Unlocking the potential of CO2 photoreduction: A hybrid experimental and machine learning approach

Controlled light-driven photocatalytic CO 2 reduction is one of the highly promising methods for the transformation of CO 2 into value-added products. The catalyst plays a pivotal role in photocatalytic CO 2 reduction by enabling efficient light irradiation, enhancing charge-carrier separation, and providing active sites for CO 2 adsorption and activation. Advanced catalysts offer superior performance due to their tailored structural and electronic properties, making them essential for achieving high conversion rates and producing value-added products like CH 4 and CO. In this work, ten tree-based machine learning (ML) regression models were systematically used to predict CO 2 photoreduction efficiency by leveraging both catalyst characteristics and experimental operational parameters. The models were further critically evaluated using multiple statistical metrics such as Coefficient of Determination (R²), Mean Square Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Median Absolute Error (MedAE) for training and testing datasets. The scatter plots show closer agreement between predicted and experimental values, and the radar plots allow comparison of model performance metrics. Among all the tested models, CatBoost (CB) model seems to perform better in terms of predictive accuracy and generalization capability, capturing complex nonlinear relationships and feature interactions in the data. To explain the model, normalized feature importance plots and SHAP analyses were performed, and both indicate that the catalyst's composition, irradiation time, and catalyst-exposed surface area are the three major driving forces of the photoreduction processes. Further experimental validation confirms that the g-C 3 N 4 /TNTA composite catalyst exhibits much higher efficiency than pure TNTAs and g-C 3 N 4 . Based on the CatBoost model, the optimal conditions to maximize the photoreduction of CO 2 are: catalyst surface area 16 cm 2 , CO 2 concentration 33 %, feed pressure 400 mbar, light power 200 W, and irradiation time 3.5 h. Under these conditions, the photoreduction efficiency is 76.64%, which is closer to the experimentally observed values. The combination of interpretable machine learning and experimental perspectives, therefore, can be seen as a scalable and effective strategy for advancing CO 2 photoreduction technologies and providing further guidance for academic research on sustainable photocatalytic systems. • ML models are trained to forecast CO 2 photoreduction on TNTAs and g-C 3 N 4 /TNTAs catalyst. • CatBoost model achieved the best predictive accuracy for the catalyst system. • SHAP analysis revealed catalyst types and irradiation time as the key controlling parameters. • Bayesian optimization identified operating conditions that maximize CO 2 photoreduction.