913.
[12] S. Zeng and C. Ying, “Separating isotope components of small abundance,” Sep. Sci. Technol., vol.
37, no. 15, pp. 3577–3598, 2002, doi: 10.1081/SS-120014445.
[13] T. E. Azizov, A. Y. Smirnov, and G. A. Sulaberidze, “Optimization of a Square Cascade of
Centrifuges for Separation of Multicomponent Mixtures of Stable Isotopes,” At. Energy, vol. 128, no.
5, pp. 291–296, Sep. 2020, doi: 10.1007/s10512-020-00690-8.
[14] S. Zeng and C. Ying, “A Second-Order Time-Accurate Method for Determination of Concentration
Distribution of Multicomponent Mixtures in Separation Cascades,” Sep. Sci. Technol., vol. 35, no. 5,
pp. 729–741, Jan. 2000, doi: 10.1081/SS-100100187.
[15] S. Zeng and C. Ying, “A Robust and Efficient Calculation Procedure for Determining Concentration
Distribution of Multicomponent Mixtures,” Sep. Sci. Technol., vol. 35, no. 4, pp. 613–622, Jan. 2000,
doi: 10.1081/SS-100100179.
[16] S. Mirjalili, S. M. Mirjalili, and A. Lewis, “Grey Wolf Optimizer,” Adv. Eng. Softw., vol. 69, pp. 46–
61, Mar. 2014, doi: 10.1016/j.advengsoft.2013.12.007.
[17] A. Naserbegi, M. Aghaie, and A. Zolfaghari, “Implementation of Grey Wolf Optimization (GWO)
algorithm to multi-objective loading pattern optimization of a PWR reactor,” Ann. Nucl. Energy, vol.
148, p. 107703, Dec. 2020, doi: 10.1016/j.anucene.2020.107703.
[18] S. Mirjalili, S. Saremi, S. M. Mirjalili, and L. D. S. Coelho, “Multi-objective grey wolf optimizer: A
novel algorithm for multi-criterion optimization,” Expert Syst. Appl., vol. 47, pp. 106–119, Apr.
2016, doi: 10.1016/j.eswa.2015.10.039.
[19] A. Naserbegi, M. Aghaie, and S. M. Mahmoudi, “PWR core pattern optimization using grey wolf
algorithm based on artificial neural network,” Prog. Nucl. Energy, vol. 129, p. 103505, Nov. 2020,
doi: 10.1016/j.pnucene.2020.103505.
[20] A. Naserbegi and M. Aghaie, “Multi-objective optimization of hybrid nuclear power plant coupled
with multiple effect distillation using gravitational search algorithm based on artificial neural
network,” Therm. Sci. Eng. Prog., vol. 19, p. 100645, Oct. 2020, doi: 10.1016/j.tsep.2020.100645.