(244d) High-Throughput Intermetallic Catalyst Evaluation for Selective Hydrogenation Reactions

Structure-property relationships between active sites and catalytic performance guides rational catalyst design. Intermetallic materials provide a unique opportunity for identifying these relationships due to their precise local atomic composition, exposing well-defined active site arrangements that can be perturbed with varying stoichiometries. Density functional theory (DFT) directly probes these perturbations and their influence on reaction energy barriers and adsorbate binding energies; however, the compositional space for intermetallics yields prohibitive requirements on computational resources for sampling possible structures. We present a multi-tiered computational workflow for the intelligent design of intermetallic catalysts for selective hydrogenation reactions.

Advancements in graph neural-networks (GNNs) for catalysis have yielded models which trained on a large DFT databases of mixed-metal surfaces adsorbed with various species. These models provide reasonable accuracy on adsorption energies considering the breadth of trained structures but fail to describe complex adsorbates within experimentally meaningful tolerances. We investigate fine-tuning to enhanced accuracy for single adsorbates. We evaluate the ability of these models, which are originally trained on stable structures, to evaluate transition state energy barriers. The training sets for these models are developed using an automated adsorbate and transition state search approach on low-index facets of stable catalysts obtained from widely utilized material databases.

We compare the use of GNNs to other high-throughput methods, including autoencoding projected densities-of-states (p-DOS) and feature selection using Iterative Bayesian additive regression trees (iBART). Whereas GNNs require only atomic configurations for predictions, there is interest to understand reactive phenomena through active site electronic structure. We leverage the GNN fine-tuning training set for p-DOS to determine the extent an “electronic” approach can predict adsorption energies in contrast to the GNN “geometric” approach. We demonstrate that p-DOS of an active site does not intrinsically encode sufficient information to be predictive on a practical, non-singly-faceted set of metallic surfaces with a complex adsorbate.