Artificial neural network modeling of photocatalytic degradation of pollutants: a review of photocatalyst, optimum parameters and model topology

ABSTRACT

In the modern world, wastewater treatment is a critical responsibility for both residential and commercial processes. This article compiled and discussed the photocatalytic degradation of organic pollutants or substances of concern using advanced oxidation processes and various catalysts with reaction conditions and environmental effects. Artificial neural networks (ANN) are also widely used to predict pollutant degradation because they can model nonlinear processes in a time- and cost-efficient manner. This study discusses the different forms of ANNs such as single-layer perceptron (SLP), multi-layer perceptron (MLP), radial basis function (rbf) and recurrent neural networks (RNN) used for predicting the degradation efficiency of the photocatalyst in the given reaction conditions for pollutant removal in textile wastewater treatment. More importantly, this article provides the critical review of the photocatalyst used, the degraded pollutant, the training algorithm, and topology of the ANN model used, as well as the input and output parameters, with a focus on the most influential parameter in the photocatalytic degradation process. This review article aims to provide the reader with a better understanding of the ANN model and its application in the field of photocatalytic degradation process optimization and sensitivity analysis of various process parameters on the degradation rate.

GRAPHICAL ABSTRACT

KEYWORDS:

• Artificial neural network
• photocatalytic reaction
• pollutants degradation
• wastewater treatment

1. Introduction

Although 71% of the earth’s surface is covered by water, the availability of safe and clean water for human consumption is one of the major challenges of our modern society.^[1] Several industries such as textile, pulp and paper and paint industries discharge their effluents into surface waters without further treatment, which leads to various health problems in humans and living beings.^[2,3] Conventional techniques such as biological wastewater treatment, catalytic vacuum distillation, adsorption, flocculation and coagulation, etc. transfer pollutants from the liquid phase to the solid phase, which requires further treatment.^[4–6] Recently, advanced oxidation processes have emerged as an alternative technique to degrade pollutants by converting them into water and carbon dioxide.^[7] Among these advanced oxidation processes, photocatalysis has emerged as a promising sustainable technology for pollutant degradation.^[8,9] Various metal oxide photo catalysts such as Fe₂O₃, SnO₂, TiO₂ and ZnO have been used for photocatalytic applications. Other researchers have shown that doping or mixing two different materials can improve their magnetic, electrical and optical properties, which play an important role in practical applications. For example, Zn/BiOBr, Bi₂WO₆/ZnWO₄, Ag/ZnO, and Sn/ZnO photocatalysts show better photocatalytic activity than their single catalyst.^[10–13]

The photocatalytic degradation process is significantly influenced by several experimental parameters such as pollutant concentration, photocatalyst concentration, pH of the initial solution, reaction temperature, light intensity and oxidant concentration.^[14–16] In the literature, the maximum degradation efficiency was achieved by optimizing the experimental parameters. In practical application, all experimental parameters are interrelated and have a complex nonlinear relationship that cannot be studied with an experimental design. In addition, developing a dynamic model is a significant task to understand the behavior of the process. Recently, researchers have used artificial neural networks as a machine learning technique to model the complex nonlinear relationships between process parameters and their outcomes. This understanding of the nonlinear relationship between the process parameters and relevant outcomes could facilitate the prediction of the process behavior and support decision-making regarding the operation of a reactor for the photocatalytic degradation of pollutants.^[17] ANN is quicker and less expensive than an experimental procedure. Artificial neural networks have been used in a variety of applications.^[18–23] A number of researchers have developed the ANN model for photocatalytic degradation of various pollutants such as maxilon blue 5 G dye, methylene blue, indole, anthraquinone, chloramphenicol, phenol, gaseous styrene, malachite green and erythrosine, etc., with single photocatalysts such as TiO₂ and ZnO.^[24–27] Researchers have also developed a model for composite or doped photocatalysts, such as Ag-doped TiO₂, Cu-TiO₂, Eu-doped GO/KSrPO₄, rGO/Ag₃PO₄/CeO_2, Co₃O₄/TiO₂ on zeolite, BiVO₄/BiPO₄/rGO heterojunctions and MgO-La₂O₃ mixed metal oxides, etc.^[28–33] The ANN model has been developed using a variety of training algorithms, including the Levenberg-Marquardt (LM) backpropagation algorithm, which quickly optimizes the network for small to medium-sized models; the Gradient Descent with Momentum algorithm, which is well-known for speeding up convergence by appending a portion of the previous update to the current one; the Scaled Conjugate Gradient algorithm, which combines conjugate gradient direction with an effective line search; the Bayesian Regularization algorithm, which guards against overfitting by adjusting the model complexity; and the Gradient Descent with Adaptive Learning Rate, which dynamically adjusts the learning rate to optimize the training process.^[17,34,35] Moreover, this progress has also been aided by the genetic algorithms, which optimize the network architecture through evolutionary processes.

The objective of this review is to discuss the applications and fundamentals of ANN in dye treatment. This article also summarizes recent studies on the modeling and optimization of the photocatalytic degradation process of pollutants by using multi-layer perceptron network, radial basis function, and recurrent neural networks. Furthermore, this study provides the detailed discussion of the potential applications of ANNs for the photocatalytic degradation of wastewater, especially phenolic compounds and dyes along with summarizing of the data on the degraded pollutant, the catalyst used, the percentage degradation, the topology of the ANN model, the input and output parameters, the correlation coefficient, and the significant influential parameters affecting the degradation efficiency.

2. Photocatalytic mechanism and kinetics

The photocatalytic reaction occurs under UV light or visible irradiation in presence of various catalysts such as CeO₂, ZnO, CuO, Mn₃O₄, TiO₂, Fe₂O₃, SnO₂, WO₃, SiO₂/TiO₂, ZnO/TiO₂ and MgO/ZnO, etc.^[36–40] The photocatalytic mechanism was carried out by generating a very strong oxidant, e.g. a hydroxyl radical as shown in the following equation:

(1) $\mathrm{Photocatalyst}+\mathrm{hv}\to \mathrm{Photocatalyst}\left(h_{VB}^{+}+{\mathrm{e}}_{\mathrm{C}\mathrm{B}}^{-}\right)$(1)

(2) ${\mathrm{H}}_2\mathrm{O}+{\mathrm{H}}_{VB}^{+}\to {\hspace{0.17em}}_{\hspace{0.17em}}^{\circ }\mathit{OH}+{\mathrm{H}}^{+}$(2)

(3) ${\mathrm{O}}_2+{\mathrm{e}}_{\mathrm{C}\mathrm{B}}^{-}\to {\mathrm{O}}_2^{\circ -}$(3)

(4) ${\mathrm{O}}_2^{\circ -}+{\mathrm{H}}^{+}{\leftrightarrow }^{\circ }\mathit{OOH}$(4)

(5) $2{\mathrm{HOO}}^{\circ }\to {\mathrm{H}}_2{\mathrm{O}}_2+{\mathrm{O}}_2$(5)

(6) ${\mathrm{H}}_2{\mathrm{O}}_2\to 2^{\circ }\mathit{OH}$(6)

Generated •OH and ${\mathrm{O}}_2^{.-}$ radicals subsequently react with the organic pollutants and degrade them to final products as follows:

