See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/357319982

Artificial neural network modeling for potential performance enhancement of

a planar perovskite solar cell with a novel TiO 2 /SnO 2 electron transport
bilayer using nonlinear prog...

Article in Energy Reports · November 2022

DOI: 10.1016/j.egyr.2021.12.010

CITATIONS READS

5 93

3 authors, including:

Innocent Oboh Nsikakabasi Okon

University of Uyo Researcher
53 PUBLICATIONS 718 CITATIONS 1 PUBLICATION 5 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Nsikakabasi Okon on 16 August 2022.

The user has requested enhancement of the downloaded file.

 Energy Reports 8 (2022) 973–988

Contents lists available at ScienceDirect

Energy Reports
journal homepage: www.elsevier.com/locate/egyr

Research paper

Artificial neural network modeling for potential performance

enhancement of a planar perovskite solar cell with a novel TiO2 /SnO2
electron transport bilayer using nonlinear programming
∗
Innocent O. Oboh, Uchechukwu H. Offor, Nsikakabasi D. Okon
Department of Chemical and Petroleum Engineering, University of Uyo, Akwa – Ibom State, Nigeria

article info a b s t r a c t

Article history: The performance of a planar Perovskite Solar Cell (PSC), having a novel TiO2 /SnO2 bilayer as Electron
Received 6 December 2020 Transport Layer (ETL), has been optimized. To achieve this, an artificial neural network (ANN) was
Received in revised form 27 November 2021 trained to predict the Power Conversion Efficiency (PCE) as a function of five technological parameters
Accepted 2 December 2021
of the PSC namely; Active layer thickness, Hole Transport Layer (HTL) thickness, HTL dopant concen-
Available online xxxx
tration, Electron Transport Layer (ETL) thickness, and ETL dopant concentration. The model was able to
Keywords: predict the data accurately, with an overall Mean Square Error (MSE) of 0.0002302, Root Mean Square
Artificial neural network Error of (RMSE) of 0.01502, Sum of Squares Error of (SSE) of 0.4712, and correlation coefficient (R)
Particle Swarm Optimization of 0.9991. Subsequently, the optimum values of independent variables giving the maximum efficiency
SCAPS were determined by applying the Particle Swarm Optimization algorithm to maximize the ANN-derived
Objective function
model. The optimum solutions predicted were then simulated using the SCAPS 1-D program, yielding
Transport Layer
a PCE of 14.94%. This paper presents a straightforward and efficient methodology by which ANN
modeling and nonlinear programming can be applied to carry out multi-property optimization of PSC
performance.
© 2021 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND
license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction buffer) and adding a window layer. One commonality between

these research efforts is the non-systematic approach employed
A deep understanding of the influence of the technological to select the final value of the parameters of interest. In each case,
parameters of PV devices on their performance is necessary if the a numerical analysis on the individual effect of each parameter
economy and performance of the PV devices are to be efficiently on the PCE was carried and then the value of the parameters
optimized. However, it is a rather arduous task to experimentally giving the best PCE was hand-picked. These values are then
analyze and identify the optimum values of device parameters reported as giving an improved or optimum PV performance.
giving maximum efficiency and for this reason, PV analysis and This hand tuning approach of improving the PCE does not take
optimization has been the subject of intensive computational into consideration the fact that the performance of the solar cell
studies. In this regard, Abdelaziz et al. (2019) studied the ef- at any instant depends not individually but collectively on the
fect of technological parameters (such as the thickness, doping parameters considered (Baloch et al., 2017). Hence, this approach
concentration, mobility and affinity of the HTL and ETL layers, cannot satisfactorily probe the feasibility region within which the
optimum point lies and any results obtained merely propose a
defect density and thickness of the active layer, defect density of
pathway to an improved performance and not necessarily the op-
the interface layers and working temperature of the solar cell)
timum performance. This paper offers a methodology for finding
on the PCE of a non-fullerene acceptor solar cell. Based on this
the optimum PCE of a solar cell device by using nonlinear pro-
analysis, it was reported that an improvement in PCE from 10.77%
gramming to vary all the considered input variables concurrently
to 14.25% was possible using a proposed set of optimum values
within technologically feasible and adjustable side constraints.
for the considered technological parameters. Among others, Baig
Planar PSCs possess a feasible and roll to roll manufactur-
(2019) optimized a p-SnS/n-CdS solar cell from a PCE of 1.5%
ing capability. However, they have significantly lower PCE and
to 14% by varying the thickness (of the absorber layer, buffer higher hysteresis compared to their mesoporous counterparts.
layer), minority carrier lifetime, doping concentration (absorber, This is due to defects in the ETL and direct contact between
the perovskite and FTO layers (Ding et al., 2018). Inserting a
∗ Corresponding author. wide band gap metal oxide semiconductor such as (ZnO, SnO
E-mail address: [email protected] (N.D. Okon). and MgO) UTL at ETLs interface has been used successfully to

https://doi.org/10.1016/j.egyr.2021.12.010
2352-4847/© 2021 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
I.O. Oboh, U.H. Offor and N.D. Okon Energy Reports 8 (2022) 973–988

nonlinear, explicit equation was extracted. This was done fol-

List of Symbols lowing the procedure outlined by Abidoye et al. (2018). The
Matlab Matrix Laboratory extracted equation was then applied as the objective function of
the optimization problem. Four different evolutionary, nonlinear
PSC Perovskite solar cell
optimization techniques were considered for optimizing the ob-
R2 Goodness of fit
jective function namely; PSO, GA, PS, and SA. Based on suggested
UTL Ultra-thin layer criteria (Doris et al., 2019), the most promising technique was
SCAPS Solar cell capacitance Simulator selected for probing the optimization search space to obtain the
PSO Particle swarm Optimization optimum values of the input parameters that give the maximum
OSC. Organic Solar Cell PCE.
Voc Open circuit voltage Few similar efforts towards multi-property optimization of
Jsc Short circuit current solar cells by nonlinear programming have been reported. Baloch
PCE Power conversion efficiency et al. (2017) showed that by interfacing Matlab with SCAPS, the
MSE Mean square error available Matlab optimization tools can be deployed to carryout
full space optimization of SCAPS simulation models. However,
RMSE Root mean square error
this model requires deep knowledge of Matlab coding and cannot
CIGS Copper Indium Gallium Selenide
be used for PV devices which cannot be modeled in SCAPS. In
CdTe Cadmium Telluride the present study, although the ANN derived objective func-
fmincon Constrained minimization function tion is obtained using data from SCAPS calculations, in the case
solver of solar cells which cannot be simulated using SCAPS (such as
IPCE Incident photon to current efficiency or Mesoporous solar cells), experimental data can also be used to
QE Quantum Efficiency develop the ANN equation which can then be used as the ob-
SSE Sum square error jective function of the optimization problem. Also, Balin et al.
SRH Shockley–Read–Hall (2019) designed and optimized a VO2 based nanostructured smart
SC Spin coating window by applying Classical trust region algorithm to carryout
constrained optimization of an ANN-derived objective function
CL Compact Layer
which was developed from FDTD calculations data. However,
FF Fill factor
considering only one technique for optimizing PV device is not
FDTD First difference time Domain modeling advisable if the best possible results are desired. In this case, the
deviation of the final predicted values of parameters from their
simulation values was significant. In the present study however,
suitable criteria were employed to decide upon the most suitable
improve band alignment, Kulkarni et al., 2016; Lu et al., 2017b; of four available optimizers in the Matlab optimization toolbox.
Xu et al., 2015), providing a base for better formation of the Based on a careful comparison, the most convincing optimizer
perovskite layer (Kogo et al., 2016), reducing the density of was selected and then by re-defining the constraints, the opti-
the trap states (Lu et al., 2017b), preventing undesirable chem- mum values giving the best tradeoff between the PCE and input
ical interactions (Kulkarni et al., 2016), improving charge ex- variables was obtained. The methodology outlined in this work
traction and suppressing charge recombination (Mohammadian- can potentially lead to more accurate and less time consuming
Sarcheshmeh and Mazloum-Ardakani, 2018; Yang et al., 2016b). optimization of PV devices. The technological parameters such as
Mozhgan et al. (2020) incorporated SnO2 and WO3 as UTLs for the layer thickness and doping concentration have a significant
optimizing TiO2 based planar PSCs. Compared to TiO2 , several influence on the cost and the ease of fabrication of the Solar
advantages have been reported for SnO2 such as faster electron cell. Hence it is safe to say that the provision for flexibility in
extraction, better energy alignment to perovskite and more effi- redefining the side constraints (upper and lower bounds) which
cient electron transfer ability as well as minimizing degradation is offered by the presented approach, allows the selection of the
of PSCs (Christians et al., 2018). WO3 is also a chemically stable most convenient tradeoff between the efficiency, cost and ease of
semi-conductor and compared to TiO2 , indicates higher electron fabrication of the PV device.
mobility (10−20 cm2 V−1 S−1 ), better transparency and a faster
kinetic process of charge transfer at its interface with perovskite 2. Materials and method
layer (Wang et al., 2015). To confirm the effect of the bilayers a
PSC without the UTL was fabricated using the same values for the 2.1. SCAPS
physical parameter as for the optimized counterpart. The results
show that incorporating the UTLs lead to improvement in PCE The simulation was performed using the SCAPS version 3.3.08.
as expected, with the TiO2 /SnO2 bilayer giving better results than It is a 1-dimensional solar cell simulator that has been developed
the TiO2 /WO3 bilayer. A SCAPS model of the fabricated device was by the University of Gent, Belgium to simulate the electrical
also developed and validated in the published work by Mozhgan characteristics (I–V , C –V , C –f , and QE.) of thin-film heterojunc-
et al. (2020). tion solar cells by solving the one-dimensional semiconductor
In this paper, the SCAPS model of the PV device was used to equations in the bulk of the layers. These equations are given by
generate data depicting the joint variation of the PCE with layer (1–3) together with the constitutive equations (4) and (5).
thickness (HTL, ETL and active layer) and doping concentration
(HTL, ETL). The SCAPS 1-D software has been extensively used Poisson’s equation
for the device simulation of CdTe, CIGS, Perovskite and other d2 (ψ) e (
p(x) − n(x) + ND − NA + ρp − ρn
)
materials solar cells giving a good agreement with experimental = (1)
dx2 ε0εr
results and providing accurate data relating considered techno-
logical parameters to the efficiency (Lockinger et al., 2018; Gupta Continuity equation for electrons and holes, (2) and (3) respec-
and Dixit, 2018; Du et al., 2016; Salah et al., 2019; Kephart tively.
et al., 2016; Brajendra et al., 2018; Bauer et al., 2017). The data djn
generated was used to develop an ANN model from which a =G−R (2)
dx
974
I.O. Oboh, U.H. Offor and N.D. Okon Energy Reports 8 (2022) 973–988

