Predicting and understanding photocatalytic CO2 reduction reaction with IR spectroscopy-based interpretable machine learning framework

Abstract

The highly selective conversion of carbon dioxide into value-added products is extremely valuable. However, even with the aid of in situ characterization techniques, it remains challenging to directly correlate extensive spectral data carrying microscopic information with macroscopic performance. Herein, we adopted advanced machine learning (ML) approaches to establish an accurate and interpretable relationship between vibrational spectral signals and catalytic performances to uncover hidden physical insights. Focusing on photocatalytic CO2 reduction, our model is shown to effectively and accurately predict the CO production activity and selectivity based solely on the infrared (IR) spectral signals, the generalizability of which is additionally demonstrated with a new Bi5O7I photocatalytic system. More importantly, further model analysis has revealed a novel strategy to steer CO selectivity, the physical sanity of which is verified by a detailed reaction mechanism analysis. This work demonstrates the tremendous potential of machine-learned spectroscopy to efficiently identify reaction control factors, which can further lay the foundation for targeted optimization and reverse design.