Artificial Intelligence-Guided Supervised Learning Models for Photocatalysis in Wastewater Treatment

Abstract

Artificial intelligence (AI), when integrated with photocatalysis, has demonstrated high predictive accuracy in optimizing photocatalytic processes for wastewater treatment using a variety of catalysts such as TiO₂, ZnO, CdS, Zr, WO₂, and CeO₂. The progress of research in this area is greatly enhanced by advancements in data science and AI, which enable rapid analysis of large datasets in materials chemistry. This article presents a comprehensive review and critical assessment of AI-based supervised learning models, including support vector machines (SVMs), artificial neural networks (ANNs), and tree-based algorithms. Their predictive capabilities have been evaluated using statistical metrics such as the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE), with numerous investigations documenting R² values greater than 0.95 and RMSE values as low as 0.02 in forecasting pollutant degradation. To enhance model interpretability, Shapley Additive Explanations (SHAP) have been employed to prioritize the relative significance of input variables, illustrating, for example, that pH and light intensity frequently exert the most substantial influence on photocatalytic performance. These AI frameworks not only attain dependable predictions of degradation efficiency for dyes, pharmaceuticals, and heavy metals, but also contribute to economically viable optimization strategies and the identification of novel photocatalysts. Overall, this review provides evidence-based guidance for researchers and practitioners seeking to advance wastewater treatment technologies by integrating supervised machine learning with photocatalysis.

1. Introduction

Artificial intelligence (AI) is transforming the modern world through its unmatched capabilities in analytical forecasting, optimization, and data-driven decision-making [1]. Its capacity to tackle complex worldwide issues is especially noticeable in the area of ecological sustainability [2,3]. Among these challenges, wastewater treatment and management stand out as vital objectives. Innovative and practical solutions are urgently needed to safeguard water resources from the increasing demands of industrialization, population growth, and climate change. AI has emerged as a powerful tool in both technology and environmental science [2,4,5]. The use of AI to optimize and enhance the performance of several wastewater treatment procedures, such as photocatalysis, has grown in recent years [6]. Machine learning algorithms and deep learning models are examples of AI-based models that have been utilized to predict and improve the photocatalytic degradation of pollutants, as well as to determine the best operating parameters [7]. Several AI models, such as support vector machines, decision trees, random forests, and neural networks, have been applied in the field of photocatalysis [8]. These models are trained on extensive experimental datasets and are capable of accurately predicting the performance of photocatalytic systems under varying operational conditions. The integration of AI with photocatalysis holds significant promise for revolutionizing wastewater treatment by enabling the design of more effective, efficient, and environmentally sustainable treatment strategies [9]. AI can lower energy usage, enhance treatment results, and optimize the design and operation of photocatalytic systems. AI may also be used to forecast the performance of novel photocatalytic materials and processes under various operating situations. Overall, the application of AI in photocatalysis is expected to revolutionize wastewater treatment and facilitates the development of more efficient and sustainable methods [10,11].

Water is one of the most vital resources for sustaining life on Earth. However, various human activities have significantly contributed to environmental degradation. The discharge of untreated or inadequately treated wastewater introduces a wide range of contaminants into natural water bodies, posing serious threats to both ecological balance and human health worldwide [12,13]. In response, integrated wastewater management has emerged as a sustainable solution, especially in ecologically sensitive regions. For example, studies conducted in the Andean páramo communities highlight the effectiveness of decentralized, nature-based treatment systems in preserving environmental integrity and promoting sustainable development [14]. To minimize environmental pollution and safeguard public health [15], it is important to learn the origins, composition, and efficient treatment techniques of wastewater. Increasing levels of urbanization, industrialization [16], and human activity have led to wastewater becoming a major global issue [17]. Pathogens in wastewater can cause waterborne illnesses like gastroenteritis, typhoid, and cholera, especially in communities lacking access to safe drinking water and sanitation [18]. Many other types of contaminants can be found in wastewater, such as microplastics, heavy metals, pathogens (viruses, bacteria), nutrients (phosphorus, nitrogen), and organic substances (pesticides, medications). Ecosystems and human health are at particular risk from each of these contaminants [19].

Wastewater treatment involves several sequential processes designed to remove contaminants and pollutants from wastewater before it is safely discharged back into the environment. These processes typically include physical, chemical, biological, and electrochemical treatment [20,21,22]. Primary sedimentation tanks operate by allowing gravity to separate solid materials from wastewater, resulting in the formation of sludge at the bottom and scum at the surface. Following this, biological treatment processes, such as activated sludge systems, trickling filters, and sequencing batch reactors, employ microorganisms to decompose organic matter into carbon dioxide, water, and biomass. Electrochemical techniques such as electrooxidation and electrocoagulation can provide environmentally suitable chemical-free wastewater treatment. They are efficient against persistent pollutants, dyes, and heavy metals while producing minimal sludge [20]. Disinfection methods such as chlorination, ultraviolet irradiation, ozone treatment, or a combination of these are then applied to ensure microbiological safety before discharge [23]. By increasing energy efficiency and pollutant removal, recent advancements in membrane technologies have enhanced wastewater treatment. Cutting-edge materials like nanotechnology and graphene oxide improve durability and lessen fouling. Superior performance is demonstrated by hybrid membrane systems that combine biological or photocatalytic processes. For real-world applications, scalability, cost, and long-term stability continue to be major obstacles [24].

The recently developed method of photocatalytic degradation presents a promising technological solution for wastewater treatment. Semiconductors are widely employed as photocatalysts due to their exceptional material properties. These include a high refractive index, strong UV absorption, excellent incident photoelectric conversion efficiency, and a favorable dielectric constant. Additionally, their outstanding photocatalytic activity, photostability, chemical stability, long-term corrosion resistance, and non-toxic nature make them ideal candidates for converting solar energy into chemical energy [25,26]. Photocatalysis, an advanced oxidation process, has garnered significant attention in wastewater treatment due to its ability to effectively degrade a wide range of organic contaminants [27]. To produce reactive species that can degrade pollutants, this technique uses photocatalysts, which are usually metal oxides like CdS, ZrO, TiO₂, and ZnO that are activated by light [28]. Despite their potential, conventional photocatalysts are constrained by issues such as electron–hole recombination and large band gaps. Recent research has focused on improving photocatalytic performance through various strategies, including doping and the development of new materials [29,30].

This review article presents a comprehensive overview of recent advances and emerging trends in photocatalysis research. It highlights the integration of AI models, particularly supervised learning techniques, for predicting photocatalytic performance and assessing the influence of various operational parameters. The review also synthesizes the current literature on AI-driven methodologies for the degradation of diverse environmental contaminants. In addition, it critically evaluates key challenges hindering the commercial scalability of photocatalytic systems, including catalyst stability, economic viability, and regulatory barriers. Finally, the article offers forward-looking perspectives aimed at guiding the future development and practical implementation of AI-assisted photocatalytic technologies.

2. Photocatalysis Mechanisms and Dynamics

Photocatalysis refers to a chemical reaction initiated by the absorption of light by one or more reactive species. Catalysts such as CeO₂, ZnO, CuO, Mn₃O₄, TiO₂, Fe₂O₃, SnO₂, WO₃, SiO₂/TiO₂, ZnO/TiO₂, and MgO/ZnO have been employed under UV or visible light to facilitate the photocatalytic degradation of pollutants [31,32]. Semiconductor materials are commonly used as sensitizers in light-induced redox reactions due to their electronic structure, characterized by a filled valence band and an empty conduction band [33,34]. Figure 1 illustrates the photocatalytic mechanism and the process through which the catalyst removes contaminants from wastewater. When the energy of incident photons equals or exceeds the band gap of the photocatalyst, the material absorbs light, exciting electrons from the valence band to the conduction band and generating electron–hole pairs. The separation of these charge carriers is essential to prevent recombination. This separation can occur through interaction with adsorbed reactants or via surface processes on the catalyst. The photogenerated holes in the valence band oxidize adsorbed water molecules, producing highly reactive hydroxyl radicals, while the excited electrons in the conduction band reduce dissolved oxygen into superoxide radicals [35,36,37,38,39,40,41]. After that, these reactive oxygen species (ROS) take part in a sequence of oxidation and reduction reactions that break down a variety of pollutants, such as organic compounds, heavy metals, dyes, biological contaminants, and pharmaceutical residues. Figure 2 illustrates the process by which contaminants are adsorbed onto the surface and then broken down by superoxide and hydroxyl radicals [39,42,43]. Despite being reusable over several cycles and not being consumed in the reaction, photocatalysts may lose some of their effectiveness over time due to sintering, agglomeration, or surface fouling.

3. Operational Parameters Effects on Artificial Intelligence-Driven Photocatalysis

The main factors affecting photocatalytic efficiency and their function in AI-based prediction models are depicted in Figure 3. The effectiveness of the photocatalytic process is significantly influenced by several variables, including the characteristics of the photocatalyst, the types of contaminants present, chemical additives, reactant concentration, catalyst dosage, temperature, light intensity and wavelength, and pH [44,45,46,47]. To forecast and maximize photocatalytic performance, these variables are crucial input parameters for evaluating and training the AI model. These parameters are commonly used as input features in AI-based models to predict photocatalytic degradation efficiency. Once processed by supervised learning algorithms, the models output predicted degradation efficiencies and identify optimal conditions for enhanced performance. The integration of such variables into AI frameworks allows for accurate modeling and optimization of photocatalytic processes in wastewater treatment applications [41,48,49]. To estimate the effectiveness of photocatalytic dye degradation, four machine learning models, Ensemble Learning Tree (ELT), GPR, SVM, and DT, were used in conjunction with GA and particle PSO. With an RMSE of 2.6410 × 10⁻⁴ and an R² value of 0.992, the ELT-PSO hybrid model outperformed DT, GPR, and SVM by far, being the most accurate of these. The most important elements influencing degradation were dye type, bandgap, initial dye concentration, solution volume, and reaction time, according to an investigation of parameter influence using partial dependence graphs, Shapley analysis, and sensitivity testing [50]. To predict the photocatalytic degradation of Rhodamine B using ZnO doped with 19 elements, eleven machine learning models were evaluated; CatBoost achieved the highest accuracy (R² = 0.96, RMSE = 0.0579). The most important factors were determined to be reaction time, dopant type, calcination temperature, and light wavelength since they directly affect charge transfer, crystallinity, and photon absorption. Reaction circumstances accounted for the greatest portion (70.7%), followed by catalyst characteristics (22.1%) and preparation conditions (7.2%) [51].