Fig. 1. A Schematic diagram of the device structure. Fig. 2. Illuminated J–V characteristics of the Simulated and experimental solar
Source: Adapted from Mozhgan et al. (2020). cell.

Table 1
djp PV simulated parameters compared with Experimental values reported in the
=G−R (3) literature.
dx
Parameters Experimental results Simulation results
Current density equation for electrons
Forward scan Reverse scan
Voc (V) 1.03 1.1 1.03
dn d∅ Jsc (mA/cm2 ) 17.68 19.61 18.57
Jn = Dn + µn n (4) FF (%) 66 32 62.4
dx dx
PCE (%) 11.96 6.97 11.91
Current density equation for holes

dp d∅
Jp = Dp + µp p (5) 2.2. Data collection
dx dx
where p and n are the are free holes and electrons densities
Very little work has been done in this area to provide a
respectively, ND is the ionized-donor density, NA is the ionized-
rich data set showing the collective effect of HTL thickness, ETL
acceptor density, G is the photo-generation rate of electron–hole
thickness, active layer thickness, HTL dopant concentration, and
pairs, R is the recombination rate, ε0 is vacuum permittivity, εr ETL dopant concentration on PCE of the novel PSC device under
is the relative permittivity, ψ is the electrical potential, e is the study. Hence, for our intents and purposes, this paper leverages
electron charge,µn and µp are the electron and hole mobilities, Jn the batch and recorder features of the conventionally accepted
is the electron current density, and Jp is the current density for SCAPS Program. The dataset used for developing the ANN model
holes. in this study was obtained from the recorded batch simulations
Together with appropriate boundary conditions at the inter- of the well-validated SCAPS model of the PSC device. This dataset
faces and contacts, a system of coupled differential equations in comprises 2047 data points. The dataset consists of five input pa-
(ψ , n, and p) is obtained by SCAPS and then a steady-state and rameters, namely HTL thickness (dHTL ), HTL dopant concentration
a small signal solution of this system is computed numerically. (NAHTL ), ETL thickness (dETL ), active layer thickness (dActive layer ),
The software accounts for deep bulk level and interface defect and ETL dopant concentration (NDETL ). The output parameter con-
recombination (non-radiative recombination) in simulated Solar sidered is the PCE. The data ranges for each parameter were fixed
cells. The recombination loss majorly present in the device under based on an analysis of the effect of each parameter on the PCE
consideration is interface recombination which is accounted for (Sections 3.1 and 3.2). A statistical description of the data accrued
by the Shockley–Read–Hall recombination (propagated by defects from SCAPS is presented in Table 2.
or traps). In other to examine the effect of one or a few parameters on
The configuration of the simulated cell is Glass substrate/FTO/ the solar cell characteristics, one can take profit from the SCAPS
TiO2 /SnO2 /Active layer/Spiro-Ometad/Au. TiO2 acts as the (ETL), batch option (Burgelman et al., 2019). When you click ‘Batch set-
Spiro-Ometad as the (HTL), Cs0.05 (MA0.17 FA0.83 )0.95 Pb (I0.83 Br0.17 )3 up’, a panel opens which allows you to select the parameters to
as the active layer, fluorine-doped Tin Oxide (FTO) as the anode be varied and the range over which they are to be varied. Up to
(front contact), and gold (Au) as the cathode (back contact). Fig. 1 nine batch parameters can be defined and varied simultaneously.
shows the main structure of the simulated cell with the different The batch recorder interface is shown in Fig. 3.
layers. During regular single shot or batch calculation, only the detail
The physical and technological parameters of the layers of panels for the last measurement points are available. To examine
the PSC defined in the simulation are reported in detail in the the OSC properties as a function of the batch parameters, you
published work. can launch a record calculation. Clicking ‘Record set-up’ allows
For detailed characterization and verification of solar cell, the you to select the properties which you want to analyze but to
illuminated J–V characteristics were simulated and the experi- launch a recorder calculation (shown in Fig. 4), one needs to
mental results of the compared specimen were plotted against click ‘calculate recorder’; a recorder calculation is then launched.
the corresponding simulation results (shown in Fig. 2) showing Cell parameters are varied according to the Batch set-up, and
an excellent match. only those simulations which are needed to calculate the asked
Also, the Photovoltaic parameters of the experimental solar properties are performed.
cell were compared with that of the simulation in Table 1 and With the Recorder facility, you can do a batch calculation
the results show that the simulated technological and physical and record selected cell properties as a function of the batch
parameter matches their real values excellently. parameter(s). For example, the cell efficiency can be recorded
975
I.O. Oboh, U.H. Offor and N.D. Okon Energy Reports 8 (2022) 973–988

Fig. 3. The batch Recorder set-up panel.

Table 2
A statistical description of the dataset used for developing the ANN.
Statistic Active layer HTL thickness HTL dopant ETL thickness ETL dopant PCE%
thickness µm µm concentration cm−3 µm concentration cm−3
Minimum 0.95 0.08 1 × 1019 0.02 1 × 1019 14
Maximum 1.25 0.15 5 × 1020 0.04 5 × 1020 15.2
Range 0.3 0.07 4.9 × 1020 0.02 4.9 × 1020 1.2
Mean 1.1 0.115 1.65 × 1020 0.03 1.65 × 1020 14.8
Standard 0.093587 0.021833 1.66896 × 1020 0.006244 1.66896 × 1020 0.359132
deviation

