(346c) Exploring Metal-Support Interactions in Catalysis with Statistical Learning

We apply statistical learning (SL) techniques to identify descriptors that can construct models for predicting metal-support interactions in catalysis that results from charge transfer. Charge transfer between the metal and the support can be tuned by substituting parent metals in the oxide surface with aliovalent dopants or by introducing co-adsorbates on the oxide surface. As a test case, we use Al, B, Li, and Na as dopants and F, H, OH, and NO₂ as adsorbates to demonstrate control of charge transfer between transition metal atoms and modified MgO(100) surfaces. We collect fundamental chemical properties of the constituent elements in the system to generate a tremendous pool (~10⁶) of possible physical descriptors via feature engineering, which are then used to construct predictive models for computing metal binding energies on the surface. SL methods, e.g., LASSO, Horseshoe prior, and Dirichlet-Laplace prior, are then trained against DFT data to select descriptors that predict metal adatom binding energy on the MgO surface. We find that the Bayesian feature selection (FS) methods, i.e., Horseshoe prior and Dirichlet-Laplace prior, yield models that are generally more accurate than those identified with the LASSO method. Furthermore, we show that Dirichlet-Laplace prior can identify features that are transferable to other relevant oxides that are not included in the training process, such as CaO(100), BaO(100), and ZnO(100). This work demonstrates the superior capability of Bayesian FS methods for constructing general models for describing the metal-support interaction.