3.1. pH

The pH significantly influences photocatalytic processes, affecting both the efficiency of photocatalysts and the degradation rates of organic compounds, and is therefore commonly included as an input in AI models [52]. Degradation rates can be affected by pH variations because they can change the way organic molecules adsorb onto the photocatalyst surfaces [53]. Marfur and his colleagues demonstrated that pH plays a crucial role in the degradation efficiency of 2,4-dichlorophenol using Cu-ZnO and Ag-ZnO photocatalysts, with both showing optimal performance in acidic conditions, particularly at pH 3 [54]. The photocatalytic activity of cerium oxide nanoparticles (CeO₂) is also strongly influenced by the pH of the synthesis medium. Studies have shown that CeO₂ synthesized at pH 8 exhibits the highest photocatalytic activity [55]. Similarly, the efficiency of dye degradation using iron oxide nanoparticles is strongly pH-dependent. For example, rhodamine B (RhB) degrades most effectively at neutral pH (6.5), while brilliant green (BG) shows maximum degradation at pH 9. In contrast, methyl red (MR) and indigo carmine (IC) degrade most efficiently under acidic conditions, specifically at pH 3. A lower pH environment reduces the interaction between the photocatalyst and positively charged dyes like RhB due to increased H⁺ concentration. Conversely, higher pH levels can negatively affect the degradation of dyes like IC due to unfavorable charge interactions [56].

3.2. Light Intensity and Wavelength

The rate of photocatalysis in certain materials is strongly affected by variations in light intensity. In AI-based predictive models, light intensity and wavelength are often used as key input parameters to determine optimal conditions for maximum photocatalytic performance. Research has shown that the photocatalytic rate constant generally increases with higher energy flux density. For instance, under intensified UV irradiation, the rate constant may rise from 0.038 to 0.055 h⁻¹, reflecting improved performance up to a certain point [57]. However, this relationship is not always linear, as some materials exhibit saturation thresholds beyond which additional increases in light intensity yield no further improvement. In specific cases, photocatalytic response is highly dependent on both the wavelength and intensity of light. For example, in the photocatalytic reduction of formic acid using graphene oxide, a high-intensity red wavelength triggered the fastest initial reaction rate, while maximum conversion was achieved under a low-intensity white light source [58]. Similarly, the photocatalytic activity of TiO₂ nanostructures was shown to be strongly influenced by light intensity, with additional interactions noted between light intensity and the stirring rate, indicating complex dependencies [59]. According to Khataee and coworkers’ study, UV light plays a critical role in photocatalytic dye degradation. Their study on C.I. Basic Red 46 (BR46) revealed that the decolorization rate increased with higher UV light intensity. This enhancement was attributed to greater photon absorption by TiO₂, which facilitated the generation of more electron–hole pairs and hydroxyl radicals, thereby accelerating the dye removal process [60].

3.3. Temperature

The rate of photocatalysis is significantly influenced by temperature. The photocatalytic activity of TiO₂-based catalysts with various cocatalysts was strongly influenced by the reaction temperature. While Pd/TiO₂ was more effective at higher temperatures up to 50 °C, after which its efficiency fell, Cu/TiO₂ demonstrated higher activity at lower temperatures, such as room temperature [61]. The impact of heat on photocatalysis showed that, when the room temperature rises to 75 or 90 °C, semiconductor photocatalysts such as TiO₂ experience a modification in their band structure. This resulted in an alteration in the action spectrum towards lower energy levels, which allowed for the photocatalytic reaction to utilize a greater number of incident photons. These datasets were used in the modeling of an AI model design for prediction [62].

3.4. Number of Catalysts

The rate at which photocatalysis occurs is largely dependent on the catalyst quantity. According to a study, increasing the catalyst concentration could speed up the reaction. In this condition, a low concentration of catalyst (0.4 g/L) increased the reaction rate up to 50% by making the wall reflectivity reach 98% [63]. Furthermore, the addition of tungsten to TiO₂ catalysts demonstrated increased photocatalytic activity with each cycle, increasing the rate by 1.7 times after the first cycle and 3.1 times after the second [64].

3.5. Concentration of Reactants

Reactant concentration is a key factor that affects the speed of photocatalysis. When employing a Pt/Ni(OH)₂/CdZnS photocatalyst for photocatalytic hydrogen generation, the reaction rate was dependent on the amounts of sodium hydroxide and ethanol in different data point sets, as varied by changing the concentration, and an AI model was used to determine the optimal performance [65]. Furthermore, in the context of a photoreactor using titanium dioxide as a photocatalyst, it was found that varying the catalyst concentration and wall reflectivity can significantly increase the reaction rate, with a 50% increase observed at low catalyst concentrations and high wall reflectivity levels [63]. Moreover, the use of an ion-concentration-polarization-assisted photocatalytic reactor demonstrated that altering the electric field through concentration polarization could lead to an 85.5% increase in the reaction rate, showcasing the impact of reactant concentration on photocatalytic efficiency [66].

3.6. Chemical Additives

The incorporation of chemical additives (doping) significantly influences the photocatalytic performance of materials like TiO₂. Various studies have highlighted the impact of additives on photocatalytic reactions. For instance, the presence of BaSO₄ in TiO₂-based materials enhanced photocatalytic activity by scattering UV light, leading to high photonic efficiencies [67]. The incorporation of both organic and inorganic additives could highlight the complex and variable impacts of additives on the overall photocatalytic process by either promoting or restricting the photocatalytic efficiency of particular pollutants [68].

3.7. Contaminants

The effectiveness of photocatalytic degradation was impacted by the presence of contaminants such as acetylsalicylic acid, ethinyl estradiol, ibuprofen, and paracetamol. It has been demonstrated that non-calcined Nb₂O₅ catalysts have better activity in reducing these new pollutants [69]. The interaction of pollutants with photocatalysts such as TiO₂ in visible light initiated the photocatalytic destruction of bacteria, showing how visible light absorption by pollutants plays a critical role in enhancing photocatalysis [70].

3.8. Properties of Photocatalysts

Materials used in photocatalysis are essential to the effectiveness of photocatalysis procedures. AI has been used in a selection of suitable materials for maximum efficiency in photocatalysis. Studies have indicated that the structure and composition have a big influence on how well they work. A photocatalysis material consisting of titanium dioxide, cerium dioxide, and an outer layer of silver exhibited antibacterial properties [71]. The photocatalytic activity of a photocatalyst was largely dependent on its surface area. Lukic et al.’s study discovered that the β-Ga₂O₃ nanoparticles’ specific surface area, which was obtained using chemical vapor synthesis, affected their photocatalytic efficacy in both overall water splitting (OWS) and aqueous methanol reforming [72]. Transition metal oxides like ZnO, TiO₂, WOx, Nb₂O₅, and CeO₂ are commonly used as photocatalysts, with composites like ZnO–ceria and titania–tungsten oxide showing enhanced photoactivity due to smaller crystallite size and increased surface area [73]. A photocatalyst’s structure and morphology were important factors in determining how active it is. Variations in the surfaces and morphologies of crystals may have a significant effect on the effectiveness of photocatalysis [74,75].

4. Shaping the Future of Photocatalytic Materials

Doping, the process of introducing impurities to enhance a material’s properties, is a key to optimizing photocatalyst performance. AI has revolutionized this field by enabling more efficient and innovative approaches to photocatalyst design and optimization [76,77]. Several doping techniques were investigated to improve the photocatalytic abilities of various substances. Doping is utilized in photocatalysts to improve band gap engineering [78], surface properties [79], charge separation [80], and light absorption [81]. The types of doping used in photocatalysis include the following: co-doping, which involves combining non-metal and metal dopants [82]; interstitial doping [83]; surface doping [46]; non-metal doping with carbon, nitrogen, or sulfur doping [84]; and metal doping with transition metal and rare earth metal doping [85]. Studies have revealed that incorporating transition metals and nonmetals into TiO₂ can greatly improve its band gap, including carrying capacity and light absorption [86]. ML was employed to predict the energy band gap of doped ZnO at various doping concentrations. Among the models tested, GPR achieved the highest accuracy (CC = 98.97%, RMSE = 0.0022, MAE = 0.0020), outperforming SVM and RF. These results demonstrate that GPR is highly effective for capturing the influence of doping on ZnO properties, emphasizing the role of model selection in designing functional photocatalytic materials for energy applications [87]. This study developed copper-doped TiO₂ (Cu–TiO₂) nanoparticles for cetirizine hydrochloride (CTZ-H) degradation. The best performance, 93% removal, was obtained with 0.5 wt% Cu–TiO₂ at pH 4.9, 100 mg/L catalyst, and 10 mg/L pollutant concentration. Among predictive models, the SVM optimized with the Improved Grey Wolf Optimizer (IGWO) showed the highest accuracy, R² = 0.9999. Degradation efficiency was mainly influenced by copper loading, solution pH, catalyst dosage, and initial pollutant concentration [88]. Research shows that by combining green-synthesized ZnO with Ni-doped ZnO photocatalysts, machine learning can be used to predict the hazardous pollutant (PRM) degradation efficiency. While linear models such as Ridge regression demonstrated less flexibility, Gradient Boosting and XGBoost optimized with GridSearchCV achieved the best performance among the models under investigation R² > 0.99, low MSE/MAE) [89]. The integration of AI, particularly ANNs, allows for efficient modeling and prediction of photocatalytic performance, significantly improving the treatment of various pollutants. Figure 4 compares the degradation efficiency of methylene blue using different AI models combined with various dopants in photocatalytic wastewater treatment. The data highlights the influence of AI-driven optimization and dopant selection on the degradation process.

