(106g) Identifying Interaction Trends between Single Metal Atoms and Modified MgO(100) Supports with Statistical Learning

In this work, we apply statistical learning (SL) to identify physical descriptors for predicting electronic metal-support interactions (EMSI) in catalysis. Charge transfer between metal atoms and their underlying oxide support influences both the binding strength of the metal to the support and the binding strength of reaction intermediates to the metal. Tuning charge transfer by introducing surface-modifying dopants or co-adsorbates is an important strategy for controlling catalytic behavior. We use SL to build models for predicting metal atom binding energy to the oxide surface as a function of readily available physical descriptors containing properties of the metal, the support, and the surface-modifiers. Models are trained against DFT datasets consisting of metal adsorption energies on modified MgO(100) surfaces featuring various aliovalent dopants (e.g., Al, B, Li, and Na) and adsorbates (e.g., F, H, OH, and NO₂). These modifications generate both electron-rich and electron-poor MgO surfaces, which in turn enhance the extent of charge transfer between the metal and the support. We also test multiple SL procedures, including LASSO, Horseshoe prior, and Dirichlet-Laplace prior, and find that features derived by Dirichlet-Laplace prior construct the most accurate predictive models (i.e., achieving RMSE of ~0.2 eV for predicted binding energies). Moreover, we find that the features selected by SL using MgO training data generally are transferable to similar oxides, such as CaO(100), BaO(100), and ZnO(100), where data for these oxides are excluded from the feature selection process. This demonstrates the robust nature of features selected by the Dirichlet-Laplace prior procedure, both in terms of model accuracy and transferability to oxide surfaces beyond those included in the training set.