Fine-Tuned Global Neural Network Potentials for Global Potential Energy Surface Exploration at High Accuracy.

Abstract

Machine learning potential (MLP), by learning global potential energy surfaces (PES), has demonstrated its great value in finding unknown structures and reactions via global PES exploration. Due to the diversity and complexity of the global PES data set, an outstanding challenge emerges in achieving PES high accuracy (e.g., error <1 meV/atom), which is essential to determine the thermodynamics and kinetics properties. Here, we develop a lightweight fine-tuning MLP architecture, namely, AtomFT, that can explore PES globally and simultaneously describe the PES of a target system accurately. The AtomFT potential takes the pretrained many-body function corrected global neural network (MBNN) potential as the basis potential, exploits and iteratively updates the atomic features from the pretrained MBNN model, and finally generates the fine-tuning energy contribution. By implementing the AtomFT architecture on the commonly available CPU platform, we show the high efficiency of AtomFT potential in both training and inference and demonstrate the high performance in challenging PES problems, including the oxides with low defect content, molecular reactions, and molecular crystals─in all systems, the AtomFT potentials enhance significantly the PES prediction accuracy to 1 meV/atom.