Table 1. Different dopants are used in methylene blue degradation by artificial intelligence photocatalysis to predict its efficiency.

Sr#	Dopant	AI Model	Degradation Efficiency	Time	Reference
1	Sulfur–nitrogen codoped Fe2O3 nanostructure	ANN	95%	5 min	[90]
2	ZnO/MgO	ANN	99%	174 min	[91]
3	Nanoscale zero-valent iron (nZVI)	ANN	100%	30 min	[92]
4	Ho-CaWO4 nanoparticles	ANN	71.17%	15.16 min	[93]
5	Graphene oxide/chitosan (GO/CS)	ANN	90.34%	125 min	[94]

shows the dopants used in methylene blue degradation for AI-based photocatalytic efficiency prediction.

5. Research Methodology

This review adopts a structured, innovation-driven approach to explore the growing integration of AI in optimizing photocatalytic wastewater treatment processes. As photocatalytic systems become increasingly complex and multi-variable, AI offers powerful capabilities in modeling, prediction, and data-driven decision-making to enhance pollutant degradation efficiency. To ensure comprehensive and reproducible findings, a four-phase systematic review protocol, comprising planning, searching, screening, and reporting, was implemented by PRISMA guidelines [95,96]. During the planning phase, the scope was refined to focus exclusively on AI-powered optimization strategies, such as support vector machines, deep learning, artificial neural networks, and hybrid ensemble models, applied in both simulated and experimental photocatalytic degradation studies. Particular emphasis was placed on identifying research utilizing AI models to address key aspects such as reaction kinetics, catalyst selection, light intensity optimization, and parameter sensitivity analysis.

A comprehensive literature search was conducted to identify relevant studies on the application of artificial intelligence (AI) in photocatalytic wastewater treatment. Three large databases were searched: Google Scholar (n = 1590), PubMed (n = 13), and ScienceDirect (n = 599), using the combined search terms “artificial intelligence” photocatalysis “wastewater treatment.” The search on Google Scholar returned far more records compared to the other databases. This aspect can be explained by the fact that it has broad indexing capability, including journals, conference proceedings, theses, and non-traditional scholarly output. In contrast, PubMed and ScienceDirect index only peer-reviewed biomedical and scientific literature. Following the consolidation of all the records (n = 2202), the cases of duplication were identified and then removed (n = 600), leaving 1602 unique articles. The remaining studies were screened using titles and abstracts, and those not relevant to the AI-based photocatalysis of wastewater treatment were discarded. Additionally, studies written in a language other than English, non-peer-reviewed articles, book chapters, editorial articles, and articles with inaccessible full texts were discarded. The inclusion criteria were specifically aimed at studies employing the use of artificial intelligence or machine learning models to optimize, predict, or improve photocatalytic processes in wastewater treatment. Experimental studies and computational modeling articles were both eligible, as long as they reported quantifiable results like degradation efficiency, optimization parameters, or mechanistic understanding. Following a stringent selection process, 230 articles were found to be eligible and were considered for the final appraisal. The stepwise screening and selection process is evident from the PRISMA diagram in Figure 5, which provides a stepwise summary of identification phases, screening, eligibility assessment, and inclusion.

5.1. Artificial Intelligence Models in Photocatalysis

AI is essential for improving photocatalysis reaction efficiency because it helps with catalyst design, performance prediction, and reaction optimization. Research has shown that AI algorithms have been used to design photocatalysts, forecast their performance, and optimize reaction conditions, all of which can save development time and costs while increasing efficiency [46,91,97]. Photocatalysis aided by AI has great potential for some industrial uses, especially in wastewater treatment and environmental purification procedures. Photocatalysts were designed and their efficacy was predicted with high accuracy through the use of intelligent algorithms, which reduced development time and costs [46]. The three categories of AI methods based on learning approaches (supervised, unsupervised, and reinforcement learning), application types (predictive modeling, optimization, and classification), and specific purposes (real-time monitoring and sensor fusion) have been developed in Figure 6. Each branch highlights commonly used techniques, showcasing the versatility and scope of AI in enhancing wastewater treatment processes. Optimization of photocatalytic processes using AI models can enhance the degradation of pollutants like MB and 2,6-dichlorophenol, resulting in cost-effective and sustainable treatment solutions [26,77].

In supervised learning (SL), a branch of algorithmic learning, models are trained on labeled datasets in order to categorize or predict data. This method is used in the field of photocatalysis to find active sites, forecast the efficiency of photocatalysts, and optimize material characteristics for particular reactions [76]. Unsupervised learning (UNSL) techniques use no previous classifications to find patterns and connections in data. Patterns in complicated datasets associated with pollutant degradation can be identified with this technique. Unsupervised learning enhances guided machine learning by offering perspectives that labeled data might not be able to provide. Reinforcement learning (RL), which provides creative ways to maximize treatment plans and boost productivity, is being included in water treatment procedures more and more. These technologies are used in many different areas of water treatment, such as improved control mechanisms for wastewater treatment and decision support systems (DSS) for drinking water treatment facilities. Furthermore, the possibility of using semiconductor photocatalysis to remediate water contaminants has been studied [98,99]. The advancement of photocatalysis has been significantly supported by AI, particularly in the design and performance prediction of photocatalytic materials. Various AI models have been applied in the context of environmental pollution treatment, aiming to enhance the efficiency and accuracy of photocatalytic processes. Among these, artificial neural networks (ANNs) and attention-based deep learning models, such as PhotoCat, have demonstrated notable success in accurately predicting photocatalytic reactions. These approaches, along with other AI techniques, offer powerful tools for modeling complex reaction pathways, optimizing operational parameters, and accelerating the development of next-generation photocatalysts [46,83]. Reinforcement learning, unsupervised learning, and supervised learning comprise the majority of AI models. Support vector machines, artificial neural networks, random forests, decision trees, gradient boosting, and XG Boost are all components of supervised learning [100,101].

5.2. Support Vector Machines

A powerful class of supervised machine learning algorithms, support vector machines (SVMs) are primarily used for tasks involving regression and classification [102]. They are valuable tools for enhancing the decolorization process in wastewater treatment applications that have handled both linearly separable and non-separable cases, resulting in accurate estimations. SVMs were used to effectively analyze process operational conditions, molecular properties, and catalytic properties of compounds under study in practical photocatalysis applications by minimizing abstract mathematical theories and providing concrete numerical data [103,104,105]. SVM emerged as a powerful tool in photocatalysis, particularly in predicting the efficiency of photocatalytic processes. Its ability to analyze and forecast pollutant degradation in photocatalytic processes was employed to forecast the breakdown of the pharmaceutically active compounds diclofenac, caffeine, ibuprofen, and organic dyes, as well as heavy metals in water [105,106]. Biochar–ZnO nanocomposites were developed and employed as photocatalysts to degrade atrazine (93%) and 2,4-D (90%) in 90 min. Several operational factors were tuned to verify their efficacy. SVM, RF, and ANN are examples of machine learning models that were used to predict deterioration performance. Of all of them, SVM was the most accurate (R² = 0.95, RMSE = 4.93, MAE = 4.38) [107]. SVMs are an area of machine learning algorithms that apply kernel functions to translate data into higher-dimensional feature spaces [108]. SVMs, as part of machine learning frameworks, can analyze large datasets to identify patterns and relationships among variables, leading to more reliable predictions of photocatalytic performance [109]. A functioning principle of support vector machines (SVMs) is shown in Figure 7. A hyperplane is built to efficiently classify several classes when dealing with linearly separable data. In situations when the dataset cannot be separated linearly, the data is projected into a higher-dimensional space using kernel functions, which allows for the construction of an appropriate hyperplane. SVM aims to find the optimal hyperplane that separates data points from different classes in a feature space. The hyperplane has been defined by Equation (3).

W^Tx + b = 0

(1)

In Equation (1), x is the input feature vector that represents the properties of the photocatalytic material and reaction conditions, and b is the bias term that modifies the position of the hyperplane. W is the weight vector that defines the orientation of the hyperplane, and W^T is the sum of all weighted vectors used in the hyperplane [110,111]. The W^Tx + b > 0 is classified as a positive class, the data point is classified into one category (e.g., high photocatalytic efficiency), or W^Tx + b < 0 is classified as a negative class; it belongs to another category (e.g., low photocatalytic efficiency). The goal of SVM is to maximize the margin between the closest support vectors of different classes. Maximizing this margin enhances the model’s robustness and generalization capabilities. The margin can be expressed as Equation (2). In Equation (3), optimal hyperplane is used to formulate the optimization problem. Once the optimization is solved, the decision function for a new point x is given in Equation (4) [109].

$$\mathrm{Margin}={\displaystyle \frac{\left|\left|W\right|\right|}{2}}$$

(2)

$$\mathrm{Minimize}=1/2{\displaystyle \frac{\left|\left|W\right|\right|}{2}}$$

(3)

f(x) = sign (W^Tx + b)

(4)

SVM’s strength is its use of kernel functions to manage nonlinear interactions. SVM may project the original feature space onto a higher-dimensional space where linear separation is possible by using a kernel technique. For nonlinear data sets, common kernel functions include the linear kernel in Equation (5), the polynomial kernel in Equation (6), and the radial basis function kernel in Equation (7). Function approximation issues are resolved by SVM using a training set that is represented as pairs of instances, (X_i,X_j)_, where X_i and X_j are the input samples [102,112].

