Machine learning approach for photocatalysis: An experimentally validated case study of photocatalytic dye degradation.

In this study, machine learning (ML) models coupled with genetic algorithm (GA) and particle swarm optimization (PSO) were applied to predict the relative influence of experimental parameters of photocatalytic dye removal. Specifically, the impact of bandgap, dye concentration, photocatalyst dosage, solution volume, specific surface area, and time duration on photocatalytic degradation rate constant of cationic dyes was discerned using selected ML models, i.e., ensembled learning tree (ELT), gaussian process regression (GPR), support vector machine (SVM), and decision tree (DT). Thus, the data points were sourced from literature studies recently published in 2024 and 2023 on materials related to working on fundamental principles of photocatalysis. The ELT-PSO hybrid model outperformed all models with R = 0.992 and RMSE = 2.6408e, followed by DT, GPR, and SVM. The partial dependence plots and Shapley's analysis demonstrate that the type of dye, bandgap, dye initial concentration, and time duration are essential parameters for photocatalytic degradation, while sensitivity analysis further displayed solution volume and time duration to be the most influential parameters for rate constant determination. The optimized ML model's prediction was also experimentally validated using as-synthesized different compositions of CuO/WO heterostructures and ZnO nanoparticles. The results suggest that an ML-optimized study can be used in designing photocatalysts with optimum properties desired for the removal of cationic dyes at high rates from wastewater, thus saving energy and cost for a sustainable environment.