(7) $\mathrm{organic}\mathrm{pollutant}+{\mathrm{OH}}^{\circ }\to \mathrm{Degraded}\mathrm{products}$(7)

(8) $\mathrm{organic}\mathrm{pollutants}+{\mathrm{O}}_2^{\circ -}\to \mathrm{Degraded}\mathrm{products}$(8)

After the excitation of the photocatalyst by light absorption, photoexcitation of electrons occurs and an electron-hole pair is generated to attack the pollutant.^[41,42] The transfer of electrons from the valence band to the conduction band enables the creation of holes in the valence band. According to Equation 2(2) ${\mathrm{H}}_2\mathrm{O}+{\mathrm{H}}_{VB}^{+}\to {\hspace{0.17em}}_{\hspace{0.17em}}^{\circ }\mathit{OH}+{\mathrm{H}}^{+}$(2) , the holes that are produced in the valence band oxidize the adsorbed water to hydroxyl radicals. Similarly, the generated electrons reduce the dissolved oxygen according to Equation 3(3) ${\mathrm{O}}_2+{\mathrm{e}}_{\mathrm{C}\mathrm{B}}^{-}\to {\mathrm{O}}_2^{\circ -}$(3) into superoxide radical anions, which generate hydrogen peroxide according to equations (4)(4) ${\mathrm{O}}_2^{\circ -}+{\mathrm{H}}^{+}{\leftrightarrow }^{\circ }\mathit{OOH}$(4) and (5(5) $2{\mathrm{HOO}}^{\circ }\to {\mathrm{H}}_2{\mathrm{O}}_2+{\mathrm{O}}_2$(5) ). The generated hydrogen peroxide further dissociates into hydroxyl radicals.^[43,44] Finally, the hydroxyl and superoxide radicals formed degraded the pollutant via Equation 4(4) ${\mathrm{O}}_2^{\circ -}+{\mathrm{H}}^{+}{\leftrightarrow }^{\circ }\mathit{OOH}$(4) and Equation 5(5) $2{\mathrm{HOO}}^{\circ }\to {\mathrm{H}}_2{\mathrm{O}}_2+{\mathrm{O}}_2$(5) to the final degradation products shown in Figure 1.

Figure 1. Photocatalytic mechanism of individual TiO₂ semiconductor photocatalyst under natural sunlight.^[43]

3. Fundamentals, Classification and Experimental Design of ANN

3.1. Artificial Neural Networks

ANN is a computational model that emulates the human nervous system in order to identify information, relationships and recognize patterns using an appropriate data set and training algorithm.^[45,46] Once the network has been trained with sufficient sample data, predictions can be made for a new input with relatively similar patterns based on prior learning.^[47,48] The ANNs have enormous potential in wastewater processing research due to their outstanding features of adaptability, non-linearity, self-learning ability, fault tolerance and continuous advancement in input to an output mapping. The trained neural networks serve as an analytical tool for the prognosis of the actual results under the influence of various operational parameters. This helps in the optimization of the operational parameters to achieve the desired outcomes. ANN has been used in wastewater treatment for modeling, prediction and optimization of pollutant removal. The artificial neuron is the basic building block of ANN, and each neuron has a transfer function, weighted input and a single output. A general neural network architecture has three layers of neurons: input layers (receive information), one or more hidden layers (perform most of the internal processing, responsible for extracting patterns), and output layers (generate and present final network outputs)^[49] as shown in Figure 2. The weighted sum of the inputs that each neuron receives activates each neuron, and the activation signal then travels via a transfer function to create a single output, as shown in Figure 2. The neural network’s overall behavior is determined by its architecture, learning rule and transfer functions.^[50,51]

Figure 2. Classification of Artificial Neural Networks.

3.2. Classification of ANNs

The feed-forward and recurrent neural networks are the two neural networks that are most frequently utilized in wastewater treatment processes as shown in Figure 2.^[52,53] A feed-forward neural network is the basic type of neural network, in which data travel in one direction, through the neural nodes and through the output nodes without receiving feedback. In contrast, recurrent neural networks use at least one feedback loop that uses outputs as inputs to the feedback process.

3.2.1. Feed-forward Neural Networks

Feed-forward neural networks are a class of neural networks that have no cyclic connection between neurons, and the flow of the information is only in the forward direction from the input layer to the output layer. The weights in the feed-forward neural networks are static as they support unidirectional forward propagation and no backward propagation. The weighted inputs will feed the activation functions, which can be used to classify or step activation functions. A feed-forward neural network can be multi-layered (radial basis function or multi-layer perceptron) or single-layered (perceptron). The feed-forward neural networks are suitable for noisy data and are very easy to maintain.^[54–56]

3.2.1.1. Single-layer perceptron (SLP)

The simplest artificial neural network is a single-layer perceptron based on a threshold transfer function. It can classify only linearly separable cases with a binary target (1,0). The local memory of the neuron will store a vector of weights for different inputs. The computation is performed to calculate the weighted sum of the input vector. This value will be fed to the activation function, and the classification will be provided based on the threshold value. The threshold and weight vectors are assigned randomly. The weight adjustment is done using Equation 9(9) $\mathrm{\Delta }\mathit{w}=\mathit{\eta }\times \mathit{d}\times \mathit{x}$(9) during the training phase.^[57]

(9) $\mathrm{\Delta }\mathit{w}=\mathit{\eta }\times \mathit{d}\times \mathit{x}$(9)

Where, d, η and x represent the difference between the predicted and desired output, the learning rate and input data, respectively.

3.2.1.2. Multi-layered feed-forward neural networks

The multi-layered feedforward networks consist of multiple layers of computational units interconnected in a feed-forward way. There are mainly two types of multi-layer feed-forward neural networks: multi-layer perceptron (MLP) and radial basis function (RBF). These two mainly differ in the activation function’s choice and the structure of their hidden layers.

A multi-layer perceptron is characterized by its use of nonlinear activation functions, which can vary but often include hyperbolic tangent, sigmoid or ReLU (rectified linear unit). MLPs can have multiple hidden layers, allowing them to model complex functions. The multi-layer perceptron model extends the SLP model by adding one or more hidden layers. It employs a backpropagation algorithm for training, which adjusts the weights of the network based on the error of the output during the training process. The designers of an MLP must set the number of hidden layers and the number of neurons per hidden layer, which are hyperparameters that affect the network’s capacity and performance.^[57–59] The output of the neurons in the MLP is typically computed using Equation 10(10) $Y_k\left(\mathit{n}\right)={\sum }_i\mathit{f}(W_{\mathit{in}}\cdot Y_{\mathit{n}-1}+b_{\mathit{in}})$(10) as follows:

(10) $Y_k\left(\mathit{n}\right)={\sum }_i\mathit{f}(W_{\mathit{in}}\cdot Y_{\mathit{n}-1}+b_{\mathit{in}})$(10)

Where, $Y_k\left(\mathit{n}\right)$ is the output of the k^th neuron in the n^th layer, $W_{\mathit{in}}$ is the weights connecting the (n-1)^th layer to the n^th layer, $Y_{\mathit{n}-1}$ is the output vector of the previous layer, $b_{\mathit{in}}$ is the bias term for n^th layer, and f is the activation function, which can vary based on the network design.