(xi,xj) = xiTxj

(5)

K(xi,xj) = (xiTxj + 1)d

(6)

$$\mathrm{K}(\mathrm{xi},\mathrm{xj})=({\displaystyle \frac{\exp \left(-y\left|\left|xi-xj\right|\right|\right)}{2}})$$

(7)

5.3. Artificial Neural Networks (ANNs)

Artificial neural networks (ANNs) are digital representations of the human brain that are made up of layers of interconnected nodes. They are made to identify trends and gain knowledge from information. The high prediction accuracy of the optimized ANN architecture suggests that it is effective in improving photocatalytic processes [113]. The ANN model is made up of an input layer that receives experimental features (such as catalyst dose and dye concentration), numerous hidden layers that perform nonlinear transformations using weighted connections and activation functions, and an output layer that predicts degradation efficiency. Bias nodes improve learning flexibility, while training changes weights to reduce errors. The study report used genetic algorithms in conjunction with ANNs to forecast and optimize dye photodegradation through nano ZnO-anchored glass fiber exposure to sun radiation [114]. In addition, ANNs were essential for forecasting photocatalytic process performance, which helped create effective environmental purification photocatalysts [46]. Researchers could improve the overall sustainability of photocatalysis processes for energy-efficient applications, anticipate degradation rates with high accuracy, and optimize process parameters by utilizing ANNs [115]. In Figure 8, the fundamental unit of an ANN was the artificial neuron, which had a single output, a weighted input, and a transfer function. Three layers make up a typical neural network architecture: input layers, which analyze information; one or more hidden layers, which carry out the majority of internal processing and are in charge of identifying patterns; and output layers, which produce and display the network’s final outputs as depicted [60,104,116]. Each neuron was activated by the weighted sum of its inputs. The activation signal then passes via a transfer function to produce a single output. The design, learning rule, and transfer function of a neural network determine its overall behavior [117]. ANNs were used for pollutant degradation forecasting in photocatalysis. ANNs offer effective modeling of nonlinear processes and reaction condition optimization in wastewater treatment [114,118]. They helped to remove pollutants from textile wastewater treatment by predicting the degradation efficiency of photocatalysts under different reaction situations [118], and they helped with the degradation of chlortoluron in wastewater using a continuous photocatalytic fixed-bed reactor using TiO₂. A maximum removal effectiveness of 94% was achieved under optimal conditions. ANN performed somewhat better than SVM when machine learning models were utilized. Combining ANN with a multi-objective genetic algorithm yielded the optimal operating conditions by striking a balance between high degradation efficiency and low energy consumption [119]. The key objective of this study was to use graphene oxide-doped titanium dioxide (GO-TiO₂) for the photocatalytic degradation of levofloxacin (LVX). The two modeling approaches that were employed were RSM and ANN. While RSM adjusted the process parameters and achieved approximately 80% degradation efficiency (R² = 0.88), the ANN model showed superior predicted accuracy with R² = 0.97. In comparison to RSM, these data show how effectively ANN predicts photocatalytic degradation performance [120]. ML was used to predict the photocatalytic effectiveness of g-C₃N₄/CdS/MoS₂ nanocomposites, reducing the requirement for resource-intensive studies. Following the evaluation of four algorithms (RF, DT, SVM, and NN), NN had the greatest accuracy (R² = 0.70, R² ≈ 0.84, RMSE ≈ 4.2). Both models anticipated methylene blue (MB) degradation under 180 min of sunshine to be between 83 and 84%, which was quite similar to the experimental figure of 86%. The results demonstrate that by effectively capturing complex nonlinear photocatalytic action, RF and NN offer reliable prediction power [121]. The dual-functional PeN@Fe₃O₄ coupled with Cd²⁺ effectively degraded the antibiotic sulfamethoxazole. The ANN model used to forecast the removal of Cd²⁺ performed exceptionally well (R² = 0.99975, MAE = 0.10906, MSE = 0.04399), suggesting great prediction accuracy and dependability. These results show PeN@Fe₃O₄ to be a sustainable and efficient technique for both organic pollutant degradation and heavy metal removal [122]. According to the study, rGO–SnS₂ nanocomposites were synthesized and shown to improve photocatalytic degradation of atrazine (ATZ) and 2,4-D in the presence of natural sunshine, removing up to 91% and 87% of the compounds, respectively. SVM, ANN, and Gaussian Process (GP) predictive modeling showed that ANN performed better than the other models (R² = 0.99, error = 0.002), offering a dependable tool for photocatalytic wastewater treatment optimization and scalability [123].

5.4. Diverse Architectures of Artificial Neural Networks (ANNs)

The neural networks that are most commonly used in wastewater treatment procedures are feed-forward and recurrent neural networks, and radial-based function [118].

5.4.1. Feed-Forward Neural Networks (FFNNs)

Feed-forward neural networks (FFNNs) are particularly effective at modeling complex, nonlinear relationships between multiple operational parameters and treatment efficiency in photocatalytic processes. By leveraging FFNNs, researchers can enhance predictive accuracy and optimize treatment conditions to improve pollutant degradation performance. FFNNs are a class of neural networks in which information flows in only one direction, from the input layer, through one or more hidden layers, to the output layer, without forming any cycles or feedback loops. This unidirectional propagation means that the network does not retain memory of past inputs, and its weights remain static once trained. Each neuron in an FFNN process weights inputs and passes them through an activation function, which can be linear, step-based, or nonlinear (such as sigmoid, ReLU, or tanh), depending on the modeling needs. FFNNs can be structured as single-layer networks (e.g., perceptrons) or as multi-layered architectures, such as multi-layer perceptrons (MLPs) or radial basis function networks (RBFNs). Their high controllability and robustness make FFNNs well-suited for handling noisy and complex datasets often encountered in environmental and wastewater treatment modeling [124,125]. The application of FFNNs in photocatalysis is one of the many sectors in which they have been thoroughly investigated. Studies on parametrizing radiative characteristics have shown that FFNNs could match complicated properties of mixtures with excellent accuracy [126]. Figure 9 shows the feed-forward neural network with various inputs (X_n) with weight (W_n) in the input layer, the number of the hidden layer (H_n), and the output layer (y) in the forward direction. Feed-forward neural networks (FFNNs) have been used to model complicated systems and forecast results based on input variables in a variety of scientific domains, including photocatalysis. The breakdown of water contaminants, including Niflumic acid (NFA), by photocatalysis has been modeled using FFNNs. In order to do this, neural networks are used instead of kinetic equations to forecast the rates of degradation depending on factors like substrate and catalyst concentrations [127]. To predict the photodegradation of phenolic compounds (PHCs) using black TiO₂ under visible light, the feed-forward neural network (FFNN/ANN) was evaluated in conjunction with RF and ERT models. The FFNN performed worse than the ERT model, producing comparatively higher errors and lower R² values, despite demonstrating an acceptable capacity for prediction. When it came to modeling the degradation of PHCs, the ERT model performed better than the FFNN, producing predictions that were more accurate and closely matched the experimental mean (73.35% vs. 72.79%) [128]. In this study, the photocatalytic degradation efficiency of boron-doped ZnO nanoparticles produced by a mechanochemical method was modeled and predicted using a single-layer feed-forward neural network (FFNN). A high coefficient of determination (R² = 0.9810), which shows good agreement between experimental and anticipated findings, was attained by the ANN model, demonstrating its exceptional predictive capacity. In comparison to undoped ZnO, the improved photocatalyst, 1 weight percent B-ZnO calcined at 700 °C, obtained 99.94% methyl orange degradation under UVA light. FFNNs can guide the optimization of photocatalytic systems for wastewater treatment, as these data demonstrate [129].

Single-Layer Perceptron (SLP)

The Single-Layer Perceptron (SLP) is a foundational artificial intelligence model inspired by the way biological neurons process information. It produces outputs as a smooth, nonlinear function of the weighted sum of its inputs. Due to its simplicity and learning capabilities, the SLP has been applied across various fields for tasks involving statistical regression and binary classification [130]. In essence, an SLP employs linear logistic regression to perform binary classification tasks (e.g., 0 or 1 outcomes), making it one of the earliest models in the development of artificial neural networks. Despite its simplicity, it serves as a fundamental building block in neural network design, and ongoing research continues to enhance its training algorithms and broaden its applicability [131]. As illustrated in Figure 10, the single-layer feed-forward neural network (SLFFNN) consists of xₙ input neurons derived from experimental photocatalysis datasets. Each input is assigned a corresponding weight wₙ, which influences the computation process. The weighted inputs are passed through a nonlinear activation function f(x) to generate the final output y, representing the predicted photocatalytic efficiency and validating experimental results. Equation (8) represents this activation function, which governs the network’s response and output generation.

y = f(x)

(8)

$$y={\sum }_{i=0}^n\mathrm{X}\mathrm{n}\mathrm{W}\mathrm{n}+\mathrm{b}$$

(9)

In Equation (9), the bias inputs are denoted by b, while the weights for the (X_n) connections are denoted by (W_n). The function f(x) exhibits a nonlinear excitation function behavior. The output of the model is expressed as in Equation (10) [132].

y = X₁W1 + X₂W₂ + X₃W_3……XnW_n + b

(10)

Multi-Layer Perceptron (MLP)

An artificial neural network with one or more hidden layers that can solve complicated issues using pattern recognition and function regression is called a multi-layer perceptron (MLP) [133,134]. The number of neurons in hidden layers and initial weights were two factors that affected the performance of MLPs. These factors could be improved using methods like hyperparameter optimization models using genetic algorithms and backpropagation [120]. Figure 11 illustrates the MLP neural network with multiple layers. Neural network theory has developed recently, showing that every multi-dimensional function can be accurately predicted with high accuracy using multi-layer feed-forward neural networks that include one hidden layer of neurons. In addition, a multi-layer identical feed backward model for ANNs was presented, along with simulation results and learning and recall techniques [135]. Many researchers have successfully used an MLPNN for a variety of purposes. Alsaffar et al.’s study, which used a multi-layer perceptron neural network to predict 1,2-dihydroxybenzene degradation in the setting of photocatalysis, revealed that the volume of the oxidant had a major impact on the degradation process [136]. The output of the neurons in the MLP is typically computed using Equation (11) as follows:

$${\mathrm{Y}}_{\mathrm{x}}(\mathrm{n})={\sum }_{i=o}^{n}f(Wn·Yn)+b$$

(11)

where Y_x(n) is the output of X_n input neurons in the nth layer. The W_n is the weight of input, y_n number of hidden layers, b is the bias, and f is the activation function, which can vary based on the network design [118]. The study examined the use of a multi-layer perceptron neural network (MLPNN) trained with Bayesian regularization for the predictive modeling of photocatalytic phenol degradation from crude oil effluent. To train the MLPNN, 26 datasets created using the Box–Behnken experimental design were employed. The input variables were irradiation time, starting phenol concentration, photocatalyst dose, and solution pH, while the output layer included phenol degradation [137]. Lenzi and coworkers reported that reactive blue dye 5G can be degraded by photocatalysis using Fe/TiO₂, Fe/ZnO, and Fe/TiO₂–ZnO. The metallic charge, calcination temperature, and time discoloration were the input data for the MLPPNN model, while the discoloration magnitude was the output value. The model’s output was outstanding, demonstrating highly precise data catalysts such as TiO₂/Fe [138].

