<p>The photocatalytic CO₂ conversion into value-added chemicals such as CH₄ and CO is a highly promising sustainable approach to meet rising energy demands while mitigating atmospheric CO₂ levels. This study investigates the prediction and optimization of CH₄ and CO production using tree-based machine learning (ML) models applied to g-C₃N₄/TiO₂ nanotube arrays (TNTAs) photocatalysts for gas-phase CO₂ conversion. To predict the photoconversion rates of CO₂ into CH₄ and CO, several tree-based ML algorithms, including AdaBoost, Bagging, CatBoost, Decision Tree, Extra Trees, Gradient Boosting, HistGradientBoosting, LightGBM, RandomForest, and XGBoost were employed. The production of CH₄ and CO (µmol/cm²) were designated as the target outputs, while input parameters included catalyst exposed surface area, initial concentration of CO₂, feed pressure, light power, and irradiation time. The performance of ML algorithms was appraised using five statistical metrics. Bayesian optimization was employed to fine-tune the hyperparameters of the machine learning model algorithms. Among the evaluated models, CatBoost (CB) performed the most accurately, with R² = 0.9887 (training) and R² = 0.9883 (test) for CH₄ production and R² = 0.9885 (training) and R² = 0.9874 (test) for CO production. Feature importance analysis and SHAP plots highlighted the significant influence of irradiation time and catalyst exposed surface area on the production efficiency of both products. Additionally, the input parameters were systematically optimized using CB model predictions, which were validated against experimental data and achieved almost similar prediction. These results support the effectiveness of ML-directed modeling in maximizing CO₂ conversion efficiency and revealing the potential of data-driven strategies in steering photocatalytic technology.</p>