The structure of RBF network is very similar to MLP in that it includes an input layer, a hidden layer, and an output layer. However, RBF typically has only one hidden layer, and it utilizes a Gaussian function to transfer the input space into a higher-dimensional space. The neurons of hidden layer of RBF are activated by the similarity between the input vector and their respective prototypes, which are determined by a process such as k-means clustering. The outputs can be achieved by Equation 11(11) $y_k={\sum }_{j=1}^L{W}_{\mathit{j}\cdot }{f}_j={\sum }_{j=1}^L{W}_j\cdot exp\left(-\frac{\parallel x-C_j{\parallel }^2}{2{\sigma }^2}\right)$(11) as follows:

(11) $y_k={\sum }_{j=1}^L{W}_{\mathit{j}\cdot }{f}_j={\sum }_{j=1}^L{W}_j\cdot exp\left(-\frac{\parallel x-C_j{\parallel }^2}{2{\sigma }^2}\right)$(11)

Where L is the number of neurons in the hidden layer, W_j is the weights of the output neurons, x is the input vector, C_j is the prototype (center) of the j^th neuron in the hidden layer, $\mathit{x}-C_j$ is the Euclidean distance between the input vector and the center of the j^th neuron, and σ is a parameter that determines the width of the gaussian function.

3.2.2. Recurrent Neural Networks (RNN)

Recurrent neural networks are an ANN, where connections between different neurons form a directed graph as shown in Figure 2. RNN are distinguished from other neural networks that use their memory to influence the current input and output using information from the previous inputs. By this RNN, we are able to identify the sequential traits and make use of patterns to forecast the likely scenario. Thus, they are suitable to model sequential data as they allow information from an unlimited number of previous steps to persist. RNNs standardized different activation functions, weights and biases to ensure that each hidden layer has the same characteristics in contrast to feed-forward neural networks. So, instead of having multiple hidden layers, it will create only one and loop over it as many times as necessary.

LSTM is the most popular RNN architecture that can keep information for a long time by default and easily learn long term dependencies. Unlike traditional RNN architectures, LSTM forgets and remembers things selectively. It uses a memory cell and a set of gates to control what information from the last sample is held in memory (input gate), the amount of data passed to the next layer (output gate) and the stored memory’s tear-down rate (forget gate).^[60–62]

4. ANN modeling of photocatalytic degradation of wastewater

The accuracy of the validation, training, and testing was investigated by the coefficient of determination (R²) as shown in Equation 12(12) $R^2=1-\left[\frac{{\sum }_{i=1}^n{\left(y_i-{\hat{y}}_{\mathit{i}}\right)}^2}{{\sum }_{i=1}^n{\left(y_i-{\bar{y}}_{\mathit{i}}\right)}^2}\right]$(12) Validation was performed by providing a new set of input data to the network. To evaluate the efficiency of the predicted responses, additional statistical analyzes such as RMSE and others were usually performed. The phases of implementation of ANN are shown in Figure 3.

Figure 3. ANN implementation flowchart of the dye degradation.

(12) $R^2=1-\left[\frac{{\sum }_{i=1}^n{\left(y_i-{\hat{y}}_{\mathit{i}}\right)}^2}{{\sum }_{i=1}^n{\left(y_i-{\bar{y}}_{\mathit{i}}\right)}^2}\right]$(12)

Here, the predicted, actual and average values are represented as ${\hat{y}}_{\mathit{i}}$, $y_i$, and ${\bar{y}}_{\mathit{i}},$ respectively and n expresses the number of data points.

Various photocatalysts like pure TiO₂; metal-based TiO₂ photocatalysts, such as Ag±doped TiO₂, Fe/TiO₂, Cu-TiO₂, TiO₂/GAC, MWCNTs – TiO₂, etc, pure ZnO, metal-based ZnO photocatalysts, such as Fe/ZnO, pumice-supported ZnO and manganese doped ZnO; and composite materials such as Fe/TiO₂–ZnO, Sn/Zn-TiO₂ and sulfur – nitrogen co-doped Fe₂O₃, etc., have been used for the photocatalytic degradation of several pollutants such as dyes: MB, indole, acid red 27 dye, malachite green dye, acridine orange dye, reactive black 5 acid red 114 C. I. Acid and Blue 9, etc. as well as colorless pollutants such as phenol, nitrophenol and 4-chlorophenol, etc.^[63–65] In the following section, a brief discussion of ANN modeling of the photocatalytic degradation process by different photocatalysts have been presented and shown in Figure 4.

Figure 4. Summary of catalysts and dyes used in ANN models research analysis.^[62–71]

4.1. Pure TiO₂ photocatalyst

Behnajady et al.^[65] developed an ANN model to show the effects of experimental parameters on the photocatalytic degradation of TiO₂ nanoparticles under UV light irradiation. Acid Red 27 was selected as the pollutant for photocatalytic degradation. TiO₂ nanoparticles were synthesized by the sol-gel method. The experimental parameters such as reflux temperature, molar ratio of titanium alkoxide, stirring speed, reflux time, and gelation pH were varied to optimize the reaction rate constants. In this study, the LM back-propagation algorithm was used for training. The analysis showed good agreement with the experimental data, a root mean square error (RMSE) value of 0.0015 with a correlation coefficient (R²) of 0.965 for an ANN structure of 5-12-1 as shown in Figure 5a. The relative importance of the five synthesis variables was analyzed and the result showed that the reflux time was the most influential parameter for the photocatalytic degradation of Acid Red 27 dye as shown in Figure 5b.^[65]

Figure 5. (a) Experimental data correlated with MSE and; (b) Relative importance of the parameters for degradation of Acid Red 27 dye.^[62]

