DOI: https://doi.org/10.29363/nanoge.interect.2022.012
Publication date: 11th October 2022
Multi-Scale modeling has been the standard to explain and predict activity and selectivity of chemical reactions during the last 20 years[1,2]. This procedure allowed us to predict semiquantitatively the activity trends, recovering the classical volcano plots. The later models lead to successful catalyst optimization using a reduced set of energy descriptors. However, the accuracy of Multi-Scale modeling confronts several limitations with complexity, mainly caused by the coverage effects, the catalyst phase or surface reconstructions, large reaction networks, and highly dynamic materials[1-3]. Statistical Learning (SL) techniques can overcome such limitations. Nevertheless, the black-box nature of most of the SL techniques hinders the physical interpretation of the results. In this work, we present a procedure to generate physical interpretable models able to correlate experimental activity and selectivity with ab-initio Density Functional Theory (DFT)-based descriptors. Here, we applied our methodology on the CH2X2 (X=Cl, Br) hydrodehalogenation reaction family catalyzed by transition metals[3]. Even if this study is based on thermochemical systems, it provides a starting point to solve more complex chemical problems, such as explaining the dynamic charge exchange of single metal atoms on Ceria[4] or electrochemistry.