An ANN simulation begins with normalization of all inputs and

outputs to the range −1 to 1 before learning their relationship.
This is done using the mapminmax function which is expressed
as,
(ymax − ymin )(x − xmin )
y= + ymin (6)
(xmax − xmin )
Where y is the normalized value of x, ymax is 1, ymin is -1, x is the
actual value of the considered variable, xmin is the minimum value
of the considered variable, and xmax is the maximum value of the
considered variable.
Regarding the learning process, ANN computes the sum of the
products of each input and their corresponding weights, and a
bias term, for each of n number of hidden layer neurons. These
n values obtained are each fed into a transfer function (each of
the same type) to get n outputs for this layer which are passed
on to the next layer (depending on the number of hidden layers)
Fig. 4. Recorder calculations. where the same process is performed until the output layer is
reached. Here, the purelin function computes the sum of the
products of these n values and their corresponding final weights
as a function of some doping concentration NAHTL , and thickness and a final bias term to give the normalized output value. To
dActive layer . (Burgelman et al., 2019). rescale the normalized output to actual value, the result from
Clicking the ‘Calculate: recorder’-button initiates a ‘batch the output layer is de-normalized by reversing the mapminmax
recording’, which can only be done when a batch is set. In this function to give actual output values from the normalized value.
case, SCAPS will ignore the action list which is set on the action The response from each neuron can be expressed mathematically
panel and determine which calculations need to be carried out to as,
record the asked properties. For instance, when a property of the Ni
( )
∑
AC-panel is to be recorded, a C –f simulation will be performed yj = f wji xi + bj (7)
by SCAPS at the frequency determined by the working point or, i=1
if a cell characteristic is to be recorded, an I–V simulation in the
range 0 V to the open-circuit voltage (Voc ), will be carried out Where x are the input parameters, i is the index for input param-
(Burgelman et al., 2019). eters, Ni is the total number of input parameters, j is the index
for hidden layer neurons, bj is the bias for hidden layer neuron j,
2.3. Artificial Neural Networks (ANNs) wji are the weights between the input and hidden layer, f is the
tan-sigmoidal or log-sigmoidal transfer function.
The biologically inspired ANNs are nonlinear, data driven, sta-
⎛ ⎞
Nh
tistical modeling tools designed for problems related to function
∑
z k = fL ⎝ wjk yj + bk ⎠ (8)
approximation and pattern classification. The term ‘‘biologically
j=1
inspired’’ is used for ANNs to describe its approach to problem
solving which is adapted from that of biological neural networks Where k is the index for a number of output parameters, bk is the
which have neurons as functional units (Jain et al., 1996). bias for the output layer, wjk are the weights between the hidden
976
I.O. Oboh, U.H. Offor and N.D. Okon Energy Reports 8 (2022) 973–988

Table 3
Parameter settings for training ANN model.
Parameters Values
Training data set 1433 (70% of data set)
Testing data set 307 (15% of data set)
Validation data set 307 (15% of data set)
Cross-validation data set 243
Number of hidden layers 1
Number of neurons in the hidden layer 1–20
Activation function (hidden layer) Tansig
Activation function (output layer) Purelin
Training Algorithm Levenberg Marquardt
Number of epochs 1000
Learning rate Default
Target goal mean square error 0
Minimum performance gradient 10−17
Maximum validation Checks 6

and output layer, Nh is the total number of hidden layer neurons, optimization problem is given. The general aim of the present
zk are the outputs of the output layer, and fL is the linear transfer study is to propose the optimum set of some technological pa-
function (Fazeli et al., 2013). rameters namely; active layer thickness, HTL thickness, ETL thick-
The neural network is trained by optimization of the weights ness, HTL dopant concentration, and ETL dopant concentration for
for each node interconnection and bias terms to achieve the low- which the PCE is maximized, using nonlinear programming. There
est possible MSE between the predicted output values and their are two main advantages to this approach:
actual values. The MSE of the network is defined by (a) The nonlinear optimization algorithms employed can probe
Eq. (2). the optimization search space for the global optimum more thor-
G Nh
oughly than the hand tuning approach previously applied
1 ∑∑[ ]2 (b) The proposed method allows flexibility in the re-definition of
MSE = yj (a) − Tj (a) (9)
2 the side constraints of the optimization problem.
a=1 j=1
The ANN derived explicit equation was taken as the objective
G is the number of training samples and yj (a) is the predicted function of the maximization problem and then the upper and
output and Tj (a) is the actual output. lower bounds of the input variables (defined in Table 2) were
set as the initial constraints. The optimization problem can be
2.3.1. Development of the ANN model defined as,
The neural fitting toolbox of MATLAB 2015a software which is
Minimize: − f (x)
equipped with early stopping regularization technique was used
to train an ANN to predict the PCE of the PSC device under study. Where: f (x) = f (dActiv elayer , dHTL , dETL , NAHTL , NDETL ) (10)
The neural fitting toolbox screens the data set into three parts: Subject to:
the training set, validation set, and test set. While the training
data set serves to develop the model, the validation data set 0.95 ≤ dactiv elayer ≤ 1.25 (11)
are used to assess the accuracy of the ANN model and avoid 0.08 ≤ dHTL ≤ 0.15 (12)
overfitting, and the testing data are used to assess the network
performance after which no further adjustments are made (Ling 0.02 ≤ dETL ≤ 0.04 (13)
et al., 2017). The default values of stopping criteria such as the 1019 ≤ NAHTL ≤ 5 × 1020 (14)
MSE, number of epochs reached and number of validation checks
1019 ≤ NDETL ≤ 5 × 1020 (15)
were used for the feed-forward, backpropagation network. In his
work only one hidden layer was investigated. Just as in Balin In Eq. (6), f(x) represents the objective or goal function where
et al. (2019) and Chaudhary et al. (2019), it is understood that x is the input vector consisting of the design variables that are
using more than one hidden layer makes the ANN-derived ob- modified to obtain the optimum. The searchable design space is
jective function rather complex and in our case, increases the defined by the upper and lower bounds of the design variables
computational cost of solving the optimization problem greatly. (inequalities (11)–(15)) known as the side constraints. The objec-
The neural fitting toolbox allows cross-validation of the trained tive function is nonlinear. The additive inverse of the objective
network with a new data set thus facilitating further confirmation function is minimized which is mathematically the same as max-
of the generalization capability of the model. The settings used for imizing f (x), (Arora, 2015). Consequently, we take the negative of
the ANN training are listed in Table 3. the resulting optimum function value as the desired maximum.
The raw input data generated form SCAPS was transformed Due to the nature of the ANN-derived equation, we require
using the natural logarithm function before training. This was an optimization algorithm for non-linear, non-continuous, non-
done to improve the accuracy of the ANN-derived model by differentiable, and unconstrained objective functions (with
reducing the skewedness of the data, thus making the pattern in bounds). The MATLAB global optimization Toolbox offers four
the data more interpretable and helping to meet the assumptions algorithms that would suffice to solve such a problem. These are
of inferential analysis (ODSC-Open Data Science, 2019). A plot the Genetic Algorithm (GA), Particle Swarm Algorithm (PSO), Sim-
of efficiency versus HTL dopant concentration before and after ulated Annealing Algorithm (SA), and Pattern-Search Algorithm
logarithmic transformation is shown in Fig. 5. (PS) respectively. The entire optimization procedure employed in
this work is illustrated in Fig. 6.
2.4. Optimization of the ANN derived model
2.4.1. Genetic algorithm
In this section, an overview of the constraints, objective func- GAs are evolutionary optimization algorithms based on Dar-
tion, and the global optimization algorithms considered for the win’s principle of survival of the fittest. In genetic algorithm,
977
I.O. Oboh, U.H. Offor and N.D. Okon Energy Reports 8 (2022) 973–988

Fig. 5. (a) Scatter plots of efficiency as a function of, (a) HTL dopant concentration (b) logarithm of HTL dopant concentration.

each individual (corresponding to a point in the search space).

The individuals who have the highest rank (fittest) are preferen-
tially selected as parents and then the next generation (Children
designs) is then reproduced (created) by randomly mixing the
selected parent designs in a process known as cross-over. The
new generation is again ranked and the process is repeated until
convergence is achieved. The basic steps for implementing the
Genetic algorithm method are outlined below (Arora, 2015).
Algorithm for GA Optimization

Step 1: Generate the initial population P(0) at random.

Step 2: For i = 0, 1, 2, . . .
Evaluate the fitness of each individual in P(i)
Step 3: Select parents form P(i)
Step 4: Apply search operators to the parents and produce gen-
eration P(i + 1)
Step 5: Check for convergence or termination condition

2.4.2. Particle swarm optimization

PSO is inspired by the collective behavior of a group of indi-
viduals such as a swarm of bees, flock of birds or a school of fish.
The algorithm starts with an initial population that is randomly
distributed throughout the search space. The PSO algorithm keeps
track of the position of each individual or particle as well as
that of the population in terms of the objective function value
where the best objective function value of the particle and that
of the group is denoted by pbest and gbest respectively. Each
individual in the group moves with a velocity that is a function
of pbest and gbest and its initial velocity and then the population
converges on the optimum by using information gained from
each individual and from the population as well. The new position
of each individual or particle is then updated from one design
iteration to the next, based on its initial position and velocity.
The value of the objective function is again calculated for new
positions and the PSO is repeated until convergence is attained.
The basic steps for executing the PSO algorithm are summarily
outlined below (Arora, 2015).
Algorithm for Particle Swarm Optimization

Step 1: Initialize imax , w1, ∅1 , ∅2 , n (population size), xi,min , and

xi,max .
Step 2: Initialize the starting position and velocities of the vari-
ables as
k = 1, . . . , n
( )
xi,k = xi,min + xi,max − xi,min ui (16)
Fig. 6. A flow chart showing the methodology employed for optimizing the PSC
in this work.
vi,k = 0

a random population is first initialized. The population is then Step 3: compute

k = 1, . . . , n
( )
ranked, based on the fitness (value of the objective function) of pi,k = f xi,k (17)
978
I.O. Oboh, U.H. Offor and N.D. Okon Energy Reports 8 (2022) 973–988

