Machine Learning-Based Prediction of Photocatalytic Degradation of Organic Dyes Under Different Artificial Light Sources

Organic dyes used in industrial processes, particularly in textile, leather, paper, and cosmetic applications, are among the most persistent and environmentally harmful pollutants due to their stable aromatic structures and resistance to natural degradation. Methylene blue (MB) and methyl orange (MO) are representative model dyes commonly used to investigate photocatalytic degradation mechanisms in aqueous systems. In this study, the photocatalytic degradation of MB and MO was systematically investigated under two artificial light sources (LED and mercury-vapor lamp) in the presence of NiO nanoparticles. A physics-consistent, data-driven modeling framework was developed to predict degradation kinetics using the logarithm of absorbance ratios (ln(At/A0). Multiple machine learning models, including Decision Tree, LSBoost, Support Vector Regression, Gaussian Process Regression, Ridge Regression, and a shallow neural network, were evaluated using a Leave-One-Group-Out (LOGO) cross-validation strategy to assess generalization across heterogeneous dye–light regimes. Model outputs were subsequently transformed back into physically measurable domains (At and degradation efficiency, D%) to ensure interpretability. Among the evaluated approaches, the Decision Tree model demonstrated superior predictive performance, stable residual behavior, and kinetic consistency. The results highlight the importance of integrating physical insight with data-driven modeling to reliably describe complex photocatalytic degradation processes.