Chandrika et al.^[66] proposed an ANN model to predict the photocatalytic degradation of Malachite Green (MG) dye using TiO₂ photocatalyst. The TiO₂ was prepared by sol-gel synthesis process. The optimum condition was identified by varying the experimental parameters such as initial 10 mL of dye concentration 10–40 ppm, catalyst loading 1–3 mg and reaction time 20–60 min. Under the optimum condition of 3 mg TiO₂ nanoparticle, 60 min of reaction time and 20 ppm initial dye concentration, the photocatalytic degradation efficiency was 90%. The back-propagation algorithm was used for ANN modeling and the optimum efficiency was predicted with a topology of 3-6-1 with an R² value of 0.9707.^[66] Dutta et al.^[34] developed an ANN model for photocatalytic degradation of reactive black 5 diazo dye by TiO₂ photocatalyst under UV light irradiation. The TiO₂ photocatalyst was purchased commercially and used for the photocatalytic reaction. The degradation efficiency was optimized by varying pH with different TiO₂ dosages and initial dye concentrations. At optimal conditions, a photocatalytic degradation efficiency of 90% was observed for reactive black dye. The ANN model was developed using the LM backpropagation algorithm with a ‘“logsig” hidden layer transfer function and shows very good accordance with the experimental data with a R² of 0.999.^[34] Garg et al.^[67] photodegraded the aqueous dye Acid Red 114 (AR114) with a commercial TiO₂ catalyst under UV light irradiation. RSM was used to optimize the photocatalytic degradation process. In this study, an ANN with the LM algorithm with a four-layer learning algorithm (back-propagation) was used. The experimental input parameters were TiO₂ dose, solution pH, initial AR114 concentration, time and area/volume, and UV light intensity. The output was the degradation and decolorization efficiency of AR114. The degradation efficiency was 100% within 150 min of the light irradiation. The ANN model structure of 6:7:2:2 with a relative correlation coefficient of 0.998 shows a very good prediction of the experimental data. The relative importance analysis of the input parameters shows that the catalyst dose is the most important parameter for both decolorization/degradation effectiveness.^[67] Ghanbary et al.^[68] have designed an ANN model for the degradation of 4-nitrophenol (4-NP) by titanium dioxide (TiO₂) nanoparticles under UV light irradiation. The TiO₂ nanoparticles were synthesized by the sol-gel method. The photocatalytic degradation efficiency was optimized to maximize the degradation efficiency by varying the experimental parameters, e.g. TiO₂ loading, UV light intensity, irradiation time, and initial concentration of 4-NP. At optimum conditions, 90% degradation was observed within a reaction time of 100 min. The predicted ANN model with topology 4:14:1 shows a very good agreement with the experimental data with a correlation correction factor of 0.9925. The relative importance analysis shows that the light irradiation time has the maximum influence on the degradation of 4-NP.^[68] Zulfiqar et al.^[69] designed an ANN model for photocatalytic degradation of phenol by TiO₂ nanoparticles under UV light irradiation. The photocatalytic degradation efficiency was maximized by varying the experimental parameters such as catalyst dose, phenol concentration, degradation time and pH of the initial solution. Under optimal conditions, a TiO₂ dosage of 1.75 g/L, a phenol concentration of 15.21 mg/L, a pH of 5.42 and an irradiation time of 540 min, the photocatalytic degradation efficiency was ~ 99.48%. The 4-8-1 structure ANN model with Levenberg- Marquardt algorithm showed a very good agreement with the experimental data with an R² ~0.999.^[69] Boutra et al.^[70] developed an ANN model with a Bayesian regularization algorithm for the photocatalytic degradation of solophenyl brown AGL dye and paracetamol by using a titanium dioxide photocatalyst under sunlight irradiation. The photocatalytic experiments were optimized using the response surface method by varying experimental parameters such as titanium dioxide dose and solution pH and pollutants concentration. Under the optimal conditions of a catalyst loading of 0.7 g/L, an initial concentration ratio of pollutants of 0.34 and a solution of pH 6.5, the degradation rate was ~ 99%. The ANN model with the optimal topology (3-4-3) gave a correlation coefficient of 0.99 for the co-degradation yield, showing excellent agreement between the predicted and observed values. The relative importance of the experimental factors was investigated and the result shows that the pollutant concentration has the greatest influence (~43%) on the photocatalytic degradation of the pollutant.^[70] Zulfiqar et al.^[71] designed an ANN model to predict the experimental data for photocatalytic degradation of Acid Orange 7 (AO7) by TiO₂-P25 catalyst. The effect of experimental parameters such as catalyst dose, pH and dye concentration on photocatalytic degradation of dye was systematically investigated. At the optimum conditions of pH 7.59, 0.748 g/L of catalyst and 28.483 mg/L of AO7, the degradation efficiency was ~ 94.97%. An ANN model with topology 3-8-1 was designed using the LM algorithm, and the result shows very good agreement with the experimental data with a correlation coefficient of 0.999.^[71] Khataee,^[72] developed an ANN model for the photocatalytic degradation of C.I. Basic Red 46 (BR46) dye by TiO₂ nanoparticles under UV-C irradiation. The effects of various operating factors such as BR46 concentration, UV light intensity and initial pH on photocatalytic decolorization efficiency were analyzed. The effect of inorganic anions such as carbonate, bicarbonate, sulfate, chloride and phosphate on the degradation of BR46 was studied. The ANN modeling structure of 4-8-1 with a backpropagation algorithm was developed for analysis. Under optimal conditions, 75% degradation was observed and the predicted ANN model with a correlation coefficient of 0.96 showed good agreement with the experimental data. The study also showed the influence of the experimental parameters on the degradation efficiency. The result showed that the initial dye concentration is the most important parameter.^[72] Oliveros et al.^[73] constructed an ANN model for the photocatalytic degradation of 2,4-dihydroxybenzoic acid by TiO₂ catalyst under UV light irradiation.^[73] Berkani et al.^[64] designed an ANN model for an industrial azo dye C.I. Basic Red 46 degradation by TiO₂ catalyst.

The experiment was conducted in a semi-pilot scale photocatalytic reactor under natural solar irradiation as shown in Figure 6. The effects of various experimental parameters such as initial dye concentration, pH and flow rate on photochemical degradation were systematically investigated. Under the optimal conditions of a dye concentration of 10.65 mg/L, pH of 10.82 and liquid flow rate of 852 L/h, the result shows a maximum degradation efficiency of 99%. An ANN model with topology 3-4-1 was developed and the result showed that the predicted model agrees very well with the experimental data, with a correlation coefficient of 0.999.^[64] The ANN model predicts the R² values in the presence of pure TiO₂ catalysts with various dyes, which are listed in Table 1. The Figure 8a provides the comparison of R² value obtained from various dye degradation processes in the presence of pure TiO₂ catalysts.

Figure 6. Experimental set-up of photocatalytic reactor under natural sunlight. (a) Schematic diagram and (b) Actual experimental Set-up. Copyrights reserved to Springer.^[64]

Table 1. ANN modeling for photocatalytic degradation of different toxic dyes and phenols by pure TiO₂ catalyst.

Pollutant Degraded	Catalyst, Synthesis method	Light Source	Degradation %, Time (min)	Most influential parameters	ANN Structure	Validation, R^2	References
Reactive black 5	TiO2, purchased commercially	UV light	90%, 120	Dye concentration	3-10-1	0.993	^[Citation31]
C.I Basic Red 46	TiO2, provided by Ahlstrom Company	Natural solar radiation.	99%, 150	Initial dye concentration	3-4-1	0.999	^[Citation64]
Acid Red 27	TiO2, sol–gel synthesis	UV-C	180	Reflux Time	5-12-1	0.966	^[Citation65]
Malachite green dye	TiO2, sol–gel synthesis	UV light	90%, 60	Initial dye concentration	3-6-1	0.971	^[Citation66]
Acid Red 114	TiO2, purchased commercially	UV light	~100%, 150	Catalyst concentration	6:7:2:2	0.998	^[Citation67]
4-nitrophenol	TiO2, sol-gel method	UV light	~90%, 100	Light Irradiation time	4:14:1	0.992	^[Citation68]
Phenol	Commercial TiO2	UV light	98.94%, 540	–	4-8-1	0.999	^[Citation69]
Solophenyl brown AGL	Commercial TiO2	Visible light	99%, 210	Pollutant’s concentration ratio	3-4-3	0.990	^[Citation70]
Acid orange 7	TiO2-P25, purchased commercially	UV light	~94.97%	–	3-8-1	0.999	^[Citation71]
C. I. Basic Red 46	Commercial TiO2	UV-C	~75%, 70	initial concentration of the dye	4-8-1	0.96	^[Citation72]
Phenol	TiO2, purchased commercially	UV light	88.3%, 300	–	5-6-4-2	0.99	^[Citation74]

