Machine learning approach for MOF-based photocatalytic CO2 reduction to fuels

Atmospheric CO 2 concentrations now exceeding 420 ppm are accelerating global warming, prompting urgent interest in solar-driven CO 2 conversion as a carbon–neutral fuel production strategy. Among emerging photocatalyst platforms, Metal-Organic Frameworks (MOFs) offer exceptional tunability through their designable organic linkers, metal ions and porous structure. Yet the complex, nonlinear relationships between their structural properties and catalytic performance remain poorly understood. Herein, we present a machine learning framework trained on 417 experimental data points compiled from the MOF-based photocatalytic CO 2 reduction literature. The dataset was randomly split into training (80%) and testing (20%) to predict and interpret the formation yields of four target products: CH 3 OH, HCOOH, CO, and CH 4 . Six tree-based ensemble models namely CatBoost, XGBoost, LightGBM, GBDT, Random Forest, and AdaBoost, were systematically compared across six performance metrics. SMOGN Augmentation was applied only to the training set. CatBoost model achieved a training R 2 of 0.988 with an RMSE of 50.21, while maintaining strong generalization on the independent test set with a test R 2 of 0.966 and an RMSE of 83.59. Furthermore, validation using an independent dataset also confirmed the excellent performance of the CatBoost model, yielding an R 2 of 0.996 with an RMSE of 30.15 for the training set and an R 2 of 0.976 with an RMSE of 70.07 for the test set. CatBoost was selected for downstream interpretation using SHAP (SHapley Additive exPlanations) analysis. SHAP enabled the quantitative identification of the catalyst properties and reaction conditions that are strongly associated with product formation. These findings establish an interpretable, data-driven roadmap for the rational design of MOF photocatalysts, enabling targeted optimization of both activity and product selectivity for sustainable solar fuel production.