Quantum-accurate Force Fields from Machine Learning of Large Materials Data

Work with Zhi Deng, Yunxing Zuo, Shyue Ping Ong

Machine learning of the quantitative relationship between local environment descriptors and the potential energy surface of a system of atoms has emerged as a new frontier in the development of interatomic potentials. We generalize the spectral neighbor analysis potential (SNAP) model to binary metal alloy, binary ionic, and high entropy alloy systems. We demonstrate that machine-learned SNAP models can achieve near-DFT accuracy in the prediction of many key properties, such as elastic constants, formation energies, melting points, phase diagram, etc. Long-time large-scale simulations are performed to provide atomistic insights into many physical phenomena, for example grain boundary diffusion, dislocation motion, among others. We also present a comprehensive evaluation of different local environment descriptors using a diverse elemental dataset comprising bcc and fcc metals and diamond group IV semiconductors. We show that all descriptors have excellent performance in predicting energies and forces, as well as a wide spectrum of properties. A general trade-off between accuracy and the computational cost of each model is discussed in the context of model selection for molecular dynamics and other applications.