4.2. Coupled TiO₂ photocatalyst

Razvarz and Jafari^[75] studied photodegradation of the Color Index Acid Yellow 23 dye (C.I. AY23) using Ag-TiO₂ photocatalysts under UV light irradiation. The experimental input parameters of dye concentration, UV light intensity, irradiation time and initial dose of nano Ag-TiO₂ were varied to optimize the maximum degradation efficiency. Under the optimal conditions, the result shows ~ 100% dye degradation at 60 min light irradiation. An ANN model with topology 4-5-1 was developed to predict the experimental data and the result shows very good agreement with a correlation coefficient of 1.00.^[75] Eskandarlo et al.^[76] studied the photocatalytic degradation of the dye acid fuchsin (AF) by coupled TiO₂/NiO nanoparticles synthesized by an impregnation method. The effects of various additives such as inorganic anions (e.g., CO₃²⁻, CH₃COO⁻, H₂PO₄⁻, Cl⁻, NO₃⁻, and SO₄²⁻), inorganic oxidants (e.g. IO₄⁻, HSO₅⁻, ClO₃⁻, H₂O₂, S₂O₈²⁻, BrO₃⁻ and S₂O₈²⁻) and transition-metal ions (e.g. Fe²⁺, Zn²⁺, Ni²⁺, Cu²⁺, Co²⁺, and Mn²⁺) on the photocatalytic degradation of AF were studied. An ANN model was developed to model the photocatalytic degradation process of AF dye. The analysis displayed that the model-predicted data is in an agreement with the experimental data. The relative importance showed that H₂PO₄⁻, Cu²⁺ and IO₄⁻ have maximum influence on the photocatalytic degradation for transition-metal ions, inorganic oxidants and inorganic anions, respectively.^[76] Khataee et al.^[77] designed an ANN model to predict the photocatalytic degradation of phenol by commercially available TiO₂ nanoparticles under UV light irradiation in combination with a photoelectron-Fenton-like process using Mn²⁺ cations as electro-Fenton catalyst. The degradation efficiency was optimized by varying experimental parameters such as initial phenol concentration, applied current, pH of the solution, type of UV light and Mn²⁺ concentration and the phenol oxidation efficiency. The ANN model used a genetic algorithm to find the optimal degradation of phenol. ANN structure of 5-12-1 with a correlation coefficient of 0.949 shows a very good prediction of the experimental value. The relative importance analysis shows that the Mn²⁺ concentration has the maximum influence on the degradation of phenol.^[77] Lenzi et al.^[24] implemented an ANN model with a backpropagation learning algorithm for photocatalytic degradation of Maxilon blue 5 G dye by ZnO, TiO₂, TiO₂–ZnO-based catalyst under UV light irradiation. The Fe/ZnO, Fe/TiO₂, and Fe/TiO₂-ZnO photocatalysts were prepared by the impregnation method, and the basic ZnO and TiO₂ were purchased commercially. The result shows that the Maxilon blue 5 G dye was completely degraded by all these catalysts within 180 min. The correlation coefficient of Fe/TiO₂, Fe/ZnO, and Fe/TiO₂-ZnO are 0.982, 0.999, and 0.981, respectively. The analysis shows that the ANN model with a correlation coefficient of 0.999 agrees well with the experimental.^[24] Mohammadi et al.^[25] designed an ANN model for decolorization of methylene blue and methyl orange using UV light irradiation. The Sn/Zn-TiO₂ photocatalyst was synthesized by the sol gel method. The experimental input parameters, e.g. dye concentration, catalyst amount, solution temperature and reaction time, etc. were varied to optimize the result (Figure 7). Under the optimum conditions, a degradation of 88.1% and 93.7% were observed for methyl orange (MO) and methylene blue (MB), respectively. To validate the experimental data, an ANN model 4-8-1 was constructed using a three-layer back-propagation algorithm. The result shows that the designed ANN model agrees with the experimental data with a coefficient performance of 0.983 and 0.988 for the MB and MO dyes, respectively. Reaction time is the most influential characteristic, with a relative value of 38% and 43% for MO and MB dyes, respectively, according to an analysis of the other major experimental parameters.^[25] Ayodele et al.^[29] employed an artificial neural network (BPANN) using the LM algorithm to determine the optimal photocatalytic degradation performance for methylene blue, anthraquinone, and indole. The experimental data for the degradation of methylene blue, anthraquinone, and indole were collected from previously published results of Merabet et al.,^[78] Khataee et al.^[79] and Sahoo et al.,^[80] respectively. The input parameters for indole degradation were UV light intensity (w/m²), stirring speed (rpm) and indole concentration (g/L). Reaction time (m), UV light intensity (w/m²), flow rate (mL/min) and initial dye concentration (mg/L) were the input parameters for anthraquinone, while catalyst dose (g/L), initial dye concentration (ppm) and pH solution for methylene blue. The analysis resulted in the optimized architectures 4-12-1, 3-16-1 and 3-14-1 for anthraquinone, methylene blue, and indole, respectively. A high correlation coefficient of 0.993, 0.961 and 0.999 were obtained for methylene blue, anthraquinone and indole, respectively, with confidence level above 95%. The analysis showed that the most influential parameters for methylene blue, anthraquinone and indole were reaction time, catalyst dosage and indole concentration, respectively.^[29] Zarei et al.^[81] designed an ANN model for the degradation of C.I. Basic Red 46 by photoelectro-Fenton (PEF) together with a photocatalytic process using commercial TiO₂ as a photocatalyst under UV light irradiation. The result shows that a maximum photocatalytic degradation efficiency ~ 98.8% was achieved by varying the experimental factors such as initial dye concentration, anode size, initial pH of the solution, applied current, Fe³⁺ concentration and type of UV light irradiation. A three-layer feed-forward backpropagation neural network model with topology 5:16:1 was developed to predict the degradation of BR46 dye solution. The results show that ANN provides a reasonable prediction coefficient of 0.986 for the experimental data. The relative importance of the experimental parameters was determined, and the result shows that the initial pH of the solution is the most influential parameter ~ 23% for the degradation of BR46.^[81] Ayodele et al.^[82] implemented an ANN model with an optimized configuration of 3-20-1, 3-5-1, 4-17-1, 3-10-1 and 4-6-1, using the LM algorithm for the photocatalytic degradation of chloramphenicol, azo dye, phenol, methylene blue, and gaseous styrene, respectively. The effect of experimental factors like organic pollutant concentration, pH of a solution, catalyst loading, oxidant concentration, light intensity, reaction time, and stirring speed on the degradation efficiency were investigated.^[82] The photocatalysts used for degradation were TiO₂, TiO₂/GAC, bismuth vanadate (BiVO₄), Cu-TiO₂ and multi-walled carbon nanotubes/titanium dioxide composite (MWCNTs – TiO₂). The high values of the results were in an agreement with the experimental data with correlation factors of 0.957, 0.998, 0.978, 0.998 and 0.997 for chloramphenicol, azo dye, phenol, gaseous styrene and methylene blue, respectively. The degradation of chloramphenicol, azo dye, phenol, methylene blue and gaseous styrene were mainly determined by the catalyst concentration, pH of the solution, phenol concentration, dye concentration and hydrothermal temperature, respectively, and the degradation rate for the photocatalytic experiment was 85.97%, 90%, 52%, 99.82%, 67.21% and 74.4%, respectively. The ANN model predicts the R² values, ANN structure and most influential parameters in the presence of coupled TiO₂ catalysts with various dyes, which are listed in Table 2. The Figure 8a provides the comparison of R² value obtained from various dye degradation processes in the presence of coupled TiO₂ catalysts.

