Learning from models: high-dimensional analyses on the performance of machine learning interatomic potentials

Abstract

Abstract Machine learning interatomic potential (MLIP) has been widely adopted for atomistic simulations. While errors and discrepancies for MLIPs have been reported, a comprehensive examination of the MLIPs’ performance over a broad spectrum of material properties has been lacking. This study introduces an analysis process comprising model sampling, benchmarking, error evaluations, and multi-dimensional statistical analyses on an ensemble of MLIPs for prediction errors over a diverse range of properties. By carrying out this analysis on 2300 MLIP models based on six different MLIP types, several properties that pose challenges for the MLIPs to achieve small errors are identified. The Pareto front analyses on two or more properties reveal the trade-offs in different properties of MLIPs, underscoring the difficulties of achieving low errors for a large number of properties simultaneously. Furthermore, we propose correlation graph analyses to characterize the error performances of MLIPs and to select the representative properties for predicting other property errors. This analysis process on a large dataset of MLIP models sheds light on the underlying complexities of MLIP performance, offering crucial guidance for the future development of MLIPs with improved predictive accuracy across an array of material properties.