A sparse Bayesian Committee Machine potential for oxygen-containing organic compounds

Abstract

Accurate and scalable interatomic potentials are essential for understanding material properties at the atomic level; however, steep computational demands often limit their application. Although recent advances in machine learning (ML) potentials have been significant, extending kernel-based models to accommodate a broad range of chemical compositions remains a major challenge. Here, we present the active robust Bayesian Committee Machine (RBCM) potential, specifically designed to handle extensive datasets encompassing hydrocarbons (in gas, cluster, liquid, and solid phases) and eight families of oxygen-containing organic compounds. By employing a committee-based approach, the RBCM circumvents the poor scaling inherent to kernel regressors, facilitating straightforward and cost-effective model expansion. Systematic benchmarking demonstrates its robustness in accurately describing complex processes such as the Diels-Alder reaction, structural strain effects, and pi-pi interactions. These results highlight the RBCM's potential as a powerful tool for developing universal, ab initio-level ML potentials that offer both transferability and scalability across diverse chemical systems.