Figure 7. Illustration in schematic form of the optimized ANN structure.^[25]

Figure 8. Comparison of the R² value obtained from various dye degradation in presence of catalysts.

Table 2. ANN modeling of photocatalytic degradation of different toxic dyes and phenols by TiO₂ based photocatalyst.

Pollutant degraded	Catalyst, Synthesis method	Light source	Degradation (%), Time (min)	Most influential parameters	ANN Structure	Validation, R^2	References
Maxilon blue 5Gdye	Fe/TiO2, Impregnation method	UV light	180	Calcination temperature	–	0.999	^[Citation24]
Methylene blue, Methyl orange	Sn/Zn-TiO2, Sol-gel	UV light	99.5%, 95.14%,	MB: reaction time MO:initial concentration	4-8-1	0.983, 0.988	^[Citation25]
Indole, anthraquinone and methyl blue	TiO2, Ag^+-doped TiO2, Commercially available, impregnation technology	UV light	~100%, 95.97%, 95%, 120, 60	Flow rate, dye concentration, and reaction time	3-14-1, 4-12-1 and 3-16-1	0.999, 0.96 and 0.993	^[Citation29]
Methylene blue	Co3O4/TiO2@Ze composite	UV-C	100%, 12 min	Physical method; hydrothermal method	5-6-1	0.943	^[Citation32]
Colour Index Acid Yellow 23	Ag-TiO2-	UV -C	~100%, 60	Dye concentration	4-5-1 4-7-1	1.00	^[Citation75]
Acid Fuchsin	TiO2/NiO, impregnation technique	UV light	~75%,40	H2PO4^−, Cu^2+ and IO4^−	–	0.949, 0.955, and 0.970	^[Citation76]
Phenol	TiO2 coupled with Mn^2+ cations, Commercially purchased	UV light	78%, 180	The effect of Mn^2+ concentration	5-12-1	0.949	^[Citation77]
C.I. Basic Red 46	TiO2 with Fe^3+ cation, Commercially purchased	UV illuminating	98.8%, 360	Initial pH	5:16:1	0.986	^[Citation81]
Chloram-phenicol, phenol, phenol, azo dye, gaseous styrene, and methylene blue	nano-TiO2, Cu-TiO2, TiO2/GAC, TiO2, MWCNTs–TiO2 composite, BiVO4, Commercially available, YOLDAS method, Hydrothermal method, commercially available, Hydrothermal treatment, polyol route.	UV, UV, UV, UV, UV, Simulated visible light	85.97%, 52%, 90%, 99.82%, 74.4%, 67.21%, 179.50, 90, 240	Photocatalyst concentration, phenol concentration, phenol con-centration, solution pH, hydrothermal temperature, and methylene blue concentration	3-20-1, 3-5-1, 3-2-1, 4-17-1, 4-6-1, 3-10-1	0.957, 0.978, 0.998, 0.998, 0.998, and 0.997	^[Citation82]
Methyl red	TiO2-graphene/chitosan, Staudenmaier method	UV light	100%, 20	Reaction time	4-8-1	0.983	^[Citation83]
Rhodamine B and Methylene Blue	TiO2 containing ZnAl layered double hydroxide, coprecipitation method	UV-A	-	Loading onto LDH	3-10-1	0.975	^[Citation84]

4.3. Pure ZnO

Amani-Ghadim et al.^[85] developed an ANN model with a three-layer multilayer perceptron backpropagation algorithm for photocatalytic degradation of C. I. Acid Blue 9 (AB9) by ZnO nanoparticles under UV light irradiation. The effect of several experimental parameters, such as light intensity, pollutant concentration, pH and ZnO content on the degradation efficiency has been calculated. AB9 dye degradation was influenced by the initial organic concentration, pH of the solution, light intensity and ZnO amount. The result shows that at the optimum condition, the degradation efficiency was ~ 90% within a reaction time of 90 min. A 5-9-1 topology ANN model with a correlation coefficient (R²) of 0.9883 shows a very good agreement with the experimental data.^[85] Khataee and Zarei^[86] developed (ANN) model for decolorizing C.I. Direct Yellow 12 (DY12) dye solution by ZnO nanoparticles under UV irradiation and photoelectron-Fenton process. The effect of experimental parameters such as initial dye concentration, applied current, initial pH of the solution, kind of ultraviolet (UV) light and initial Fe³⁺ concentration on photocatalytic degradation of DY12 was studied. At optimum conditions, the degradation efficiency was 96.7% within 6 h of reaction time. An ANN model was designed to predict the decolorization efficiency of the DY12 solution. The model structure with a topology of 5:14:1 and a correlation coefficient of 0.980 shows a very good agreement with the experimental data. The relative influence of experimental parameters were calculated and analysis indicated dominance of the initial pH factor on the degradation of Direct Yellow 12 dye solution.^[86] Also, the ANN model predicts the R² values, ANN structure and most influential parameters in the presence of pure ZnO catalysts with C.I. Acid Blue 9 and C.I. Direct Yellow 12 dyes, which are listed in Table 3. The Figure 8b provides the comparison of R² value obtained from various dye degradation processes in the presence of pure ZnO catalyst.

Table 3. ANN modeling for photocatalytic degradation of different toxic dyes by pure ZnO catalyst.

Pollutant degraded	Catalyst, Synthesis method	Light Source	Degradation efficiency (%), Time (min)	Most influential parameters	ANN Structure	Validation, R^2	References
C. I. Acid Blue 9.	ZnO Precipitation method	UV-C	~90%, 90	–	5-9-1	0.988	^[Citation85]
C.I. Direct Yellow 12	ZnO, Sol-gel dip coating method	UV C	96.7%, 360	Initial pH	5:14:1	0.980	^[Citation86]

