Abstract:
We present a comprehensive framework for the rational design of homogeneous catalysts, integrating crystal structure-derived X-ray absorption fine structure (XAFS) spectroscopy with machine learning to predict and validate reaction performance. Starting from the crystal structures (CIF files) of well-characterized Rh(III) catalysts, we systematically extract XAFS descriptors—including XANES edge features, white-line intensity, and EXAFS k- and R-space parameters. Machine learning models are subsequently trained on these metal-centric descriptors to predict reaction yield, regioselectivity, and stereoselectivity with remarkable accuracy. Furthermore, we evaluate and reveal which descriptor source (crystal structure-derived, DFT-derived, or DeepFit-extracted) provides the optimal training accuracy for selectivity prediction. By employing regression models, we elucidate the positive or negative impacts of specific structural features on catalytic performance, thereby establishing a strong causal relationship between catalyst structure and reactivity. Through the identification of the most influential dynamic structural features—particularly metal-ligand coordination changes that directly correlate with observed catalytic outcomes such as yield and selectivity—we successfully design novel catalysts predicted to exceed the performance of known systems. This approach not only demonstrates significant theoretical importance by deepening our understanding of structure-activity relationships but also offers immense practical utility in accelerating the discovery of high-performance catalysts.
Keywords – Descriptors, Machine learning, Crystal, DFT, Catalyst design, Asymmetric catalysis.