Data‐driven insights into photocatalytic dye degradation: Ensemble learning and <scp>SHAP</scp> ‐based mechanistic interpretation

Abstract BACKGROUND Photocatalytic dye degradation remains difficult to predict because performance depends on a complex interplay among catalyst composition, electronic structure, and operating conditions, while the available literature is often inconsistent in experimental design and reporting. In this study, a literature curated dataset containing photocatalytic degradation experiments from 305 studies was systematically analyzed using interpretable machine learning to reveal quantitative structure–activity relationships. The dataset integrates compositional descriptors including photocatalyst type, active metal, and support materials, with physicochemical and operational variables such as bandgap, surface area, catalyst dosage, reaction time, dye concentration, and solution volume. RESULTS Among the evaluated models, ensemble methods, particularly Random Forest, achieved the best predictive performance ( R 2 ≈ 0.63, MAE ≈ 0.01), indicating a strong ability to capture nonlinear dependencies within heterogeneous literature data. Feature importance analysis and SHapley Additive exPlanations consistently identified reaction time, photocatalyst composition, and active‐metal species as the dominant predictors of degradation efficiency. Bandgap energy and surface area showed secondary but mechanistically meaningful contributions. The model also reproduced key kinetic behaviors reported experimentally, including saturation trends with irradiation time and the nonlinear influence of bandgap on degradation performance. CONCLUSION Despite limitations arising from literature heterogeneity and missing experimental descriptors, the results demonstrate that interpretable machine learning can convert fragmented experimental data into predictive and mechanistically meaningful insights. This framework provides a practical bases for the data‐driven design and optimization of next‐generation photocatalysts, while also underscoring the importance of more standardized experimental reporting to improve model reliability, reproducibility, and transferability. © 2026 Society of Chemical Industry (SCI).