− new
E −Eold
( )
Step 4: Compute else if e Eold
>u
( )
pbesti,k = pi,k and gbesti = minimum pbesti,k (18) then Eold = Enew (30)
The location of pbesti,k and gbest is given by pxik and gix
else xi+1 = xi − ∆xi (31)
Step 5: Update velocity
Step 7: Go to Step 4 until termination criterion (function not
Vi+1,k = w1 vi,k + ∅1 px,k − xi,k ui + ∅2 gi,x − xi,k ui
( ) ( )
improving for certain number of Iterations) is satisfied
(19) Where xi,min and xi,max are the bounds of the variable xi and ui
Step 6: Update position are random number generated between 0 and 1 (uniform distri-
bution). The energy Eold of this variable is given by its objective
xi+1,k = xi,k + vi+1,k (20) function value f (xi ), the energy state for the new point is Enew ,
∆xi is the perturbation of xi in its neighborhood, and ε is a small
Step 7: Update fitness number fixed at the start of the simulation.
( )
pi+1,k = f xi+1,k (21)
2.4.4. Pattern search
Step 8: If pi+1,k < pbesti,k PS follows the general form of most optimization methods:
given an initial guess at a solution x0 and an initial choice of
then pbesti+1,k = pi+1,k (22) a step length parameter ∆0 > 0. Pattern search methods only
require simple, as opposed to sufficient, decrease on the objective
Step 9: Update function. This weaker condition is possible because we require
that the step be defined by ∆k and the pattern Pk and we place
( )
gbesti+1 = minimum pbesti+1,k (23)
certain mild conditions on both the Exploratory moves and the
Step 10: If i < imax then increment i and go to Step 5, else stop. way in which we update ∆k to guarantee global convergence.
A pattern Pk is defined by two components, a real nonsingular
Where xi,min and xi,max are the bounds of the variable xi and matrix B ∈ Rn×n and an integer generating matrix Ck ∈ Z n×p ,
ui are random numbers generated between 0 and 1 (uniform where p > 2n. In addition, the columns of Ck must contain a core
distribution). Here i, is the iteration number, and wi , ∅1 , and pattern (Mesh) represented by Mk ∈ M ∁Z n×n and its negative
∅2 are the tuning factors of the algorithm. −Mk , where M is a finite set of nonsingular matrices. A pattern
Pk is defined by the columns of the matrix Pk = BCk . Since both
2.4.3. Simulated annealing B and Ck have rank n, the columns of Pk span Rn .The steps are of
As implied by its name, simulated annealing is an optimization the form sk = ∆k Bck , where ck ∈ Ck . The basic steps for executing
technique which is analogous to the process of annealing of solids a General Pattern Search are outlined below (Torczon, 1998).
where a solid is heated and then allowed to slowly cool until
Algorithm for General Pattern Search
the minimum energy of molecules is reached. In comparison, the
optimization problem represents the solid which is transformed Step 1: Initialization: for k = 0, 1, 2,. . .
to an ‘‘ordered state’’ or a desired optimized state. When heated, Step 2: Check for convergence
the molecules are free to move and the freedom gets restricted Step 3: Compute f (x)
as the solid is cooled. In like fashion, SA optimization algorithm Step 4: Determine a step sk using Exploratory Moves (∆k , Pk )
allows ‘‘hill climbing’’ when the temperature is high. That is, these
points that are in the near vicinity of the search point, but have a If f (x) > f (xk + sk ) , then xk+1 = xk + sk . (32a)
higher objective function value can still be selected with certain Otherwise xk+1 = xk . (32b)
probability. This allows the algorithm to escape from local optima
and thus increase the chances of finding a global optimum. The Step 5: Update (∆k , Pk )
basic steps for implementing SA are as follows (Arora, 2015),
2.4.5. Selection criteria
Algorithm for Simulated Annealing To select the most compatible global optimization algorithm
for our problem, four criteria were considered, namely; the ability
Step 1: Initialize ε and variable bounds xi,min and xi,max
of the algorithm to converge to the optimum value, the ability
Step 2: Compute the starting value of the variables as of the algorithm to generate new design alternatives, the com-
putational time required (number of function counts required
( )
xi = xi,min + xi,max − xi,min ui (24)
for convergence), and the reproducibility of the optimum result
Step 3: Compute obtained (Doris et al., 2019). The four algorithms were assessed
based on all four criteria and the highest performing option was
Eold = f (xi ) (25) chosen.
Step 4: Compute It is noteworthy here, that the Genetic algorithm is a con-
strained optimization solver and thus will require linear and/or
∆x1+i = ε xi ui and xi+1 = xi + ∆xi (26) nonlinear constraints along with the upper and lower bounds.
This issue can be addressed by simply using empty square brack-
if ∆x1+i exceeds bounds then ets in those fields where the constraints are required. This implies
that for this problem, such constraints are not applicable. Also, to
( )
xi+1 = xi,min + xi,max − xi,min ui (27)
improve the predicted optimum function value, hybrid functions
Step 5: Compute were used along with the tested algorithms. The hybrid function
is an optimization function that runs after the main optimization
Enew = f (xi+1 ) (28) algorithm terminates and serves to improve the probability that
Step 6: If Enew < Eold the main algorithm converges to a global optimum value. The
hybrid function uses the final point from the main algorithm as
then, Eold = Enew (29) its initial point.
979
I.O. Oboh, U.H. Offor and N.D. Okon Energy Reports 8 (2022) 973–988

Fig. 7. Effect of active layer thickness variation on (a) Jsc (b) Voc (c) FF and (d) PCE.

Table 4
Variation of series resistance and fill factor with active layer thickness recorded
from SCAPS.
d (nm) Rs ( cm−2 ) FF (%)
550 1.20 62.4
610 1.47 62.0
660 1.73 61.7
720 2.00 61.3

3. Results and discussion

3.1. Effect of thickness of active layer on OSC performance Fig. 8. Dark J–V characteristic for various values of active layer thickness.

As the thickness of the active layer is increased from 400 nm

to 2000 nm, Jsc and PCE increase in a similar manner reaching 3.2. Effect of transporting layers characteristics
19.97 mA/cm2 and 12.65% respectively at 675 nm (shown in
Fig. 7a and 7d) and after this point; the rate of increase begins In this section, we examine the effect of thickness and doping
to decline. From Fig. 7d, the PCE begins to drop after reaching concentration, of HTL and ETL on PCE. The results are shown in
a thickness of 1774 nm. This can be explained thus, at first, Fig. 9. The thickness has been varied from 1 nm to 100 nm for
increasing the thickness of the active layer nearly up to 675 nm ETL and from 10 nm to 500 nm for HTL. The results illustrated in
leads to more light absorption and therefore more generation of Fig. 9b and 9c show a slight decrease in the PCE with an increase
hole–electron pairs which enhances Jsc and PCE. At this thickness, in the thickness of both ETL and HTL due to an increase in the
the charge carrier diffusion length is still greater than the active series resistance.
layer thickness and the induced electric field is enough to drive The PCE reaches a peak value of 11.91% at 19.2 nm for HTL and
excess carriers to the corresponding electrodes. However, for ab- a maximum value of 12.03% at 1 nm for ETL. Lower bounds of (80,
sorber layer thicknesses greater than 675 nm, the induced electric 20) for HTL and ETL and upper bounds of (150, 40) for HTL and
field begins to decrease and recombination becomes influential. ETL respectively, are set for optimization purposes. These ranges
There will be greater recombination of charges before they reach are selected to give higher performance and ease of fabrication.
the electrodes thus hindering further increase in Jsc and PCE. Furthermore, the doping concentrations: NA and ND of HTL and
The variation of the dark I–V characteristics for various values ETL, respectively have been varied from 1017 cm−3 to 1021 cm−3 .
of the thickness of the active layer is displayed in Fig. 8. An The results are illustrated in Fig. 9a and Fig. 9d.
increase in thickness causes in a decrease in dark saturation cur- An improvement in PCE is obtained with increasing doping
rent. However, this rate of decrease gradually drops as thickness concentration reaching a maximum value of 13.51% and 12.04%
increases. So, the Voc variation (shown in Fig. 7b) can be said to for HTL and ETL, respectively at a doping concentration of 1021
be dependent on dark saturation current which in turn depends cm−3 . Increasing the doping concentration of the transporting
on the thickness (Abdelaziz et al., 2019). As illustrated in Fig. 7c, layers increases their conductivity. This leads to a decrease in the
as the thickness increases, the series resistance increases which series resistance which in turn leads to a rise in FF and a rise in
reduces the F.F as shown in Table 4. the PCE. Lower bounds of (1019 , 1019 ) for HTL and ETL and upper
For the optimum range of thickness which gives the best bounds of (5 × 1020 , 5 × 1020 ) for HTL and ETL respectively are
performance, a thicker absorber layer absorbs more light which set for optimization purposes. The lower and upper bounds are
leads to the generation of carriers but for thickness values close selected to give a balance between the cost and efficiency of the
to or greater than the diffusion length, there will be more re- analyzed solar cell device.
combination and less collection (requiring careful engineering
to increase chances of carrier collection). Hence, we prescribe 3.3. Assessment of model performance
a thickness range of 950 to 1250 nm for our lower and upper
bound which gives a balance between the diffusion length and The basis for gauging the predicting power of a trained ANN
absorption length. should be its generalization ability. It follows therefore, that the
980
I.O. Oboh, U.H. Offor and N.D. Okon Energy Reports 8 (2022) 973–988

Fig. 9. Variation of Efficiency with (a) HTL dopant concentration (b) HTL Thickness (c) ETL Thickness (d) ETL Dopant concentration.

