Benchmarking Machine Learning Models for Predicting Lithium Ion Migration

Abstract

The development of fast ionic conductors to improve the performance of electrochemical devices relies on expensive high-throughput (HT) density functional theory (DFT) calculations of transport properties. Machine learning (ML) can accelerate HT workflows but requires high-quality data to ensure accurate predictions from trained models. In this study, we introduce the LiTraj dataset, which comprises 13,000 percolation and 122,000 migration barriers, and 1700 migration trajectories, calculated for Li-ion in diverse crystal structures using empirical force fields and DFT, respectively. With LiTraj, we demonstrate that classical ML models and graph neural networks (GNNs) for structure-to-property prediction of percolation and migration barriers can distinguish between “fast” and “poor” ionic conductors. Furthermore, we evaluate the capability of GNN-based universal ML interatomic potentials (uMLIPs) to identify optimal Li-ion migration trajectories. Fine-tuned uMLIPs achieve near-DFT accuracy in predicting migration barriers, significantly accelerating HT screenings of new ionic conductors.