Distinguishing Among High Activity Electrocatalysts: Regression vs Classification

Abstract

Computational design of high-functioning electrocatalysts is an integral part of catalyst discovery. However, as we approach the limits of performance imposed by scaling relations, finding active catalysts using computational methods needs to account for uncertainty within these methods. In particular, density functional theory calculations followed by thermodynamic analysis have been the main tool to rationalize trends in activity and selectivity of electrochemical reactions. In this work, we analyze this problem using three important reactions—hydrogen evolution, oxygen reduction, and oxygen evolution reactions. Using error estimation capabilities within the Bayesian error estimation functional (BEEF), we find based on distribution of adsorption energies for intermediates there is significant uncertainty in determining whether a material belongs to the strong binding or weak binding leg of an activity volcano. This results in bimodal (or multimodal) distributions for the predicted activity of high-performing catalyst materials. For example, there is only 75% agreement as sampled through GGA-level with BEEF that Pt3Ni(111) yields higher ORR activity than Pt(111). To circumvent this issue, we approach this delineation problem as a classification task from both legs of the activity volcano. For each reaction, we identify a pair of reference catalysts on opposite legs of the volcano such that they have a similar predicted activity. By calculating an interval confidence value for a catalyst of interest to lie in the descriptor energy interval set by reference catalysts, we can assess whether a material exhibits higher activity than both references as suggested by activity volcano trends. For oxygen reduction, reference catalysts Pt(111) and Au(100) are well-established to lie on the strong and weak binding legs, respectively, and predict OH* adsorption of Pt3Ni(111) bound by both with 100% confidence. With similar reasoning, we choose several reference pairs for hydrogen evolution and oxygen evolution. By evaluating the best trade-off between the number of catalysts to distinguish and accuracy, we suggest that this method can be used universally to distinguish catalysts more robustly near the top of an activity volcano. This becomes even more important for assessing activity of multielectron electrochemical reactions.