Fig. 10. The influence of the number of hidden layer neurons on cross validation (a) MSE and (b) Correlation coefficient (R).

relative importance which is in resonance with the theoretically

expected contribution of each variable to the value of the output.
Hence, a sensitivity analysis was conducted on the best results to
further detect weaknesses in the models. The sensitivity analysis
results for several realizations of the 5-7-1 network configuration
were compared with the inferential analysis form Sections 3.1
and 3.2. The ANN with cross-validation MSE of 1.529 × 10−3 .
showed the best match with the expected results and was thus
Fig. 11. The best architecture of the backpropagation network. selected. For this ANN, the training process was truncated at 6
validation checks and 181 epochs. It was therefore concluded,
that the 5-7-1 configuration is the best neural network for the
performance of a particular ANN during cross-validation with un- modeled system since it possesses the best prediction capability.
familiar data should be the benchmark for assessing the quality of The optimal architecture of the ANN model is shown in Fig. 11.
an ANN network. There is no standard technique for choosing the The model predicted values fit quite well with the actual
appropriate number of hidden neurons, so the ANN model was values for both training, testing, and validation sets as can be seen
trained with sequential numbers of hidden neurons and the con- in their correlation coefficients (R) of 0.9992, 0.999, and 0.9987
figuration which gave the best performance upon cross-validation for the training, validation, and testing data, respectively (shown
was selected. in Fig. 12). The ANN derived model can be summarily expressed
In this regard, the number of neurons was varied from 1 by Eq. (33).
through 20 for the single hidden layered network. Each topology ⎧ ⎡ ⎛ ⎞⎤ ⎫
was repeated 30 times to minimize random initialization of the 7 ⎨
∑ ∑5 7
∑ ⎬
weights which causes random correlation. After repeated trials, it η= purelin ⎣LWj,1 ⎝ tansig(xi ∗ IWij + b1 )⎠⎦ + b2
was observed that the lowest MSE and highest R values for cross
⎩ ⎭
j=1 i=1 j=1
validation were produced by the network with seven neurons
(33)
in the hidden layer. This is illustrated in Fig. 10a and 10b. In
Fig. 10a and 10b, the cross validation performance corresponding Eq. (33) represents a trained ANN model correlating the five input
to the best training performance was plotted against sequentially parameters to the PCE (η).
incremented number of hidden layer neurons. The purelin function gives a linear relationship between the
For an ANN model to be useful for optimization, it is only input and the output, with the algorithm being purelin(n) = n,
reasonable that the model should allot to each input variable, a while tansig is a hyperbolic tangent sigmoid transfer function and
981
I.O. Oboh, U.H. Offor and N.D. Okon Energy Reports 8 (2022) 973–988

Fig. 12. Regression plots of ANN predicted PCE versus the actual PCE.

Table 5
List of aggregate weights and biases used for the ANN model in Eq. (33).
Input layer weight matrix Input layer bias Hidden layer Output bias
vector weight vector vector
Wij
j=1 j = 2 j = 3 j = 4 j = 5 b1 W2 b2
28.669052 −0.543831 −17.755598 −1.252405 11.252986 −26.098511 0.087710 −0.299481
12.274599 −0.132103 36.273538 −0.182863 −0.056163 −32.656469 0.046729
10.275873 0.271306 −10.299286 0.666327 −0.144383 −3.883817 −0.035892
3.707060 −0.015362 0.226731 0.144391 −0.818371 0.750819 2.307570
−3.570771 0.013466 −0.236239 −0.151887 0.841071 −0.771435 2.339044
0.068188 0.004852 1.139217 0.016780 −0.020376 0.602967 1.030396
−1.475345 0.004248 −0.228030 0.026589 0.006401 −1.500353 −0.375515

is defined by Eq. (34). SSE are used to assess the performance of the ANN model. This
2 is summarized in Table 6. From Table 6, the assessment is based
Tansig(n) = (34) on the cross-validation MSE. Based on this, a combination of the
(1 + exp (−2∗n)) − 1
lowest obtained MSE, SSE, and RMSE values, as well as the highest
LW and IW are the weights of connections from the input layer R value obtained for cross-validation, makes the model the best
to the hidden layer and from the hidden layer to the input layer, choice.
respectively. The values in Table 5 can be used to predict the PCE
using Eq. (33). 3.4. Sensitivity analysis of the developed model
Furthermore, the value of xi in Eq. (33) represents the individ-
ual data points for each of the input variables where x represents Sensitivity analysis aims to identify how dependent the out-
the input variables namely active layer thickness, HTL thickness, put is, on the value of each input. In this sense, a sensitivity
HTL dopant concentration, ETL thickness, and ETL dopant con- analysis, therefore, helps to spot the weak points of a model
centration; N represents the number of neurons (in this instance (Lawson and Marion, 2008). In comparison to the other variables,
is seven); j represents the number of input variables, (in this the contribution of an independent variable to the prediction of
instance are five) b1 represents the bias of the hidden layer, and the output variable is known as the relative importance of that
b2 represents the bias of the output layer. Using the procedure variable. Various methods exist in the literature for calculating
outlined by Abidoye et al. (2018), a nonlinear equation (which is the relative importance of input variables. Examples include the
quoted in the Appendix A of this work) was extracted using the use of partial derivatives, connection weights algorithm, Lek’s
weights and biases of the developed ANN model. profile method, Garson’s algorithm, etc. However, for this study,
Fig. 13 shows a comparison between actual output values and the connection weights algorithm was employed. The choice is
model-predicted output values. based on the results of previous work in this area, by Olden
From Fig. 13, the model output from ANN shows a good et al. (2004), who compared different techniques for assessing
match with the actual data. This indicates that the model maps the contributions of input variables in ANNs. Their work showed
each point in the search space quite well. However, to measure that of all the methods, the connection weights algorithm was
numerically, how well the model’s output matches the target data the least biased. According to the connection weights algorithm
values, the following statistical models namely R, MSE, RMSE, and as proposed by Olden and Jackson (2002), the relative importance
982
I.O. Oboh, U.H. Offor and N.D. Okon Energy Reports 8 (2022) 973–988

Fig. 13. Comparison between model-predicted efficiency and actual efficiency values.

Table 6
Summary of ANN model performance.
Performance Training Validation Testing Overall Cross-
metric validation
R 0.9992 0.999 0.9987 0.9991 0.9972
MSE 0.0002071 0.0002722 0.00029595 0.0002302 0.001529
RMSE 0.01439 0.0165 0.0172 0.01502 0.03910
SSE 0.4239 0.5572 0.6058 0.4712 0.03715

Table 7
Connection weights products, sums, and rank of inputs.
Active Layer HTL HTL dopant ETL ETL dopant
thickness thickness Concentration Thickness Concentration
Hidden layer 1 2.514588 −0.047699 −1.55736 −0.109845 0.987009
Hidden layer 2 0.573582 −0.006173 1.695032 −0.008545 −0.002624
Hidden layer 3 −0.36883 −0.009737 0.369667 −0.023916 0.005182
Hidden layer 4 8.554303 −0.035450 0.523197 0.333192 −1.888449
Hidden layer 5 −8.35219 0.031498 −1.23826 −0.355270 2.473078
Hidden layer 6 0.070261 0.004999 1.173845 0.017290 −0.020996
Hidden layer 7 0.554014 −0.001595 0.085629 −0.009984 −0.002404
Sum 3.545731 −0.064158 1.051754 −0.157082 1.550797
Rank 1 5 3 4 2

of the input variable x can be computed as the sum of products direction in which each input affects the output. A positive sign
of final weights of the connections from input neurons to hidden indicates that increasing this input variable will increase the
neurons with the connections from hidden neurons to output output parameter, while a negative sign indicates that increasing
neurons for all input neurons. The connection weights from input this input variable will decrease the output variable.
neurons to hidden neurons and, connection weights from the With this understanding, it can be surmised that an increase
hidden to output neurons for all the input variables are presented in active layer thickness, ETL dopant concentration, and HTL
in Table 9. The relative importance of a given input variable can dopant concentration would lead to increased PCE. However, the
be defined as shown in Eq. (35). negative sign for the ETL thickness and HTL thickness indicates
that increasing these variables would lead to a drop in the PCE.
m
∑ These findings are corroborated by the results from the inferential
.RIx = wxy wyz (35)
analysis in Sections 3.1 and 3.2 (see Table 7).
y=1

