Many-Body Interatomic Potential with Bayesian Active Learning, an Application to SiC

Machine learning interatomic potentials (MLIPs) have high efficiency and quantum accuracy to model atomic interactions and simulate atomic level processes. Active learning methods have been developed to train MLIPs efficiently. Among them, Bayesian active learning (BAL) utilizes uncertainty quantification as an acquisition threshold. In this work, we present a highly efficient BAL workflow, where MLIPs is constructed using Gaussian process (GP) kernels based on the atomic cluster expansion (ACE) descriptors which is trained efficiently with MPI parallelization. A high-performance mapping of the potential and an approximation of the uncertainty of the trained GP are developed. We demonstrate that our workflow is orders faster compared to BAL with exact GPs.<br/>As an application, we train a MLIP model for silicon carbide (SiC), a wide-gap semiconductor with diverse applications in power electronics, nuclear physics and astronomy. Particularly, the phase transition of SiC under high pressure is investigated, and is captured during active learning facilitated by the uncertainty prediction of the model. We demonstrate that the trained MLIP reaches excellent agreement with the density functional theory and outperforms the empirical potentials in the prediction of elastic and thermal properties of pristine bulks, as well as the enthalpy under different pressures ranging from 0-150 GPa. The highly efficient active learning workflow can be easily extended to other systems, accelerate material discovery and facilitate the development of quantum technologies.