5.4.2. Radial Basis Function (RBF)

The input layer, hidden layer, and output layer that make up an RBF network’s structure have immediate resemblances with those of MLPs, but RBFs usually have only one hidden layer and convert the input space into a higher-dimensional space with the help of a Gaussian function [139]. Radial basis functions (RBFs) have been widely applied in several domains, demonstrating their effectiveness in tasks including particle coagulation evolution, neural network training, and differential equation solution [140]. Through the use of RBFs’ distinct features, scientists could be able to create fresh ideas for photocatalysis procedures, taking advantage of the flexibility and computing power that RBF approaches provide in a variety of fields [141,142]. The Gaussian function (activation function) is indicated by the following Equation (12):

$${\mathrm{O}}_{\mathrm{j}\left(\mathrm{X}\right)} = \exp {\displaystyle \frac{\_\left|\right|Xj\_Cj\left|\right|}{\infty j}}$$

(12)

where Cj and ∞j are the center width and the peak width, respectively [143]. The input nodes in the first layer send unweighted inputs to each node in the hidden layer, and each hidden node has an RBF as the transfer function. The outputs of these nodes are then weighted and added together to create the final output, and the output nodes that calculate the weighted sum of the outputs from the hidden nodes are in Equation (13).

$${\mathrm{Y}}_{\mathrm{I}} ={\sum }_{I=1}^n\mathrm{W}\mathrm{j}\mathrm{i}\mathrm{O}\mathrm{j}\left(\mathrm{x}\right)$$

(13)

where wji represents the weights of the connection between the hidden layer, i, and output layer, j, and Oj(x) is obtained from above [144]. RBF architecture is shown in Figure 12, which contains an input layer that is processed through a single hidden layer with a Gaussian as an activation function and gives output. The input layer, hidden layer, and output layer are the three primary layers that make up the network. The input layer, which is displayed on the left, is made up of nodes with the labels I1, I2, I3, and In. Each represents a feature or variable from the input data. Without carrying out any calculations, these nodes are used to send raw data into the network. The network may learn intricate patterns and representations from the input data due to the hidden layer. The output layer, which has nodes with labels, receives the signals after they have been processed. The task at hand determines how many nodes this layer contains [145].

Dashti and coworkers demonstrated the photodegradation of pesticides by ZnO-based photocatalysts for water treatment by different AI models. The input parameters included light source, dopant mass ratio, pesticide concentration, solution pH, catalyst dose, and irradiation duration, while considering physicochemical characteristics like molecular weight and water solubility. Compared to other models, RBF had the highest modeling accuracy, with an average absolute relative deviation (AARD) of 4.80% and the highest determination coefficient R² = 0.978 [11].

5.4.3. Recurrent Neural Networks (RNNs)

RNNs contain a feedback loop that returns data to the hidden layer, unlike feed-forward neural networks (FFNNs), which have a one-way data flow (from input to output through hidden layers). This loop produces a time lag effect, which helps the RNNs in remembering previous time steps. RNNs are a type of deep learning architecture that is designed to handle sequential or time-series data. Unlike feed-forward networks, RNNs have internal memory that allows them to retain information from previous inputs, capturing time-dependent relationships in the data [146]. This property makes RNNs ideal for modeling dynamic processes like pollutant degradation, where photocatalysis effectiveness can fluctuate over time due to changing environmental conditions, catalyst behaviors, or contaminant concentration. RNNs improve prediction accuracy in complicated, time-dependent degradation systems by taking into account both current and historical input sets. RNNs are widely used in a variety of sectors, one of which is photocatalysis. RNNs are used in the wastewater treatment context to predict how effectively photocatalysts will degrade in terms of removing contaminants from textile waste [118]. RNNs were able to keep up context over time by keeping track of information about past inputs in a hidden state. All things considered, RNNs are essential for process optimization, result prediction, and efficiency enhancement in a wide range of applications, such as spectroscopic analysis and photocatalysis [147]. Furthermore, RNNs were effective in predicting the electromagnetic spectrum evolution of pulses in nonlinear periodically poled waveguides, showing excellent agreement with previous numerical models and providing flexibility in the analysis of optical pulse propagation [148]. The study shows that biochar made from tea waste is a low-cost, efficient adsorbent for the simultaneous removal of aromatic compounds, dyes, and agrochemicals from wastewater. Overall elimination efficiency for combined pollutants was 82.66% under ideal conditions (5 mg mL⁻¹ dose, pH 2, 60 min). When RNN and CatBoost machine learning models were used to forecast removal performance, RNNs demonstrated higher accuracy, with R² = 0.960 [149]. An RNN’s structure, comprising an input layer, hidden layers, and an output layer, is depicted in Figure 13. RNNs, in contrast to simple neural networks, feature feedback connections that help them store information from earlier stages. The RNN receives an input that updates its hidden state h_t based on the previous hidden state h_t- and the current input. Mathematically, this can be represented in Equation (14).

h_t = σ(W_hh_t − 1 + W_xx_t + b)

(14)

where h_t is the hidden state, and x_t is the input at time step t. The W_h and W_x are weight matrices, b is a bias term, and σ\sigma σ is an activation function. The output of each layer can be calculated by Equation (15).

y_t = g(w_yh_t + c)

(15)

where y_t is the output at time step t, w_y is a weight matrix for the output layer, is the output bias, h_t is the hidden state and g is often softmax for classification tasks [150]. Long short-term memory (LSTM) networks are a complex variation of ordinary RNNs that were created to address their shortcomings, especially the vanishing and ballooning gradient issues in lengthy sequences. The memory cell and gating mechanisms (input gate, forget gate, and output gate) that LSTMs have enable them to selectively store, update, or discard data over time steps [137]. Because of their ability to better represent long-term dependencies than conventional RNNs, LSTMs are a preferred option for time-series prediction tasks in photocatalysis and other fields [151,152].

5.5. Tree-Based Models

The efficiency of combining decision trees, XGBoost, random forests, and gradient boosting in machine learning for classification and prediction problems has been thoroughly investigated. These algorithms are particularly known for their versatility and robustness across a wide range of data domains, including photocatalysis [153]. The input for these models is used for predicting the effect. While random forests and gradient boosting improve performance through collective approaches, decision trees make up the basis of the model. These models help identify key variables influencing degradation, such as light irradiation time and catalyst type, which are crucial for enhancing photocatalytic efficiency [154]. Rhodamine-B photocatalytic degradation efficiency employing TiO₂ catalysts is predicted using a machine learning model. The efficacy of XGBoost and other models, such as random forest and gradient boosting decision tree, was confirmed by validation against a variety of composite catalysts [135,155]. Decision trees are straightforward yet effective models for issues involving regression and classification. Although they are simple to understand, they may overfit, particularly when dealing with complicated data [156]. Combining the predictions of several trees, random forests, and a combination of decision trees reduces overfitting and enhances durability and generalization. Gradient boosting creates models one after the other, fixing the errors of the models that came before it. This approach enhances the model’s precision and transparency. XGBoost is a robust gradient-boosting solution that is renowned for its effectiveness and speed. It can handle big datasets well because of a complex algorithm, which makes it appropriate for data with high dimensions, like those seen in photocatalysis [157,158,159]. Apart from conventional ensemble techniques, Extreme Gradient Boosting (XGB) has been acknowledged as a scalable and extremely effective gradient boosting solution, widely used in a variety of prediction tasks [160]. In Figure 14, the effectiveness of pollutant removal from wastewater may be predicted using tree-based models, including decision trees, random forests, XGBoost, and gradient boosting models. Every model removes the bias in prediction before it occurs due to efficiency in its design. Parameters including pH, temperature, catalyst dose, pollutant concentration, and irradiation time are frequently included as input variables for these models. The output variable is the pollutant’s expected removal efficiency, which may be expressed as a percentage removal or residual concentration. Through the examination of the connections between these input variables and the output variable, these models can offer important information about the ideal operating parameters for effective pollutant removal in wastewater treatment [153,161]. The degrading efficiency of ketoconazole (KTC) was predicted using a variety of machine learning models, such as SVM, ANN, and tree-based techniques. With R² values of 0.992–0.998 and low RMSE and MAE scores, GBR outperformed the others, demonstrating its higher predictive accuracy. Reaction time was shown to be the most significant element, and GBR’s outstanding performance is ascribed to its capacity to capture intricate nonlinear interactions between parameters, including catalyst dosage, pH, and reaction time [162].

5.6. AI Model Development and Benchmarking in Photocatalytic Treatment

The effective use of AI in wastewater treatment is largely dependent on careful model design, which includes the development of features, data selection, model architecture, and assessment techniques. A comprehensive dataset has been compiled, incorporating experimental data points from literature sources, along with extensive datasets obtained from experiments and published studies, to provide data for AI models as input for prediction and optimization of parameters [53,135]. The dataset needs to be preprocessed and cleaned before any AI model is trained. To guarantee data quality, preprocessing procedures are essential and include resolving missing values, encoding categorical variables, normalizing numerical characteristics, and eliminating outliers. To improve statistical reliability, the preprocessed dataset is then divided into training, validation, and testing subsets, usually in a 70:15:15 [163] ratio, or put through k-fold cross-validation. Using hyperparameter optimization methods like Grid Search or Random Search, models are refined after being trained on the training set. In order to choose a suitable AI model, important hyperparameters like the spread factor and the maximum number of neurons had to be systematically adjusted. These adjustments were essential for improving the predicted accuracy of the model [11]. In order to optimize photocatalyst discovery and utilization, selecting software for AI-driven photocatalysis requires integrating machine-learning models with decision-making frameworks. Techniques for software selection ensure that the tools chosen meet the specific needs of the research. Python-based program, Origin Lab, Neuro Selection, Java Pro, and R statistical tools are among the most often used programs based on model prediction. The most popular software and Python libraries for creating AI models that optimize for experimental factors influencing the photocatalysis mechanism are displayed in

Table 2. Specialized AI/ML platforms for materials science.

