General reactive element-based machine learning potentials for heterogeneous catalysis

Abstract

Developing truly universal machine learning potentials for heterogeneous catalysis remains challenging. Here we introduce our element-based machine learning potential (EMLP), trained on a unique random exploration via imaginary chemicals optimization (REICO) sampling strategy. REICO samples diverse local atomic environments to build a representative dataset of atomic interactions, making the EMLP inherently general and reactive, capable of accurately predicting elementary reactions without explicit structural or reaction pathway inputs. We demonstrate the generality and reactivity of our approach by building a Ag-Pd-C-H-O EMLP targeting Pd–Ag catalysts interacting with C/H/O-containing species, achieving quantitative agreement with density functional theory even for complex scenarios such as surface reconstruction, coverage effects and solvent environments, cases for which existing foundation models typically fail. Our method paves the way to replace density functional theory calculations for large and intricate systems in heterogeneous catalysis, and offers a general framework that can readily be extended to other catalytic systems, and to broader fields such as materials science. It is challenging to design machine learning potentials for heterogeneous catalysis that are universal, reactive and have high accuracy. Now, an element-based machine learning potential relying on a random exploration via an imaginary chemicals optimization sampling strategy is put forward, and is successfully demonstrated for a range of applications.