Data Efficient and Stability Indicated Sampling for Developing Reactive Machine Learning Potential to Achieve Ultra-long Simulation in Lithium Metal Batteries

Modelling the formation of solid-liquid interphase (SEI) is challenging as its strict requirement with both simulation accuracy and length. Machine learning potential (MLP) based molecular dynamics (MD) simulation is expected to play a role in this field while currently its use is hindered by sampling efficiency and simulation stability. In this work, we tackle the two challenges together. We propose the stability-indicatedsampling (SIS) algorithm for efficiently sampling training data using physical information (temperature). Unlike previous strategies, our method does not need prior knowledge of reaction networks or training multiple MLPs for uncertainty estimation. Compared with the recent proposed methods HAIR and DP-GEN, our approach gives significant improvement of sampling efficiency with less requirements with the initial training data, to realize > 10 ns MLPMD simulation using ab initio MD (AIMD) trajectory of just a few ps. We introduce the concept underlying instability consistency by showing the accuracy of reaction mechanisms and radial distribution function (RDF) can be improved by SIS-MLPMD, although their information is not explicitly used in our sampling decision. Furthermore, we show that long-time MLPMD simulation of Lithium metal battery (LMB) can not only reproduce some well-known SEI components including LiF, Li2O, LiOH, LiS and the incomplete N-S breaking in highconcentration systems, but also ionic aggregation structures of LiF, which is not shown in our AIMD training data but matches previous results of electrochemical impedance spectroscopy. Our work is expected to help accelerate future investigations, especially for studying long-time (≥ ns scale) reaction dynamics in interfacial problems.