Where RIx is the relative importance of the input variable x, y is

3.5. Optimization results
the total number of hidden neurons, and z is the output neurons.
The products of the connection weights have been summed and
the rank of the input variables presented in Table 12. A large To assess the suitability of the considered algorithms for the
sensitivity to a variable implies that the system’s performance defined optimization problem, four criteria have been selected
will greatly change with a small variation in the variable and in Section 2.3. The computational time was estimated from the
vice versa. From this comparison, it follows that of all the input number of function evaluations (Function Count) required to
variables, the active layer thickness has the greatest impact on obtain the optimum values. To ascertain the capability of the al-
the PCE followed by the ETL dopant concentration, HTL dopant gorithm to converge to the maximum efficiency, the entire search
concentration, ETL thickness, and the HTL thickness respectively. space was explored and the optimum results were compared in
Furthermore, it must be said here that in using the connection Table 8. Table 8 shows that the G.A and PSO algorithms predicted
weights algorithm, the absolute values are used in determining approximately similar final results higher than the results from
the relative importance of the input variable. The positive or SA and PS although the results from SA required the highest
negative sign of the sum of the products is indicative of the computational time, requiring 3896 function counts and 3770
983
I.O. Oboh, U.H. Offor and N.D. Okon Energy Reports 8 (2022) 973–988

Table 8
Comparison of the optimum results gotten from each algorithm using the entire search space as upper and lower
bounds.
Parameter G.A PSO SA PS
examined (Patternsearch) (fmincon) (Patternsearch) (none required)
NAETL (cm−3 ) 4.61 × 1020 5 × 1020 5 × 1020 5 × 1020
dETL (µm) 0.02 0.02 0.02 0.02
NAHTL (cm−3 ) 2.39 × 1020 3.3 × 1020 5 × 1020 5 × 1020
dHTL (µm) 0.08 0.08 0.08 0.08
dactiv elayer (µm) 1.244 1.25 1.117 1.117
ηmax 15.274 15.287 15.227 15.227
F-Counts (required to 3770 438 3896 556
reach optimum)
No. of iterations 74 12 3770 88

Table 9
Comparison of optimization results with the re-defined optimization problem.
Parameter Side constraints Tested Algorithms
Lower Upper G.A PSO SA PS
bounds bounds (Patternsearch) (fmincon) (Patternsearch) (none required)
NAETL (cm−3 ) 1019 3 × 1020 6.99 × 1019 5.89 × 1019 3 × 1020 3 × 1020
dETL (µm) 0.02 0.04 0.02 0.04 0.04 0.04
NAHTL (cm−3 ) 1019 2 × 1019 1.98 × 1019 2 × 1019 2 × 1019 2 × 1019
dHTL (µm) 0.08 0.15 0.15 0.15 0.15 0.15
dactivelayer (µm) 0.95 1.2 1.2 1.2 1.12 1.2
ηmax – – 14.584 14.588 14.58 14.58
(F-Counts required – – 3850 88 2586 129
to reach optimum)
No. of iterations – – 76 40 2500 20

Table 10
The final results from PSO optimization.
Parameter NAETL (cm−3 ) dETL (µm) NDHTL (cm−3 ) dHTL (µm) dactivelayer (µm) ηmax % f-Counts NoIterations
19
Lower Bounds 10 0.02 1 × 1019 0.08 0.95 – – –
Optimum 1019 0.04 1020 0.148 1.138 14.99 171 32
Values
Upper Bounds 2 × 1020 0.04 1020 0.15 1.15 – – –

iterations to find the global optimum. Subsequently, the ability of fabrication, and PCE, we report a simulated optimum efficiency
of the algorithms to generate new alternatives was assessed. of 14.94%.
One of the key contributions of this study is to show how the
performance of the novel OSC structure can be improved with
3.6. Feasibility of the proposed optimization scheme
flexibility in the re-definition of the multivariable optimization
problem. Therefore, the algorithm chosen must be capable of
generating new design alternatives upon re-definition of the side The major cause of the deviation of observed PCE of solar
constraints. To test for this capability of the algorithms, the upper cells from their theoretically estimated values is the drop in the
bounds of the side constraints were adjusted and the new results estimated Voc upon fabrication, while the J sc usually remains
were compared in Table 9. From Table 9, the PSO algorithm shows settled around the values theoretically estimated (Alharbi and
the most innovative response to the change in the bounds of the Kais, 2015; Alharbi et al., 2015; Hossain et al., 2015). However,
problem. Also, the particle swarm algorithm offers the highest a different trend obtains for PSCs. Jung et al. (2019) reported a
maximum efficiency within the lowest computational time (88 certified Jsc value of 24.92 mA/cm2 which was significantly less
F-Counts). Hence, the particle swarm algorithm is chosen as the than the theoretical value of 29.51 mA/cm2 (Alharbi and Kais,
Algorithm for solving the optimization problem. 2015,?). This drop in J sc is majorly due to the reduced thickness
The PSO algorithm was deployed using MATLAB scripts of of the absorber layer in PSCs to curtail the effects of non-radiative
codes. Further re-definition of the problem was done to obtain the recombination (Johnston and Herz, 2015). These effects occur due
lowest values of the input variables possible without reducing the to decadence in crystal quality and can be mitigated by improv-
efficiency below 15%. The final results are presented in Table 15. ing the growth process quality, maximum charge extraction and
The swarm size was set to 5 and fmincon was used as the hybrid minimum recombination for transport mediums. That is, such
function. effects are non-intrinsic, can be minimized and therefore do not
For all other settings, their default values were used. To give determine the feasibility limits of the PSC PCE. More so, in the
the results a firmer footing, the results were simulated in SCAPS PSC simulation model optimized in this work, the effects of non-
which yielded an Efficiency of 14.94%, Voc = 1.0326 V, radiative recombination are already accounted for (Layer and
Jsc = 22.68 mA/cm2 , and FF = 63.81%. It stands to reason interface defects). Therefore, the simulated optimum values of
that the difference between the PSO predicted PCE in Tables 13 technological parameters and the corresponding PCE reported in
and 15, (from 14.99% to 15.29%) is not worth the correspondingly this study are based on minimum defects state which have been
significant increase in optimum Dopant concentrations of the achieved in practice (Mozhgan et al., 2020).
Transport layers, HTL thickness, and active layer thickness. Hence Comparing the simulated optimized PV parameters with their
to achieve the best tradeoff between the cost of fabrication, ease previous values without optimization (shown in Table 11) it is
easy to see that the rise in PCE is mainly due to the rise in
984
I.O. Oboh, U.H. Offor and N.D. Okon Energy Reports 8 (2022) 973–988

Table 11
A Comparison of the PV parameters of the reported experimental PSC simulation with results from the proposed optimization scheme.
Reference Device structure Voc (V) Jsc (mA/cm2 ) FF % PCE %
Mozhgan et al. (2020) FTO/TiO2 /SnO2 / Perovskite 1.03 18.57 62.4 11.91
Spiro-Ometad/Au
This study FTO/TiO2 /SnO2 / Perovskite 1.033 22.68 63.81 14.94
Spiro-Ometad/Au