4.4. Coupled ZnO photocatalyst

Abdollahi et al.^[87] developed the ANN model for photocatalytic degradation of m-cresol to determine the optimum efficiency. Precipitation method was utilized to synthesize the photocatalyst manganese/ZnO nanoparticles to degrade m-cresol under visible-light irradiation. The experimental parameters like light irradiation time, the concentration of m-cresol, pH of the solution and photocatalyst amount (input parameter) were varied to optimize the degradation efficiency (output parameter). To obtain the best results, ANN was trained using a variety of algorithms, including batch back propagation (BBP), quick propagation (QP), incremental back propagation (IBP), and LM. The smallest value of RMSE was an indicator for choosing each algorithm’s topologies. Based on the smallest RMSE value, the optimized topologies are IBP-4-15-1, QP-4-8-1, LM-4-10-1 and BBP-4-6-1. Among these structures, QP-4-8-1 showed absolute average deviation, the minimum RMSE and maximum R² value. Finally, QP-4-8-1 was used as the optimum model for the validation test. Further, the relative importance of the experimental parameters has also been analyzed and the result shows that pollutant concentration has the maximum influence on the photocatalytic degradation of m-cresol.^[87] Yusuffs et al.^[88] designed an ANN model structure 3-10-1 with a genetic algorithm to photo catalytically treat textile industry effluent by using PUM-ZnO photocatalyst under sunlight irradiation. The three parameters, catalyst dose, irradiation time and pH, were varied to obtain maximum degradation efficiency. At optimum conditions, the maximum degradation efficiency was observed ~ 91% within 45 min of irradiation time. The result shows a very good prediction of experimental data by the designed ANN model with a high coefficient of determination value (0.993).^[88] Dhiman et al.^[89] developed an ANN model for photocatalytic degradation of Acridine Orange dye by ZnO nanoparticles. ZnO@pAAm-g-GG nanoparticles were synthesized with polyacrylamide grafted guar gum polymer (pAAm-g-GG) along with prudent green chemistry method. An evaluation was performed to assess the effect of experimental variables such as AO dye concentration, contact time and catalyst dose on the photocatalytic degradation of Acridine Orange dye. The Response Surface Methodology involving Box-Behnken Design was used to evaluate the optimum process parameters to achieve maximum degradation efficiency. At the optimum condition of 10 mg/L dye concentration, 0.2 g/L catalyst dose and 210 min of irradiation, there was a maximum of 98% photocatalytic degradation. The developed ANN model with structure of 4-7-1 shows a good validation of experimental data with an R² value of 0.9997.^[89] Kıransan et al.^[90] developed an ANN model for Disperse Red 54 (DR54) decolorization by ZnO/Montmorilonite (ZnO/MMT) catalyst under UV-C, UV-A and UV-B radiation. The effect of various operational parameters such as UV light regions, initial dye concentration, nanocomposite dosage, and catalyst reusability was investigated. The result displayed that under UV-C radiation, the highest decolorization efficiency of ~ 90% was achieved within 10 min of light irradiation. An ANN model with three-layered feed forward back propagation with a structure of 3-10-1 was designed. The result shows agreement with experimental data with a correlation co-efficient of 0.999. Analysis shows that catalyst loading has the highest influence of 49.21% on the degradation of Disperse Red 54.^[90] Lenzi et al.^[24] implemented the ANN model with a backpropagation learning algorithm to photo catalytically degrade the maxilon blue 5 G dye by TiO₂, ZnO, TiO₂-ZnO-based catalyst under UV light irradiation. The photocatalysts Fe/TiO₂-ZnO, Fe/ZnO and Fe/TiO₂ were prepared by the impregnation method, where basic ZnO and TiO₂ were purchased commercially. The result shows that the maxilon blue 5 G dye was completely degraded within 180 min by all these catalysts. The analysis shows that the ANN model was in good agreement with experimental data with a correlation co-efficient of 0.9996.^[24] Shfeizadeh et al.^[91] performed photocatalytic degradation of methyl orange dye under UV irradiation by the chitin/TiO₂/ZnO, chitin/TiO₂ and chitin, composites synthesized by the sol-gel method and developed an ANN model to predict the experimental data. The effect of the experimental parameters such as initial dye concentration, catalyst loading, solution pH and initial concentration of hydrogen peroxide was investigated. At optimum experimental conditions, the maximum degradation efficiency was observed ~ 61% for 4 hours. An ANN model with a structure of 4-4-1 shows a very good agreement with experimental values with a correlation co-efficient of 0.975. The high correlation coefficient for chitin/TiO₂/ZnO, chitin/TiO₂ and chitin composites are 0.996, 0.995 and 0.991, respectively. The result shows that the catalyst loading has the maximum influence of 38.86% on the photocatalytic degradation of methyl orange.^[91] Ozbay et al.^[92] have investigated the photodegradation activity of Sunfix red S3B reactive dye using polyaniline-modified TiO₂ and ZnO under visible range. The polyaniline-modified TiO₂ and ZnO nanocomposites were synthesized using an in situ polymerization method. The photodegradation efficiency was investigated by varying the irradiation time, initial dye concentration, photocatalyst amount, and irradiation period. The ANN model with back propagation feed-forward network has been developed to determine the photodegradation efficiency. The results show good prediction for both nanocomposite types. The high correlation factor of 0.9 and 0.87 for PANI/TiO₂ and PANI/ZnO, respectively, and the ANN structure of 3-9-1 shows good agreement with the experimental data.^[92] The ANN model was able to predict the R² values, ANN structure and most influential parameters in the presence of coupled ZnO catalysts with various dyes, which are listed in Table 4.

Table 4. ANN modeling for photocatalytic degradation of different toxic dyes by coupled ZnO catalyst.

Pollutant Degraded	Catalyst, Synthesis method	Light source	Degradation efficiency (%), Time (min)	Most influential parameters	ANN Structure	Validation, R^2	References
Maxilon blue 5G dye	Fe/TiO2–ZnO Fe/TiO2, and Fe/ZnO impregnation method	UV light	100%, 180	reaction time	–	0.999	^[Citation24]
Methylene blue and methyl orange	Sn/Zn-TiO2, sol–gel method	UV light	93.7%, 88.1%, 9	reaction time	4-8-1	0.983 and 0.988	^[Citation25]
Reactive black 5	Cu-doped ZnO, mild hydrothermal technique	UV light	–	–	4-6-1	0.925	^[Citation30]
m-cresol	Manganese doped ZnO, Precipitation method	visible-light irradiation.	100%, 360	m-cresol’s concentration	4-8-1	0.994	^[Citation87]
Textile Effluent C	Pumice-supported ZnO, impregnation method	visible light	99.46%; 45.85	Catalyst dosage, irradiation time, pH	3-10-1	0.993	^[Citation88]
Acridine Orange dye	ZnO, green chemistry approach	UV light	98%, 210	–	4-7-1	0.999	^[Citation89]
Disperse Red 54	ZnO/Montmorilonite Chemical method	UV illuminating	~90%, 10	Catalyst dose	3-10-1	0.999	^[Citation90]
Methyl orange dye C	Chitin/TiO2/ZnO, sol–gel method	UV irradiation	61%, 240	Catalyst loading	4-4-1	0.975	^[Citation91]
Sunfix red S3B dye	Polyaniline-modified TiO2 and ZnO, Polymerization method	Visible light			3-9-1	0.90, 0.87	^[Citation92]
Acid orange 7	Ni- doped ZnO, mild hydrothermal condition.	UV light	93%, 90		4-7-1	0.991	^[Citation93]

The Figure 9b provides the comparison of R² value obtained from various dye degradation processes in the presence of coupled ZnO catalysts.

Figure 9. Comparison of the R² value obtained from various dye degradation; (a) presence of coupled TiO₂ catalysts; (b) presence of coupled ZnO catalysts.

Pirsaheb et al.^[93] photocatalytically degraded acid orange 7 by using NiO-ZnO photocatalyst and the process was modeled by ANN. Analysis shows that experimental values are supporting the ANN model tropology 4-7-1 with a R² values of 0.991 as shown in Figure 10.^[93]

Figure 10. Degradation efficiency of acid orange 7 as a function of (a) catalyst dose at pH 5 and dye concentration 60 mg/L; (b) dye concentration at pH 7 and a catalyst dose of 0.8 g/L.^[93].

