(45g) Improved Catalyst Predictions with Machine Learning Coupled Alchemical Perturbation Density Functional Theory

First principles Kohn-Sham density functional theory (KS-DFT) calculations hinder large-scale searches for new and sustainable heterogeneous catalysts due to significant computational expense. Alchemical perturbation density functional theory (APDFT) is a promising method that provides rapid predictions of catalyst descriptors based on perturbations of electrostatic potentials that arise from isoelectronic transmutations made to a reference catalyst. While straightforward and simple, first-order APDFT approximations are largely inaccurate when large and/or many transmutations are made. Here, we first discuss systematic trends in errors related to these factors for binding energy (BE) prediction of CH_x, NH_x, and OH_x adsorbates on hypothetical alloy variations of fcc Pt(111) surfaces. Based on these observations, we demonstrate our approach fingerprinting locations of transmutations in these hypothetical alloys, and we show how to correct APDFT errors with machine learning models trained using about 3600 BE data points. Our models provide improved prediction accuracy with errors decreased by as much as an order of magnitude, and this approach extends the breadth of APDFT predictions, allowing rapid and accurate BE predictions on many more alloys made by numerous transmutation combinations.