Machine Learning Potential Model Based on Ensemble Bispectrum Feature Selection and Its Applicability Analysis

With the continuous improvement of machine learning methods, building interatomic potential (MLP) based on datasets from quantum mechanics calculations has become an effective technical approach to improving accuracy classical molecular dynamics simulation. The Spectral Neighbor Analysis Potential (SNAP) is one most commonly used potentials. It uses bispectrum encode local environment each atom in lattice. hyperparameter jmax controls mapping complexity and precision between descriptor. As increases, description will more accurate, but number parameters descriptor increase dramatically, increasing computational complexity. In order reduce without losing accuracy, this paper proposes a two-level ensemble feature selection method (EFS) for descriptor, combining perturbation selector strategy. Based proposed method, subset selected original dataset dimension-reduced MLP. application validation, data Fe, Ni, Cu, Li, Mo, Si, Ge metal elements are train linear regression model SNAP predicting these metals’ atomic energies forces them evaluate performance subsets. experimental results show that, compared features qSNAP, training our EFS qSNAP than SNAP. Compared with existing when size 0.7 times that features, SSWRP strategy can achieve best terms stability, achieving average stability 0.94 across all datasets. reduced by about half, prediction 30%.