4.5. Other photocatalysts

Kakhki et al.^[35] developed an ANN model for photo-catalytically degrading methylene blue by a sulfur – nitrogen doped Fe₂O₃ catalyst. The catalyst was prepared by a simple chemical method. The photocatalytic efficiency was optimized by varying the nanoparticles dose, the dye concentration, pH and the light dose. The result shows that the degradation was ~ 95% within a reaction time of 5 min. A three-layer ANN model with a genetic algorithm was generated. The ANN structure consisted of 3-18-1 with a correlation coefficient of 0.921 which shows a very good prediction of experimental data. The relative influence of the experimental parameters on MB degradation was calculated and the result shows that the dye concentration has the maximum influence of ~ 46%.^[35] Wang et al.^[9] developed an ANN model to simulate the photocatalytic performance of BaAl₂O₄: Ce and Mn-Ce-co-doped BaAl₂O₄ nanocomposite for MB degradation under simulated sunlight irradiation. The catalysts were synthesized by γ-ray irradiation-assisted polyacrylamide gel method. The experimental parameters such as dye concentration, catalyst content, irradiation time and pH value were systematically varied to optimize the photocatalytic efficiency. At the optimum condition, 1 g/L catalyst content, 5 mg/L dye concentration, 13 pH value and 4 h irradiation time, maximum photocatalytic degradation efficiency of ~ 92% was obtained. An ANN model was designed to validate the experimental data, and the result shows that the experimental and predicted values are highly consistent with a correlation coefficient of 0.994.^[94] Yang et al.^[95] constructed an ANN model for photo catalytically degraded Congo red dye by ternary metal selenide nanocomposites supported by chitosan microspheres (FeNiSe-CHM). The photocatalyst was synthesized by the solvothermal method. At the optimum condition of pH 6.0, 60 ppm dye concentration and 0.2 g catalyst dose, 99% photocatalytic degradation was observed for Congo red dye under sunlight irradiation for 140 min. An ANN model with a topology of 4-2-1 was employed to predict the experimental data, which shows a very good validation with a correlation coefficient of 0.99. The operational parameters for photocatalytic degradation are catalyst dose, dye concentration, pH, and reaction time. Among all the parameters catalyst dose is the most influential in dye degradation.^[95] The ANN model is capable to predict the R² values, ANN structure and most influential parameters in the presence of various photocatalysts catalysts with various dyes, which are listed in Table 5.

Table 5. ANN modeling for photocatalytic degradation of different toxic dyes by various other photocatalyst.

Pollutant Degraded	Catalyst	synthesis method	Light source	Degradation efficiency (%), Time (min)	Most influential parameters	ANN Structure	Validation, R2	References
Methylene blue	sulfur–nitrogen co-doped Fe2O3	a simple and efficient method	Simulated visible light	~95%, 5	concentration of the dye	3-18-1	0.921	^[Citation32]
Methylene blue	BiVO4/BiPO4/rGO hetrojunctions	Hydrothermal	UV–visible light	94.4	pH	–	09944	^[Citation33]
Methylene blue	BaAl2O4:Ce:Mn composite	a gamma-ray irradiation assisted polyacrylamide gel method	Xenon lamp	92%, 360	–	–	0.994	^[Citation94]
Congo red dye	(FeNiSe-CHM) NPs	solvothermal method	sunlight light	90%, 140	Catalyst dose	4-2-1	0.99	^[Citation95]
Tartrazine and patent blue	MgO-La2O3 mixed metal oxides	layered double hydroxide	UV–visible light	97.4%, 150	irradiation time		0.98	^[Citation96]
methylene blue	Montmorillonite		UV–visible light	91.3%, 90	pH	–	0.9987	^[Citation97]

5. Conclusion and future perspectives

Effective photocatalytic degradation of textile wastewater is significantly influenced by several process parameters such as pollutant concentration, the oxidizing agent, photocatalyst loading, pH of the solution, reaction temperature, light intensity and light irradiation time. The optimum condition of these parameters is difficult to achieve using experimental design due to time and cost consumption in the experimental process, and the complex non-linear relationship between these parameters. ANN model is widely used in the research community for understanding this non-linear relationship among process parameters along with the determination of each parameter’s influence on the degradation process to design an optimal reaction process for pollutant degradation in less time and cost. In this review paper, we have studied the literature from the past 14 years to reveal the potential of ANN to model the photocatalytic degradation processes in order to predict the degradation of pollutants in the given reaction conditions. In general, findings of this work show the critical role of the ANN model in the optimization of process parameters for efficient degradation of pollutants. From this review, it could be concluded that the MLP network is the most widely used for the pollutant degradation among different categories of ANNs. For training, backpropagation algorithm is widely used in the existing literature. However, the recent studies have also reported the use of faster algorithms such as LM, and scaled conjugate Gradient. In addition, the recent research also suggests the use of evolutionary algorithms such as genetic algorithms (GA) and sparrow search algorithms (SSA) for optimization of neural network architecture and training time.

Abbreviation

Abbreviations	=	Explanation
ANN	=	Artificial Neural Network
$\mathrm{\Delta }\mathit{w}$	=	Weight adjustment
$\mathit{\eta }$	=	Difference between the predicted and desired output
$\mathit{d}$	=	Learning rate
$\mathit{x}$	=	Input data
RBF	=	Radial basis function
MLP	=	Multi-layer perceptron
ReLU	=	Rectified linear unit
SLP	=	Single Layer Perceptron
f	=	Activation function
$W_{\mathit{in}}$	=	Weights connecting
jth	=	Neuron in the hidden layer
$Y_{\mathit{n}-1}$	=	Output vector of the previous layer
$b_{\mathit{in}}$	=	Ias term for nth layer
Wj	=	Weights of the output neurons
Cj	=	Prototype(center) of the jth neuron
x	=	Input vector
σ	=	Parameter that determines the width of the gaussian function
RNN	=	Recurrent Neural Networks
LSTM	=	Long short-term memory
R2	=	Coefficient of determination
n	=	Number of data points
RMSE	=	Root-mean-square deviation
LM	=	Levenberg-Marquardt
MG	=	Malachite Green
AR114	=	Acid Red 114
4-NP	=	4-nitrophenol
AO7	=	Acid Orange 7
BR46	=	C.I. Basic Red 46
C.I. AY23	=	Colour Index Acid Yellow 23
AF	=	Acid fuchsin
MO	=	Methyl orange
BPANN	=	Back Propagation Artificial Neural Network
PEF	=	Photoelectro-Fenton
LDH	=	Layered double hydroxides
AB9	=	C. I. Acid Blue 9
DY12	=	C.I. Direct Yellow 12
UV	=	Ultraviolet
QP	=	Quick propagation
BBP	=	Batch back propagation
IBP	=	Incremental back propagation
pAAm-g-GG	=	Polyacrylamide grafted guar gum polymer
DR54	=	Disperse Red 54
MMT	=	Montmorilonite
PANI	=	Polyaniline

Acknowledgments

The authors are grateful for the financial support provided by the Slovenian Research Agency through the core research funding (P1-0134) and through the projects focused on photocatalysts (J2-4444, J2-4441, J2-4401, N2-0310, and J7-4638). Additionally, Project No. SRG/2019/001732 received funding through the Science and Engineering Research Board, New Delhi, India.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Correction Statement

This article has been republished with minor changes. These changes do not impact the academic content of the article.