Platform/Tool	Type	Key Features	Application Areas	Ref (Accessed on 8 September 2025)
Materials Project	Database and ML Tools	Materials database based on DFT with machine learning methods for predicting properties	Battery materials, thermodynamic stability, and crystal structures	https://next-gen.materialsproject.org/ml
AFLOW	Framework and Database	Data production for material attributes that is automated	Analysis of mechanical properties, symmetry, and high-throughput screening	https://aflowlib.org/
Citrination	AI/ML Platform	Materials discovery powered by machine learning (Citrine Informatics)	Material selection, property prediction, and process optimization	https://citrine.io/
NOMAD	Repository and Toolkit	ML-ready data and analysis tools for materials	Analysis of electronic structures and dataset benchmarking	https://nomad-lab.eu/nomad-lab/
Matminer	Python Library	Tools for materials data feature extraction and machine learning	Feature development and data mining	https://pypi.org/project/matminer/
Atomate	Workflow Automation	Handles procedures for high quantities of materials	Electronic framework and synthesis path modeling	https://atomate.org/
ASE	Python Library	Setting up and analyzing an atomic model	Simulations using DFT/MD and modeling of the manufacturing process	https://pypi.org/project/ase/
MODNet	ML Model Framework	Property prediction using materials descriptors	Supervised learning from structured data	https://github.com/ZHKKKe/MODNet
DeepChem	ML Library	Deep learning in chemistry and materials science	Quantum chemistry and molecule/material property prediction	https://deepchem.io/
Open Catalyst Project	Dataset + Model	ML models for catalyst discovery	Reaction energy prediction and catalyst screening	https://opencatalystproject.org/
MEGNet	Graph Neural Network	GNN models for material property predictions	Prediction from atomic connectivity graphs	https://opencatalystproject.org/
QMOF Database	Dataset + Models	Data about MOFs from quantum chemistry	Gas storage and electronic and thermal properties	https://github.com/Andrew-S-Rosen/QMOF
SISSO	Compressed Sensing ML Tool	Identifying descriptors using sparsifying agents	Predicting material efficiency with regression modeling	https://github.com/rouyang2017/SISSO
SchNetPack	Deep Learning Model	Learning fundamental interactions from start to finish	Dynamics of molecules and property forecasting	https://schnetpack.readthedocs.io/en/latest/
TPOT	AutoML Tool	Optimizing ML pipelines with genetic programming	Predicting the behavior of materials and creating auto models	https://epistasislab.github.io/tpot/latest/
Polymer Genome	ML Platform	ML for predicting polymer properties	Electrical power, thermal energy, and mechanical characteristics	https://www.polymergenome.org/
BigSMILES	Polymer Representation	Polymer uniform representation for machine learning input	Modeling the structure and properties of polymers	https://olsenlabmit.github.io/BigSMILES/docs/line_notation.html

and

Table 3. General-purpose AI/ML frameworks in materials science.

Framework/Tool	Type	Key Features	Use in Materials Science	Ref (Accessed on 8 September 2025)
Scikit-learn	ML Library (Python)	Traditional machine learning algorithms (e.g., SVM, RF, PCA)	Quick prototyping, grouping, regression, and classification	https://scikit-learn.org/stable/
TensorFlow	Deep Learning Framework	Building neural networks and machine learning from scratch	Virtual models, GNNs, and deep neural networks for predicting properties	https://www.tensorflow.org/
Keras	High-level DL API	Neural network protocol that is simple to use (TensorFlow as backend)	CNNs/RNNs for structure–property estimates and spectral information	https://keras.io/
PyTorch	Deep Learning Framework	A flexible deep learning system based on graphs	GNNs and deep learning for simulated and structured recognition	https://pytorch.org/
XGBoost	Gradient Boosting Framework	Effective gradient boosting for classification and regression	Quick and precise material property prediction	https://xgboost.ai/
LightGBM	Gradient Boosting Framework	Acceleration and memory efficiency optimization	Predicting properties with high-dimensional information	https://lightgbm.readthedocs.io/en/stable/
Auto-sklearn	AutoML Library	Hyperparameter-tuning and robotic machine learning	Quick creation of forecasting models	https://pypi.org/project/auto-sklearn/
DGL	GNN Framework	Performing deep learning using graphs (supports PyTorch/TensorFlow)	Crystal framework and interatomic interactions modeling	https://www.dgl.ai/
PyCaret	ML Library	ML model assessment and training made simpler	Rapid implementation of models and assessment	https://pycaret.org/

[164,165]. The process for using AI in photocatalysis is shown in Figure 15. Data collection, including database information, experimental investigations, and book reviews, is the first step in the process. A software design is created, and an appropriate AI model is chosen based on this data. In order to guarantee forecast accuracy and dependability, the model parameters are subsequently refined and contrasted with experimental parameters in an ongoing feedback loop.

When AI datasets are trained and analyzed with models that predict efficiency and are reasonably priced, experiments are avoided. These metrics provide information about the model’s durability and accuracy. A benchmarking table, which lists the datasets used, model types, input features, and performance metrics attained under uniform settings, is frequently provided in high-quality studies for a clear comparison [166]. A significant drawback of current AI-based photocatalytic wastewater treatment research is its omission of external validation, which is crucial to guaranteeing the model’s performance in practical settings. To address this, researchers should employ transfer learning strategies, conduct a sensitivity analysis to assess the model’s generalization ability, and incorporate external testing using separate datasets. Moreover, explainable artificial intelligence (XAI) techniques such as SHAP (Shapley Additive Explanations) and LIME (Local Interpretable Model-Agnostic Explanations) can be used to determine which input features have the most impact on the predictions, hence improving the transparency and interpretability of AI models [167]. SHAP works well for elucidating AI models in scientific fields, fostering more transparency, understanding, and trust. The model’s output is decomposed into each input variable’s additive contribution. Each feature’s contribution to the prediction, relative to a baseline (average) value, is computed by SHAP. For instance, SHAP explains the proportion of a model’s predicted 80% dye degradation that can be attributed to factors like time, dye concentration, and catalyst amount [163,168]. The efficacy of the models was evaluated by contrasting their predicted values and fit with the experimental data. The outcome metrics for the training and prediction sets were the root mean square error (RMSE), sum of squares error (SSE), sum of absolute error (SAE), mean squared error (MSE), and coefficient of correlation (R²) [47,169]. Ahmed and coworkers described the use of an experimental dataset from the degradation of methylene blue (MB) in wastewater and applied various AI models to predict maximum degradation efficiency. The study utilized the Scikit-learn library in Python to build the models and evaluated their predictive performance using metrics such as R², MSE, RMSE, MAE, and CV R² [158]. More than a thousand tests were conducted to investigate the effectiveness of photocatalytic water purification, resulting in a substantial dataset. ANN demonstrated the highest prediction accuracy among the machine learning models used (R² = 0.970, MAE = 2.70, MSE = 5.88). According to the SHAP study, the most important variables were the type of pollutant, its concentration, and the amount of catalyst used; the best catalyst was CuO/TiO₂. Strong CuO–TiO₂ interactions were validated by molecular dynamics simulations, confirming experimental results and pointing to a solid framework for sustained photocatalysis optimization [47].

Oviedo and coworkers reported algorithms such as neural network algorithms and Xtreme Gradient Boosting to create prediction models for photocatalytic deterioration. Using Spearman correlation, the link between the variables influencing RhB degradation was examined, providing information on catalyst and dye concentrations [170]. Sohrabi and coworkers reported gaining insight into the parameters influencing photocatalytic efficiency and, to enhance the design of photocatalysts, the study uses sensitivity analysis and machine learning. The prediction of deterioration efficiency was examined using a number of machine learning models, such as Polynomial Regression, support vector regression (SVR), XGBoost, random forest, gradient boosting, and artificial neural networks (ANNs). The ANN model outperformed the others, with a mean squared error (MSE) of 5.88, a mean absolute error (MAE) of 2.70, and a coefficient of determination (R²) of 0.970 [47].

$${\mathrm{R}}^2=1-\left({\displaystyle \frac{\sum (y_i-{\mathrm{ŷ}}_{i })}{\sum (y_i-\mathrm{ȳ})}}\right)$$

(16)

$$MAE=\left({\displaystyle \frac{1}{n}}\right)\sum |y_i-{\mathrm{ŷ}}_i|$$

(17)

$$RMSE=\surd \left[\left({\displaystyle \frac{1}{n}}\right)\times {\sum (\mathrm{y}\mathrm{ᵢ}-\mathrm{ŷ}\mathrm{ᵢ})}^2\right]$$

(18)

$$MSE=\left({\displaystyle \frac{1}{n}}\right)\times {\sum (\mathrm{y}\mathrm{ᵢ}-\mathrm{ŷ}\mathrm{ᵢ})}^2$$

(19)

$$CV {\mathrm{R}}^2=\left({\displaystyle \frac{1}{k}}\right)\times \sum {{\mathrm{R}}^2}_i$$

(20)

5.7. Hybrid AI Models

To improve the efficacy and efficiency of pollutant degradation, hybrid AI models are being included in photocatalysis for water treatment more and more. These models handle the complexity of water contaminants and treatment parameters by combining a variety of computational methodologies to optimize the photocatalytic processes [50,171,172]. Although supervised learning models like SVM, ANN, and tree-based algorithms are the main focus of this research, hybrid approaches are also becoming more popular due to their capacity to capture detailed nonlinear interactions. Among these, frameworks for the cascaded adaptive neuro-fuzzy inference system (ANFIS) are effective in environmental modeling. When dealing with nonlinear and multivariable datasets, ANFIS outperformed single-stage ANFIS in terms of prediction accuracy. The shown resilience of cascaded ANFIS in environmental systems underscores its potential as a future research avenue for combining hybrid AI with photocatalysis, even though its applications in photocatalytic wastewater treatment are limited [173]. Similar to this, hybrid deep learning algorithms that combine long short-term memory (LSTM) and artificial neural networks (ANNs) have been used to predict photocatalytic outcomes with remarkable accuracy (R² > 0.99, RMSE < 0.03), indicating their potential to improve wastewater purification technologies and increase the efficiency of Se(IV) reduction [174]. In order to improve simulation robustness, improve capture degradation mechanisms, and optimize process parameters, the study builds on this by proposing a hybrid modeling approach for photocatalysis in advanced water treatment that integrates quantitative-structure activity/property relationships (QSAR/QSPR). The efficacy of photocatalytic systems is eventually increased by this integration, which also improves the predictive performance of photocatalytic models and offers mechanistic interpretability [175].

