(662f) Improving the Accuracy of ML-Models for Catalysis through Bulk Electronic Structure Descriptors

The discovery of high performing catalysts for the oxygen evolution reaction (OER) and oxygen reduction reaction (ORR) is of high importance for advancing clean hydrogen technologies. Current state-of-the-art catalysts contains precious metals, such as iridium, rhodium, and platinum, and affordable and abundant alternatives with sufficient stability and catalytic activity are needed for large scale implementation. For the OER and ORR, the O* and OH* surface adsorption energies obtained from density functional theory (DFT) have been demonstrated to be key descriptors for activity. However, the computational cost and effort of surface calculations is a challenge for the efficient prediction of catalytic performance across a larger material space. Therefore, surrogate and machine learning models can enable a faster exploration for materials of larger complexity.

In this talk I will show how electronic and structural descriptors derived from bulk DFT calculations can improve ML models for OER/ORR catalysis. First, I will discuss the application to O and OH adsorption energy predictions, demonstrated for a new, consistent dataset of adsorption on unary (AxOy) oxides spanning the entire transition metal series. Building on our recent work [1], we extend our ICOHP-based model to a ML-based prediction of adsorption energies across multiple oxidation states with a MAE of ~0.2 eV for the O-OH and OH descriptors. Next, I will show the application to experimental electrocatalysis targets where ML models are enhanced by integrating experimental parameters with DFT-based bulk descriptors. Last, I will discuss computational and experimental datasets availability via catalysis-hub.org.

These results can enable a more efficient screening of materials active for OER/ORR on the bulk level of DFT computation that can significantly reduce the computational cost and improve the generalizability of ML models.

[1] Comer et al. JPCC 2022 126 (18)