Machine Learned Force Field for Simulating Structural Phase Transformation under High Pressure

Hongxiang Zong^{1, 2}*，Long Zhao¹, Graeme Ackland², Xiangdong Ding¹

¹ Xi’an Jiaotong University, Xi’an,710049, China；

² Univeristy of Edinburgh, Edinburgh, EH9 3FD, UK

ABSTRACT: Atomic simulations provide an effective means to understand the underlying physics of martensitic transformations under extreme conditions. However, this is still a challenge for certain allotropic materials due to the lack of an accurate classical force field. Based on machine learning (ML) techniques and domain knowledge of structural phase transformations, we develop a hybrid ML-AIMD method in which interatomic potentials describing phase transformations can be learned with a high degree of fidelity from a sampling of ab initio molecular dynamics simulations (AIMD). Using Zirconium, Potassium and solid hydrogen as model systems, we demonstrate the feasibility and effectiveness of our approach. Specifically, the ML interatomic potential correctly captures energetic and structural properties as verified by comparison to experimental and DFT data. Molecular-dynamics simulations successfully map out their pressure-temperature phase diagram. Furthermore, the strategy is promising for finding novel phenomena of materials under extreme compression.

Figure 1. Scheme of machine learning based atomic simulation of phase transformation behaviors.

Keywords: machine learning; multi-scale simulation; phase transformation; high pressure