Active learning accelerated exploration of single-atom local environments in multimetallic systems for oxygen electrocatalysis

Hoje Chun,Jaclyn Lunger,Jeung Ku Kang,Rafael Gómez-Bombarelli,Byungchan Han

DOI: https://doi.org/10.26434/chemrxiv-2024-6skxn

2024-03-14

Abstract:Single-atom catalysts (SACs) exhibit high activity for a wide range of sluggish reactions and allow performance tunability at atomic-level through the selection of central metals, ligand environments, and secondary metal sites. However, the design space with varying structures and compositions significantly hinders the fast and accurate identification of desired multimetallic SACs. In this work, we demonstrate a self-driving computational strategy for exploring binary metallic sites with varying combinations of 3d transition metals and different ligand environments, resulting in over 30,000 single atom sites for the electrochemical catalysis of the oxygen reduction and evolution reactions (ORR/OER). This approach is based on the density functional theory (DFT) calculations of binding energies and atomic descriptors as target properties and utilizes an equivariant graph neural network (GNN) as a surrogate model for predicting DFT labels directly from the atomic structure. The chemical environments learned by the GNN lead to capturing composition-structure-property relationships for the ORR/OER activity and selectivity. Active learning facilitates the investigation of the search space by balancing the exploration of unseen atomic structures with the exploitation of the active ones. The computational predictions of promising Co-Fe, Co-Co, and Co-Zn metal pairs are consistent with the state-of-the-art results of experimental measurements reported in the literature. The GNN-based analysis of multiple active sites on a single surface for target catalytic reaction can be extended to a broader class of multi-element high entropic materials systems.

Chemistry

What problem does this paper attempt to address?

This paper focuses on accelerating the search for efficient single-atom catalysts (SACs) in multi-metal systems, particularly for oxygen electrocatalysis. The research team proposes a self-driving computational strategy based on active learning, which effectively explores single-atom sites in different metal combinations and ligand environments to optimize SACs for oxygen reduction reactions (ORR) and oxygen evolution reactions (OER).

They developed an equivariant graph neural network (GNN) as a surrogate model, which directly predicts density functional theory (DFT) labels from input ideal structures to identify active and selective SACs. By computationally-guided local environment tuning, different single-atom metals and ligands were selected. The paper also emphasizes the importance of multiple active sites on a single surface for target catalytic reactions and demonstrates the scalability of this approach to a wider range of multicomponent high-entropy materials systems.

The active learning strategy balances exploration of unseen atomic structures with utilization of known active structures, enabling efficient search of a large materials design space with various configurations and compositions. The predicted results for Co-Fe, Co-Co, and Co-Zn metal pairs mentioned in the paper are consistent with experimental measurements, indicating high activity and stability.

In summary, this work provides a novel computational approach to rapidly discover efficient and tunable single-atom catalysts for ORR/OER using machine learning and active learning techniques, which is of significant importance for energy conversion and storage reactions.