Reliable Li-ion conductivity with efficient data generation and uncertainty estimation toward large-scale screening

Abstract

High ionic conductivity is a crucial property required in solid electrolytes (SEs) for all-solid-state batteries with high rate capability. However, the high computational cost of conventional ab initio molecular dynamics (AIMD) simulations necessitates the use of small cells, short simulation times, and high temperatures, often leading to the overestimation of the ionic conductivity. Here, we exploited a universal machine learning interatomic potential to accurately calculate ionic conductivities and activation energies of Li-ion conductors at a much lower cost. By employing the pretrained M3GNet, we developed a novel workflow that eliminated the need for AIMD to generate datasets for finetuning. Through uncertainty estimation, we showed that the finetuned M3GNet exhibited excellent reliability in MD simulations even without active learning, accurately predicting the ionic conductivities and activation energies. Our approach was successfully incorporated into the screening pipeline, which led to the identification of eight promising solid electrolyte candidates out of 4,285 materials. These new candidates, along with high synthesizability, meet the key properties required for SEs. The proposed workflow is anticipated to expand the application of universal machine learning interatomic potential in materials discovery.