Machine Learning Enabled Design and Experimental Validation for Optimum CO2 Reduction to Methanol by MIL-53(Al) Based Heterojunction Photocatalysts

Abstract The rational design of photocatalysts for carbon dioxide (CO2) reduction remains challenging due to the complex interplay between electronic structure, surface properties, and reaction conditions. Data-driven approaches offer a promising route to identify key governing parameters; however, experimental validation of such predictions remains limited. In this study, machine learning-guided predictions for photocatalytic CO2 reduction to methanol are experimentally validated using MIL-53(Al) based composites. Prior analysis employing SHAP (shapley additive explanations) and partial dependence methods identified band gap, support characteristics, and light source as dominant factors influencing methanol yield, while the present experimental validation focuses on catalyst composition, band gap narrowing, and charge-transfer improvement under controlled UV irradiation. Guided by these insights, MIL-53(Al) is selected as the support and engineered into Ag2O and Bi2O3 codoped composites to probe the predicted structure activity relationships. Structural and compositional characterization using powder X-ray diffraction (XRD), scanning electron microscopy with energy dispersive spectroscopy (SEM-EDS) and high-resolution transmission electron microscopy (HRTEM) confirmed successful composite formation and homogeneous dopant distribution, while Fourier transform infrared spectroscopy verified the preservation of functional groups relevant for CO2 adsorption. Optical analysis revealed a substantial band gap reduction from 3.65 eV in pristine MIL-53(Al) to 2.62 eV in the codoped composite, consistent with machine learning predictions linking reduced band gap to enhanced photocatalytic performance. Brunauer–Emmett–Teller (BET) analysis further indicated increased surface area and porosity, supporting the identified importance of the support structure. Electrochemical measurements, including cyclic voltammetry, electrochemical impedance spectroscopy, Mott–Schottky and Bode analysis, demonstrate improved charge separation, enhanced carrier mobility and reduced interfacial resistance. These synergistic effects result in the highest methanol yield of 151.67 μmol g–1 h–1 for the optimized MIL-53(Al)/Ag2O/Bi2O3 composite under UV irradiation. Thus, this approach of machine learning guided experimental validation workflow provides a scalable strategy for sustainable CO2 conversion.