6. Innovative Catalysts for Effective Wastewater Purification

6.1. Titanium Dioxide (TiO₂)

Photocatalysts such as TiO₂ can facilitate photoreactions by enhancing the absorption of light by the reacting species. TiO₂ is preferred because of its low prices, stability, and efficiency; various polymorphs, such as rutile, brookite, and anatase, have distinct characteristics. TiO₂’s main drawback is its wide band gap, which limits its photocatalytic activity to ultraviolet light, which makes up a very small portion of the solar spectrum. The modifications of TiO_2, such as changing the band gap and crystal size, can increase photocatalytic efficiency [86,176]. Environmental applications such as air filtration and solar energy conversion value TiO₂ for its nontoxicity, affordability, and efficiency. Although anatase had a large band gap that limited light absorption, it was the favored polymorph due to its greater photocatalytic activity and durability. TiO₂ could be made to absorb visible light more efficiently in a variety of applications by doping it with metals and nonmetals to modify the band gap [86].

Table 4. Titanium dioxide (TiO₂) used as a catalyst to degrade different pollutants using AI models.

Sr#	Pollutant Degraded	Structure of Pollutant	Catalyst	AI Method Used	Light Source	Factor Effects	Ref.
1	Methylene Blue		Fe3O4TiO2 Ag magnetic nanocomposite	ANN	UV lamp	Initial dye concentration, pH, and temperature	[177]
2	Malachite green dye		TiO2	ANN, BBD	UV light	Catalyst dosage, time, and initial dye concentration	[178]
3	Selenium Se (VI)	_	TiO2BiOBr fabric	ANN, LSTM	Visible light	Machine learning models, recyclability, light, and catalyst stability	[174]
4	Congo red		Bi–TiO2 nanomaterials	RF, GBDT, XG Boost	Visible light	Initial dye concentration, pH, and temperature	[179]
5	Acid red 14 (AR14)		TiO2	RBF, ANFIS	UV light	Irradiation time, flow rate, and catalyst concentration	[180]
6	Tetracycline (TC)		Cobalt atoms as a co-catalyst on TiO2 nanosheets	ANN, ANFIS	UV, visible light	Co catalyst dose, time of irradiation	[181]
7	Polycyclic aromatic hydrocarbons (PAHs)	_	Rutile TiO2	ANN	UV-C light	UV wavelength, temperature, and catalyst concentration	[182]
8	Beta-naphthol		(TiO2) NPs	ANN-PSO and ANFIS-PSO	UV light	Aeration rate, acidity, catalyst content, and impurity concentration	[183]
9	Amoxicillin (AMX)		Ni2P–TiO2 (NPT)	ANN	UV irradiation	pH, catalyst dose, and irradiation time	[184]
10	Rhodamine B (RhB)		Nd-TiO2	ANN	UV light, solar light	Doping concentration, light source, presence of humic acid, and CaCl2	[185]
11	Tartrazine		TiO2	ANN	Solar light	pH, TiO2 concentration, initial pollutant concentration, and solar radiation intensity	[186]
12	2,4-dichlorophenol (2,4-DCP)		(Fe3O4/TiO2/Ag), TiO2	SGB, ANN, ANFIS, GA-ANFIS, PSO-ANFIS	UV, visible light	pH, temperature, light intensity, pollutant nature, and type of catalyst	[32]
13	Cetirizine hydrochloride		(Cu–TiO2) nanoparticles	SVM with IGWO	UV, visible light	Copper loading, solution pH, catalyst dosage, and initial pollutant concentration	[88]

shows the Titanium dioxide (TiO₂) used as a catalyst to degrade different pollutants using AI models.

6.2. Zinc Oxide (ZnO)

AI improves reaction results and synthesis conditions to greatly increase the efficiency of ZnO photocatalysis processes. AI demonstrates encouraging progress in improving the photocatalytic effectiveness of zinc oxide (ZnO) photocatalysts by adjusting their band gap, especially in visible light [187]. The improvement is important because ZnO’s large band gap naturally reduces its ability to absorb UV light, which restricts the material’s application in photocatalysis [188]. AI models, particularly ANNs and GA, have been pivotal in predicting optimal conditions for ZnO NP synthesis and photocatalytic processes [189].

Table 5. Zinc oxide catalysts that degrade different pollutants using an AI model.

Sr#	Pollutant Degraded	Structure of Pollutant	Catalyst	AI Method Used	Light Source	Factor Effect	Ref.
1	Methylene blue (MB)		ZnO/MgO	ANN	UV light	Catalyst dose, initial dye concentration level, and photodegradation time	[91]
2	Yellow 84 dye		(ZnO) NPs	(GPR), RBF, ANF, MLP	UV light	Catalyst dose, initial dye concentration level, and photodegradation time	[190]
3	Low-density polyethylene (LDPE)	_	FKMW-ZnO NPs	RBFNNES	_	Pollutant concentrations, catalyst concentrations, time, and pH	[145]
4	Eosin Y		ZnO/SnO2 nanocomposites	FFNN	UV light	Crystallite size, surface area, absorption edge, TOC values, time of reaction, and catalyst concentration	[191]
5	Methylene Orange		ZnO-MgAl layered double hydroxides	ANN	UV light	Temperature, irradiation time, catalyst amount, and dye concentration	[192]
6	Red dye		ZnO	ANN, ANFIS	UV light	pH, amount of ZnO, and initial dye concentration	[193]
7	Acid blue 113 (AB113) dye		ZnO	ANN	Ultrasound irradiation	Reaction time, pH, ZnO dosage, ultrasonic power, and persulphate dosage	[194]
8	Rhodamine 6G dye		TiO2-ZnO/BAC composite	ANFIS, ANN	UV light	Coupling of material, support of material, and light intensity	[195]
9	p-cresol		ZnO	ANN	UV irradiation	Irradiation time, pH, photocatalyst amount, and concentration of pollutant	[196]
10	Metronidazole (MNZ)		ZnFe12O19/BiOI nanocomposite	ANN	UV irradiation	Contaminant concentration, pH, nanocomposite dosage, and retention time	[115]
11	Pesticide photodegradation	_	ZnO	CSA-LSSVM, RBF, PSO-ANFIS, MLP-ANN	UV, VIS light	Irradiation time, pH, light source, dopant mass proportion, catalyst dose, and starting pesticide concentration	[11]
12	Tetracycline		CdS/ZnO nanosheets	ANN and GBRT	VIS light	Temperature, pH, and light intensity	[197]
13	Phenol		ZnOFe2O3	ANN, MLPNN	Solar irradiation	Initial pollutant concentration, photocatalyst dosage, photocatalysis irradiation time, and solution pH	[137]

shows the Zinc oxide catalysts that degrade different pollutants using an AI model.

6.3. Cadmium Sulfide (CdS)

AI is increasingly being integrated into photocatalysis for wastewater treatment, particularly using CdS-based materials. Cadmium sulfide (CdS) photocatalytic activity is highly dependent on its bandgap energy in a number of settings. Different crystalline phases of CdS have been seen; the bandgap of cubic CdS is 2.24 eV, whereas that of hexagonal CdS is 2.17 eV. Because of its higher energy carrier migration and lower mixing rates, the cubic phase exhibits stronger photocatalytic performance and is hence more useful for hydrogen evolution reactions when subjected to visible light [198]. Furthermore, improved photocatalytic degradation of pollutants is linked to the creation of heterogeneous structures with reduced band gaps, suggesting that lower bandgap values can promote higher light absorption and catalytic activity. In general, changes in structure that control bandgap energy were essential for optimizing CdS photocatalysts in a variety of applications [199]. Research on CdS nanocomposites, such as chitosan/polyaniline/CdS, employed ANNs to predict decolorization efficiency. This indicates that AI can effectively model and optimize photocatalytic reactions involving Cd [200].

Table 6. Cadmium sulfides used as catalysts to degrade different pollutants using the AI model.

Sr#	Pollutant Degraded	Structure of Pollutant	Catalyst	AI Method Used	Light Source	Factor Effect	Reference
1	Cefoperazone		CdSg-C3N4	ANN	Visible light	pH, irradiation period, and catalyst dose	[201]
2	Methylene blue		Nano CdS diatomite	ANN	UV light	Composite weight, pH level, dye concentration, and light intensity	[202]
3	Reactive Blue 19 (RB19)		CS/PAni/CdS nanocomposite	ANN	Visible light	Light intensity and nanocomposite dosage	[200]
4	Tetracycline (TC)		CdS/ZnO nanosheets	ANN and GBRT	UV and ultrasonic light	Catalyst surface area and light source	[197]
5	Tetracycline (TC)		CdS/TiO2 nanosheets/graphene nanocomposites	ANN, ANFIS	Visible light	CdS molar ratio, surface area, and light intensity	[203]
6	Cefazoline (CFZ)		CdS-ZnFe2O4 nanocomposites	ANN	Visible light	pH, time, catalyst concentration, and pollutant concentration	[204]

shows the cadmium sulfides used as catalysts to degrade different pollutants using the AI model

6.4. Tungsten Trioxide (WO₃)

AI is increasingly being integrated into the field of photocatalysis; in particular, tungsten trioxide (WO₃) is an increasingly common substance for photocatalysis because of its advantageous qualities, such as a band gap of 2.5 to 2.7 eV, high stability, and excellent electrochemical properties. Its uses include pollutant degradation, CO₂ reduction, and water splitting, making it a diverse option for sustainable energy solutions [205]. Doping with hydrogen and noble metals (silver) improves solar energy usage and broadens light absorption into the near-infrared spectrum. When coupled with other materials like hydroxyapatite (HAp), WO₃ has shown outstanding photocatalytic activity, increasing the breakdown efficiency of contaminants like MB to 88.689% [206,207].