J sc and FF. From Section 3, the following effects of the change

in technological parameters on the PSCs PCE can be deduced.
As expected, the increased doping concentrations lead to higher
conductivity which leads to a drop in the series resistance of
the PSC thereby causing the FF to rise. On the other hand, the
increase in the active layer thickness causes an increase in the
series resistance and therefore a drop in FF. In this instance of
a PSC device, the effect of increase in the doping concentrations
more than counterbalances the effect of the increase in active
layer thickness, leading to a slight increase in the FF. More so,
as established in Section 3.1, the rise in J sc value compared to
the already existing device performance can be attributed to
the increased thickness of the absorber layer of the proposed
optimization scheme. Considering all these, the results presented
are considered explainable and feasible for fabrication. Hence
the proposed methodology is well illustrated for the case study. Fig. 14. A comparison of illuminated J- V characteristics of the PSC device before
The major physical and technological parameters alongside the (blue) and after (green) optimization. Calculated by SCAPS.. (For interpretation
optimum parameters of the optimized solar cell are reported in of the references to color in this figure legend, the reader is referred to the web
version of this article.)
the Appendix section of this paper for interested experimentalists
and as for the details of the operating conditions and procedures
for fabrication the solar cell, one can turn to the published work
by Mozhgan et al. (2020). Recommendation
I recommend that there be an experimental campaign provid-
4. Conclusion ing a rich data set that can be used for the intents and purposes
of research work such as this.
The thin-film solar cell structured Glass substrate/FTO/TiO2 /
SnO2 /Perovskite/Spiro-Ometad/Au was optimized using the PSO CRediT authorship contribution statement
algorithm. The photovoltaic performance of the device has been
analyzed and optimized to an efficiency of 14.94% (from an initial Innocent O. Oboh: Conception or design of the work, Method-
value of 11.91%) with ology, Writing - review & editing. Uchechukwu H. Offor: Con-
Voc = 1.033 V, Jsc = 22.68 mA/cm 2 , and F.F = 63.81% for, ception or design of the work, Methodology, Writing - review &
dactiv elayer = 1.138 µm, dETl = 0.04 µm, dHTL = 0.148 µm, editing. Nsikakabasi D. Okon: Conception or design of the work,
NAHTL = 1019 cm−3 , and NDETL = 1020 cm−3 . Data curation, Data analysis and interpretation, Methodology,
Fig. 14 shows the illuminated J–V characteristics of the PSC Writing - original draft, Writing - review & editing.
device before and after optimization.
A similar comparison of the dark J–V characteristics of the PSC Declaration of competing interest
device is given in the appendix section of this work. The approach
employed in this study allows flexibility in the re-definition of the The authors declare that they have no known competing finan-
multivariable optimization problem to accommodate constraints cial interests or personal relationships that could have appeared
like the ease of fabrication and cost of material resources which to influence the work reported in this paper.
are collectively dependent on the layer thickness and dopant
concentration. More so, no additional sets of simulations are Data availability statement
required when re-defining the optimization problem.
The procedure outlined in this work is very fast. Upon the The generated training and cross-validation data, and MATLAB
training of the ANN, the PSO problem converged to the global op- script used for PSO are too long to be quoted directly in this work
timum within a maximum time of 150 ms on a computer (2 GHz and can be provided swiftly upon request. All other data relevant
Intel Core i3-5005U processor with 4 GB RAM). In comparison, to this work are tabulated in the work.
carrying out multi-property optimization solely by simulations
would require several minutes of trial and error analysis. Acknowledgments
Furthermore, the nonlinear programming approach used in
this work can be used not only for the optimization of technolog- The authors would like to appreciate the management of the
ical parameters but also for intrinsic properties of the OSC layers Department of Chemical and Petroleum Engineering, University
such as valence band effective density of states, bandgap energy, of Uyo, Akwa-ibom State, Nigeria for providing an enabling en-
electron affinity, etc. vironment to carry out this research. Also, the authors would
This work will serve as a very helpful precursor for any exper- like to appreciate Dr. Marc Burgleman for promptly providing the
imentalist seeking to fabricate a PSC with optimum performance. current version of SCAPS (version 3.308) for use in this research.

985
I.O. Oboh, U.H. Offor and N.D. Okon Energy Reports 8 (2022) 973–988

Appendix A

See Figs. 15 and 16.

986
I.O. Oboh, U.H. Offor and N.D. Okon Energy Reports 8 (2022) 973–988

Table 12
Material parameters set for simulation of the Solar cell under study at 300 K and with A.M 1.5 G illumination.

S/N Parameters Spiro-Ometad Perovskite TiO2 SnO2 (UTL)

HTL (Active layer) (ETL)

1. d (nm) 148 (Opt) 1138 (Opt) 40 (Opt) 10 (exp)

2. Eg (eV) 2.91 1.5 (Shen et al., 3.2 (exp) 3.5 (Qiu et al.,
(Adhikari et al., 2017) 2018)
2016; Ng et al.,
2018)
3. χ (eV) 2.2 3.93 (Shen et al., 4.89 (exp) 4.69 (exp)
(Adhikari et al., 2017)
2016; Ng et al.,
2018)
4. εr 3 (Shen et al., 6.5 (Da et al., 31 (Berberich and 7.25
2017) 2018) Bell, 1940) (Yildirim et al.,
2014)
5. NA (cm−3 ) 1.0 × 1019 0 (Shen et al., 0 (exp) 0 (exp)
(Opt) 2017)
6. ND (cm−3 ) 0 (Shen et al., 0 (Shen et al., 1.0 × 1020 2.42 × 1019 (exp)
2017) 2017) (Opt)
7. µn (cm2 /V s) 1.0 × 102 2 (Shen et al., 1.0 × 104 2.4 × 102 (Jiang
(Shen et al., 2017) 2017) (Shen et al., 2017) et al., 2017;
Van Zeghbroeck,
1997)
8. µp (cm2 /V s) 2 (Shen et al., 2 (Shen et al., 1.0 × 104 2.2 × 102 (Jiang
2017) 2017) (Shen et al., 2017) et al., 2017;
Van Zeghbroeck,
1997)
9. NC (cm−3 ) 1 × 1020 5.0 × 1018 (Shen 6.67 × 1020 3.61 × 1018
(Shen et al., 2017) et al., 2017) (Kormann et al., (Button et al.,
1988) 1971)
10. NV (cm−3 ) 1 × 1020 (Shen 5.0 × 1018 6.67 × 1020 3.61 × 1018
et al., 2017) (Shen et al., 2017) (Kormann et al., (Button et al.,
1988) 1971)
11. A 1.0 × 104 (fitting) 4.0 × 104 (fitting) 1.0 × 104 1.0 × 104 (fitting)
(1/cm eV( 21 )) (fitting)
12. B 0 (fitting) 0 (fitting) 0 (fitting) 0 (fitting)
(eV( 12 )/cm)

Fig. 16. A plot of the Quantum efficiency (IPCE) against the photon energy of
incident light.
Fig. 15. A comparison of dark J–V characteristics of the PSC device before (red)
and after (blue) Optimization. Calculated by SCAPS.. (For interpretation of the
references to color in this figure legend, the reader is referred to the web version
of this article.) References

Abdelaziz, W., Shaker, A., Abouelatta, M., Zekry, A., 2019. Possible efficiency-
Appendix B. Supplementary data boosting of non-fullerene acceptor solar cell using device simulation. Opt.
Mater 91, 239–245.
Abidoye, L.K., Mahdi, F.M., Idris, M.O., Alabi, O.O., Wahab, A.A., 2018. ANN-
Supplementary material related to this article can be found derived equation and its application in the prediction of dielectric properties
online at https://doi.org/10.1016/j.egyr.2021.12.010. of pure and impure CO2 . J. of Cleaner Prod. 175, 123–132.

