Statistical learning (SL) has been introduced in material science for various applications, such as predicting various material properties and screening large composition spaces for material candidates with certain desirable properties. In this work, we develop new feature selection (FS) strategies for predicting the extent of metal-support interactions in heterogeneous catalysis. Metal-support interaction, quantified here as the binding energy of a metal on an oxide support, is dominated by charge transfer and is extremely sensitive to the composition of the support. Given the large number of physical influencers that control charge transfer, understanding metal-support interactions is an ideal test bed for developing FS methods. Our training data consists of metal binding energies on oxide supports that are computed with density functional theory (DFT). We populate a feature space of candidate descriptors built from fundamental properties of supported metals, oxide supports, support dopants, and surface adsorbates that are readily available in the literature. These features are then engineered to generate a feature space pool containing over one million candidate features. Using the FS methods, such as least absolute shrinkage and selection operator (LASSO), Horseshoe prior, and Dirichlet-Laplace prior, we can identify the correlated features with metal binding energy and construct the predictive model. We find that Bayesian methods, such as Horseshoe prior and Dirichlet-Laplace prior, can derive more accurate models than LASSO. This behavior is further investigated with statistical simulations, which show that the performance of LASSO is limited when handling a highly correlated feature space leading to a high rate of false positives. Horseshoe prior displays improved characteristics when handing the correlated feature space, but is too aggressive in feature selection leading to a high rate of false negatives. Dirichlet-Laplace prior balances these tendencies, and therefore derives the most accurate models. Not only are the models accurate, we also show that descriptors built using Dirichlet-Laplace prior feature selection are transferable to oxide supports that were not contained in the original training and feature selection procedures. This demonstrates the robust nature of the selected features for describing the underlying physics.

Developing statistical learning tools for understanding metal-support interactions in catalysis

ACS Spring Meeting 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.