A Sparse Bayesian Committee Machine Potential for Hydrocarbons

Abstract: Accurate and scalable universal interatomic potentials are key for understanding material properties at the atomic level, a task often hindered by the steep computational scaling. Although recent developments of machine learning potential has made significant progress, the flexibility and expansion to a wide range of compounds within a single model seems still challenging to build in particular for the kernel-based models. Here, we introduce the Bayesian Committee Machine (BCM) potential to handle a wide array of hydrocarbons in various phases (gas, clusters, liquid, and solid). The BCM potential leverages the committee approach to bypass the poor scaling of kernel regressors when dealing with large datasets. Its committee-based structure allows for easy and cost-effective expansion, maintaining both transferrability and scalability. Demonstrating its robustness, the BCM accurately models challenging examples like the Diels-Alder reaction, structural strains, and pi-pi interactions. Our systematic benchmarking positions the sparse BCM potential as a promising candidate in the quest for a universal ab initio machine learning potential.