987
I.O. Oboh, U.H. Offor and N.D. Okon Energy Reports 8 (2022) 973–988

Adhikari, K.R., Gurung, S., Bhattarai, B.K., Soucase, B.M., 2016. Comparative study Kephart, J.M., McCamy, J.W., Ma, Z., Ganjoo, A., Alamgir, F.M., Sampath, W.S.,
on MAPbI3 based solar cells using different electron-transporting materials. 2016. Band alignment of front contact layers for high-efficiency CdTe solar
Phys. Status Solidi (C) 13, 13–17. cells. Sol. Energy Mater. Sol. Cell 157, 266–275.
Alharbi, F.H., Kais, S., 2015. Theoretical limits of photovoltaics efficiency and Kogo, A., Ikegami, M., Miyasaka, T., 2016. A SnO x –brookite TiO2 bilayer electron
possible improvements by intuitive approaches learned from photosynthesis collector for hysteresis-less high efficiency plastic perovskite solar cells
and quantum coherence. Renew. Sustain. Energy Rev. 43, 1073–1089. fabricated at low process temperature. Chem. Commun..
Alharbi, F.H., et al., 2015. An effcient descriptor model for designing materials Kormann, C., Bahnemann, D.W., Hoffmann, M.R., 1988. Preparation and char-
for solar cells. Npj Comput. Mater. 1 (15003). acterization of quantum-size titanium dioxide. J. Phys. Chem. 92 (18),
Arora, R.K., 2015. Optimization Algorithms and Applications. vol. 14–16, Vikram 5196–5201, 6700.
Sarabhai Space Centre Indian Space Research Organization, Trivandrum, Kulkarni, A., Jena, A.K., Chen, H.W., Sanehira, Y., Ikegami, M., Miyasaka, T., 2016.
India, pp. 139–160. Revealing and reducing the possible recombination loss within TiO2 compact
Baig, Faisal, 2019. A Numerical Analysis for Enhancement of Thin Film Solar layer by incorporating MgO layer in perovskite solar cells. Sol. Energy 136,
Cells. (Ph.D. thesis). Institute of design and fabrication, Valencia. 379–384.
Balin, I., Garmider, V., Long, Y., Abdulhalim, I., 2019. Training artificial neu- Lawson, D., Marion, G., 2008. An introduction to mathematical modeling. https:
ral network for optimization of nanostructured VO2 -based smart window //people.maths.bris.ac.uk/~madjl/course_text.pdf (Accessed 10 June 2019).
performance. Research article. Opt. Express 27 (16). Ling, M., Luo, X., Hu, S., Gu, F., Lytton, R.L., 2017. Numerical modeling and
Baloch, A.A.B., Aly, S.P., Hossain, M.I., et al., 2017. Full space device optimization artificial neural network for predicting J-integral of top-down cracking in
for solar cells. J. Sci. Rep. 7, 11984. http://dx.doi.org/10.1038/s41598-017- asphalt concrete pavement. In: Transportation Research Board 96th Annual
12158-0. meeting (No. 17-05318).
Bauer, A., Sharbati, S., Powalla, M., 2017. Systematic survey of suitable buffer Lockinger, J., Nishiwaki, S., Weiss, T.P., Bissig, B., Romanyuk, Y.E., Buecheler, S.,
and high resistive window layer materials in CuIn1−x Gax Se2 solar cells by Tiwari, A.N., 2018. Tio2 as intermediate buffer layer in Cu(In, Ga)Se2 solar
numerical simulations. Sol. Energy Mater. Sol. Cell 165, 119–127. cells. J. Sol. Energy Mater. Sol. Cell. 174, 397–404.
Berberich, L., Bell, M., 1940. The dielectric properties of the rutile form of TiO2 . Lu, H., Tian, W., Gu, B., Zhu, Y., Li, L., 2017b. TiO2 electron transport bilayer for
J. Appl. Phys 11 (10), 681–692. highly efficient planar perovskite solar cell. Small 13 (38), 1701535.
Brajendra, S.S., Vivek, G., Amitesh, K., Vishnu, A., 2018. Band alignment of Cd-free Mohammadian-Sarcheshmeh, H., Mazloum-Ardakani, M., 2018. Recent advance-
(Zn, Mg)O layer with Cu2 ZnSn (S, Se)4 and its effect on the photovoltaic ments in compact layer development for perovskite solar cells. Heliyon 4
properties. Opt. Mater. 84, 748–756. (11), e00912.
Burgelman, M., Decock, K., Alex, N., Johan, V., Erschraegen, V., Stefaan, D., 2019. Mozhgan, K.O., Mehdi, R., Ahmad, K., Nima, T., Mehran, M., Mozhdeh, F.,
Scaps manual version 3. pp. 1–5, 308. Firouzeh, E., 2020. Enhanced performance of planar perovskite solar cells
Button, K.J., Fonstad, C.G., Dreybrodt, W., 1971. Determination of the electron using TiO2 /Sno 2 and TiO 2 /Wo3 bilayer structures: Roles of the interfacial
masses in stannic oxide by submillimeter cyclotron resonance. Phys. Rev. B layers. J. Solar Energy 208, 697–707.
4 (12), 4539. Ng, C.H., Ripolles, T.S., Hamada, K., Teo, S.H., Lim, H.N., Bisquert, J., Hayase, S.,
Chaudhary, J.K., Liu, J., Skon, J., Chen, Y.W., Kanth, Heikkonen, J., 2019. Optimiza- 2018. Tunable open circuit voltage by engineering inorganic cesium lead
tion of silicon tandem solar cells using artificial neural networks. In: Lecture bromide/iodide perovskite solar cells. Sci. Rep. 8 (1), 2482.
Notes in Computer Science. LNAI 11927, pp. 1–12. http://dx.doi.org/10.1007/ ODSC-Open Data Science, 2019. Transforming skewed data for machine
978-3-030-34885-4_30, 2019. learning. www.medium.com/@ODSC/transforming-skewed-data-for-
Christians, J.A., Schulz, P., Tinkham, J.S., Schloemer, T.H., Harvey, S.P., machine-learning-90e6cc364b0.
de Villers, B.J.T., Sellinger, A., Berry, J.J., Luther, J.M., 2018. Tailored interfaces Olden, J.D., Jackson, D.A., 2002. Illuminating the black box: understanding
of unencapsulated perovskite solar cells for > 1, 000 h operational stability. variable contributions in artificial neural networks. J. Ecol Model 154,
Nat. Energy 3 (1), 68. 135–150.
Da, Y., Xuan, Y., Li, Q., 2018. Quantifying energy losses in planar perovskite solar Olden, J.D., Joy, M.K., Death, R.G., 2004. An accurate comparison of methods for
cells. Sol. Energy Mater. Sol. Cells 174, 206–213. quantifying variable importance in artificial neural networks using simulated
Ding, B., Huang, S.Y., Chu, Q.Q., Li, Y., Li, C.X., Li, C.J., Yang, G.J., 2018. Low data. J. Ecol Model 178, 389–397.
temperature SnO2 -modified TiO 2 yields record efficiency for normal planar Qiu, L., Liu, Z., Ono, L.K., Jiang, Y., Son, D.Y., Hawash, Z., He, S., Qi, Y.,
perovskite solar modules. J. Mater. Chem. A 6 (22), 10233–10242. 2018. Scalable fabrication of stable high-efficiency perovskite solar cells and
Doris, E., Philip, F., Thomas, V., Clemens, M., Steffen, F., Thorsten, P., Martin, S., modules utilizing room temperature sputtered SnO2 electron transport layer.
2019. A systematic approach for the selection of optimization algorithms Adv. Funct. Mater 1806779.
including end-user requirements applied to box-type boom crane design. Salah, M.M., Hassan, K.M., Abouelatta, M., Shaker, A., 2019. A comparative study
Article. Appl. Syst. Innov. 5–7. of different ETMs in perovskite solar cell with inorganic copper iodide as
Du, H., Wei-Chao, W., Jian, Z., 2016. Device simulation of lead-free Ch3 nh3 SnI3 HTM. Optik 178, 958–963.
perovskite solar cells with high efficiency. Chin. Phys. B 25, 108802. Shen, H., Jacobs, D.A., Wu, T., Duong, T., Peng, J., Wen, X., Fu, X., Karuturi, S.K.,
Fazeli, H., Soleimani, R., Ahmadi, M.A., Badrnezhad, R., Mohammadi, A.H., 2013. White, T.P., Weber, K., 2017. Inverted hysteresis in CH3 NH3 PbI3 solar cells:
Experimental study and modelling of ultra-filtration of refinery effluents role of stoichiometry and band alignment. J.Phys. Chem. Lett. 8 (12),
using a hybrid intelligent approach. J. Energy Fuels 27 (6), 3523–3537. 2672–2680.
Gupta, G.K., Dixit, A., 2018. Theoretical studies of single and tandem Cu2 Torczon, Virginia, 1998. Pattern search methods for nonlinear optimization.
ZnSn(S/Se)4 junction solar cells for enhanced efficiency. Opt. Mater 82, SIAG/OPT views and News.6.
11–20. Van Zeghbroeck, B.J., 1997. Bravais lattices. Principles of semiconductor devices.
Hossain, M.I., Alharbi, F.H., Tabet, N., 2015. Copper oxide as inorganic hole Wang, K., Shi, Y., Dong, Q., Li, Y., Wang, S., Yu, X., Wu, M., Ma, T., 2015. Low
transport material for lead halide perovskite based solar cells. Sol. Energy temperature and solution-processed amorphous WOX as electron-selective
120, 370–380. layer for perovskite solar cells. J. Phys. Chem. Lett 6 (5), 755–759.
Jain, A.K., Mao, J., Mohiuddin, K.M., 1996. Artificial neural networks: A tutorial. Xu, X., Zhang, H., Shi, J., Dong, J., Luo, Y., Li, D., Meng, Q., 2015. Highly efficient
Computer 3, 31–44. planar perovskite solar cells with a TiO2 /ZnO electron transport bilayer. J.
Jiang, Q., Zhang, L., Wang, H., Yang, X., Meng, J., Liu, H., Yin, Z., Wu, J., Zhang, X., Mater. Chem. A 3 (38), 19288–19293.
You, J., 2017. Enhanced electron extraction using SnO2 for high-efficiency Yang, G., Tao, H., Qin, P., Ke, W., Fang, G., 2016b. Recent progress in electron
planar-structure CH(NH2 )2 PbI3 -based perovskite solar cells. Nat. Energy 2 transport layers for efficient perovskite solar cells. J. Mater. Chem. A 4 (11),
(1), 16177. 3970–3990.
Johnston, M.B., Herz, L.M., 2015. Hybrid perovskites for photovoltaics: Charge- Yildirim, M.A., Yildirim, S.T., Sakar, E.F., Ates, A., 2014. Synthesis, characterization,
carrier recombination, diffusion, and radiative efciencies. Acc. Chem. Res. 49, and dielectric properties of SnO2 thin films. Spectrochim. Acta Part A Mol.
146–154. Biomol. Spectrosc. 133, 60–65.

988

View publication stats