Li-P-S Electrolyte Materials as a Benchmark for Machine-Learned Interatomic Potentials

Abstract

With the growing availability of machine-learned interatomic potential (MLIP) models for materials simulations, there is an increasing demand for robust, automated, and chemically insightful benchmarking methodologies. In response, we here introduce LiPS-25, a curated benchmark dataset for a canonical series of solid-state electrolyte materials from the Li2S-P2S5 pseudo-binary compositional line, including crystalline and amorphous configurations. Together with the dataset, we present a suite of performance tests that range from conventional numerical error metrics to physically motivated evaluation tasks. With a focus on graph-based MLIP architectures, we run numerical experiments that assess (i) the effect of hyperparameters and (ii) the fine-tuning behavior of selected pre-trained ("foundational") MLIP models. Beyond the Li-P-S solid-state electrolytes, we expect that such benchmarks and their code implementations can be readily adapted to other material systems.

Abstract: arXiv descriptive metadata (CC0). Thank you to arXiv for use of its open access interoperability. This service was not reviewed or approved by, nor does it necessarily express or reflect the policies or opinions of, arXiv.