Infrared Spectra‐Based Machine Learning Framework for Photocatalytic Reaction and Performance

In experimental catalysis research, conventional trial‐and‐error approach still dominates during the exploration of the intricate relationship between microscopic mechanisms and macroscopic performance, which unfortunately is time‐consuming and resource‐intensive, placing additional hurdles on catalyst high‐throughput screening. Alternatively, machine learning (ML) methods offer new possibilities to largely accelerate the entire process due to their capability of extracting hidden patterns. Herein, we exploit machine‐learnt vibrational spectroscopy to establish a direct, predictable, interpretable, and transferable relationship between infrared (IR) signals of adsorption with catalytic performance. Taking photocatalytic NO oxidation reaction as a typical case, the framework is shown to effectively and accurately predict the nitrate formation based solely on the IR spectral signals of NO adsorption, the generalizability of which is additionally demonstrated with a new CaCO 3 ‐decorated g‐C 3 N 4 system. The physical sanity of this model can be rationalized by the extracted insights well–aligned with mechanistic understandings, and its transferability is illustrated with a further prediction on NO removal activity. The ML‐based approach reduces experiment time by 3.5 times and enables quantitative determination of catalytic performances, thereby expanding the applicability of traditional spectroscopic techniques for catalyst screening and prediction.