Efficient Sampling of Structural Configurations with A Universal Materials Graph Neutral Network Potential

Training structure is one key component for fitting robust machine learning interatomic potentials (ML-IAPs). It is usually sampled by short trail runs of <i>ab initio</i> molecular dynamics together with manual selection by intervals. This traditional sampling method suffers from high computational cost and low diversity. The as-trained ML-IAPs are not reliable for severe temperatures at or above melting temperatures, thus prohibiting its accurate simulation for amorphous structures and calcination conditions. On the other hand, active learning (AL) strategy has been proposed and verified to effectively sample distinctive configurations in extreme and target scenarios, thus improving the reliability of ML-IAPs. However, AL still requires a good initial training set and intrinsically have to conduct <i>ab initio</i> calculations as well as ML-IAP fitting iteratively. In our study, we propose a method to broadly sample training configurations with a universal graph neutral network potential. Our strategy can achieve highly reliable ML-IAPs with one single iteration of optimization, and it is generally applicable to 89 elements in the period table.