Table 7. Tungsten trioxides (WO₃) used as catalysts to degrade different pollutants using the AI model.

Sr#	Pollutant Degraded	Structure of Pollutant	Catalyst	AI Method Used	Light Source	Factor Effect	Reference
1	Malachite green		ZnWO4	ANN	Uv	pH, contact duration, nano adsorbent dosage, initial MG concentration, and temperature	[208]
2	Cefixime		WO3/Co-ZIF nanocomposite	ANN, SVR	UV	pH level, reaction time, and catalyst amount	[209]
3	Methylene blue (MB)		BiFeO3-WO3 nanocomposite	AI(SVM)	Sunlight	Light intensity, dye concentration, and temperature	[105]
4	Methylene blue (MB)		CuWO4@TiO2	HGB model	UV	Light intensity, dye concentration, and temperature	[158]

shows the Tungsten trioxides (WO₃) used as catalysts to degrade different pollutants using the AI model.

6.5. Cesium Dioxide

The integration of AI in CeO₂ photocatalysis is revolutionizing the field by enhancing defect engineering, optimizing reaction conditions, and improving predictive capabilities. AI techniques, particularly ML, are being employed to address the challenges associated with photocatalytic processes [83]. ML algorithms have been utilized to control defect structures in CeO₂ nanostructures, crucial for their catalytic properties [210].

Table 8. Cesium dioxide as a catalyst that degrades different pollutants using an AI model.

Sr#	Pollutant Degraded	Structure of Pollutant	Catalyst	AI Method Used	Light Source	Factor Effect	Reference
1	Acid Orange 7 (AO7) dye		CeO2	ANN	UV light	Reaction time and pH	[88]
2	Methylene blue (MB) dye		rGO/Ag3PO4/CeO2 nanocomposite	ANN	Visible light	Initial dye concentration, pH, reaction time, and temperature	[211]
3	Cefazoline (CFZ)		CdS-CaFe2O4-CP	ANN	Visible light	pH of the solution	[204]

shows the Cesium dioxide as a catalyst that degrades different pollutants using an AI model.

6.6. Zirconium Oxide (ZrO₂)

This is a significant milestone, as it combines artificial intelligence with zirconium oxide photocatalysis for highly efficient and optimized water treatment using intelligent systems and advanced materials. Recent studies highlight the development of zirconium-based photocatalysts, such as zirconium oxide–carbon composites and zirconium-doped chromium IV oxide, which demonstrate effective degradation of organic compounds under various light conditions.

Table 9. Zirconium oxide as a catalyst that degrades different pollutants using AI models.

Sr#	Pollutant Degraded	Structure of Pollutant	Catalyst	AI Method Used	Light Source	Factor Effect	Reference
1	Organic dyes *	_	Zr-MOF	ANN	-	pH, contact time, MOF content, and dye concentration	[212]
2	Congo red		ZrO2/PdO-NPs	RF, XGB, GBR	Visible	pH, CR concentration, catalyst concentration, and irradiation time	[213]
3	Carbamazepine		TiO2/ZrO2 nano composite	ANN	UV radiation	Initial concentration of pollutant, pH of the solution, catalyst concentration, and time of UV irradiation	[125]
4	Basic red 46		CuO-doped ZrO2–ZnO (ZZC) nanocomposites	ANN	LED visible irradiation	Initial concentration of pollutant, pH, time, and catalyst loading	[214]
5	Amoxicillin (AMX)		ZrO2	ANN	UV light	pH, catalyst dose, and time	[215]

* Wastewater containing multiple dyes was used.

shows the Zirconium oxide as a catalyst that degrades different pollutants using AI models.

6.7. Limitations and Challenges of Data Sparsity and Overfitting in AI-Driven Photocatalysis: Mitigation Strategies

Even if artificial intelligence has the potential to revolutionize photocatalytic wastewater treatment, several inherent limitations must be carefully considered to ensure the scalability, interpretability, and reliability of AI-driven models. The expense, effort, and complexity of synthesizing and characterizing photocatalytic materials limit the availability of complete, high-quality experimental datasets. Consequently, data sparsity remains a major challenge. Predictive accuracy may suffer, and model training may be affected by this lack of data [216]. Model overfitting is a serious issue, especially when using deep learning techniques with small datasets. This can lead to poor generalization to new data since the model may collect noise rather than key patterns. The limited generalizability of AI models across various photocatalysts is another significant drawback. This is because models that have been trained on small or chemically homogeneous datasets frequently perform unreliably when applied to structurally or functionally varied systems [217]. These problems highlight the necessity of strong mitigation measures, such as regularization methods, data augmentation, transfer learning, cross-validation, and the addition of mechanistic descriptors that capture basic physicochemical characteristics. To move AI from a helpful computational tool to a trustworthy prediction framework that can hasten the logical design of next-generation photocatalysts for sustainable wastewater treatment, these constraints must be solved [218]. AI applications in photocatalysis face several challenges, including data sparsity, overfitting, limited generalizability, poor interpretability, lack of reproducibility, weak experimental integration, and dataset bias. The inability to interpret models is a significant issue, especially with complicated deep learning frameworks that frequently act as “black boxes.” This transparency undermines confidence in model-driven decision-making by making it more difficult for researchers to extract significant chemical or mechanistic insights from AI forecasts. Additionally, because much research lacks standardization in feature selection, dataset curation, and evaluation measures, reproducibility is still a major problem. These variations limit the creation of generally recognized benchmarks and make it challenging to compare models across various research teams. The poor integration of AI models with actual experimental procedures is another drawback [98,219]. The practical usefulness of many AI studies is limited because they are mostly theoretical or computational and have received little validation through laboratory experimentation. Additionally, if the training datasets have preferences towards particular catalyst families or reaction conditions, computational bias may be introduced, which would ultimately decrease the fairness and applicability of the model across a variety of systems. Standardized datasets, transparent model development, interdisciplinary cooperation, and a closer connection between AI predictions and experimental validation are all necessary to overcome these more general constraints [220,221]. Along with these challenges, it is important to highlight that overfitting and data sparsity can undermine the predictive accuracy of AI-based photocatalytic models. Transfer learning, which enables models trained on larger, related datasets to be adapted for specialized systems, offers a robust solution to these issues. This approach has been shown to significantly boost performance in materials science applications [222]. Concurrently, rigorous cross-validation methods (such as 10-fold cross-validation) [141] help prevent models from identifying spurious correlations and have proven effective in validating robustness in studies of photocatalytic degradation. When combined with regularization and kinetic feature selection, these strategies ensure improved generalizability and predictive reliability. As a result, such targeted mitigation methods are critical for reducing bias and achieving reproducible outcomes in real-world applications [223].

7. Conclusions

This review offers a comprehensive overview of current methods, materials, and model performances while thoroughly exploring how AI is transforming the field of photocatalytic wastewater treatment. Researchers have significantly enhanced their ability to understand, predict, and optimize photocatalytic degradation processes by using supervised machine learning techniques, including SVM, ANN, and tree-based algorithms. When trained on key operational parameters such as the type of photocatalyst, presence of dopants, light conditions, and types of contaminants, these AI models have shown exceptional effectiveness in capturing complex, nonlinear interactions that traditional methods often miss. AI has played a vital role in improving the performance of materials like TiO₂, ZnO, CdS, ZrO₂, WO₃, and CeO₂, along with their doped versions, by guiding experimental design and parameter selection. These materials have consistently demonstrated increased activity. Additionally, by integrating multiple studies into a cohesive synthesis, this review addresses a critical gap in the literature by highlighting both promising results and limitations related to model evaluation and data standardization. Nevertheless, there are still significant research gaps. Reliable model validation requires external datasets. Integration of AI with digital representations can facilitate scalable optimization and real-time monitoring. To increase trust and openness, explainable AI is required. When combined, these developments will position AI as a major facilitator of sustainable and intelligent photocatalytic wastewater treatment. Moving forward, prioritizing open data sharing and code accessibility is essential for ensuring transparency, reproducibility, and accelerated scientific progress in AI-driven photocatalytic research. Establishing benchmarking procedures, publishing reliable code, and providing datasets will enable the scientific community to validate, replicate, and continually improve models, thereby increasing confidence in reported results.

8. Future Perspectives

The analysis’s conclusions offer important information that may help determine the focus of future research, direct the creation of successful AI-driven solutions, and support more sustainable water management techniques. Future research should consider the following:

• As the current model algorithms are only designed for particular stages of the wastewater treatment process, more organized and comprehensive algorithms involving the prediction of pollutant levels, the enhancement of operating parameters, the monitoring of maintenance processes, and the reuse of recyclable polluting substances should be developed.
• Developing combination algorithms and combined optimization approaches that can better monitor and control wastewater treatment systems.
• Using artificial intelligence, which is currently understudied, to design novel, useful materials for wastewater treatment.
• Strong cross-disciplinary collaboration will be necessary for future developments in AI-driven photocatalysis to connect computational predictions with experimental validation successfully. For the results of models to be chemically interpretable, explainable AI must be implemented.
• At the same time, basic research needs to be prioritized to support both deeper mechanistic insights and real-world applications. More replicable and optimized photocatalytic systems will be made possible by striking a balance between theoretical knowledge and innovation. Ultimately, this integrated strategy has the potential to revolutionize sustainable chemical synthesis, wastewater treatment, and energy storage.