Research Data Supporting "Li-P-S Electrolyte Materials as a Benchmark for Machine-Learned Potentials"

This repository contains code and data for a manuscript to be submitted shortly.

Its purpose is to provide readers with access to the models and analysis code required to reproduce the results, as well as to make the dataset and related tasks open-access for benchmarking machine-learned interatomic potentials for the Li-P-S system.

---

Repository Structure

LiPS-25 Dataset

• The LiPS-25 benchmark dataset contains ~13,000 structures of solid-state electrolyte materials along the Li₂S–P₂S₅ pseudo-binary compositional line. It includes both crystalline and amorphous configurations, providing comprehensive coverage of the Li-P-S phase space.

• The full dataset can be found in data/lips-25.

• A filtered subset of Li₇P₃S₁₁ structures used for fine-tuning experiments is available in data/fine-tuning.

• All datasets come with predefined train, validation, and test splits.

Benchmark Tasks

• We provide a suite of four performance tests in the benchmarking folder, ranging from conventional numerical error metrics to physically motivated evaluation tasks.
• Each task includes Python notebooks or LAMMPS input scripts to facilitate reproducibility and benchmarking.

Models

• MACE hyperparameter sweep models: Provided in models/mace-models in both graph-pes format (x5 repeats) and LAMMPS-compatible versions.
• Fine-tuned foundation models: Available in models/fine-tuned for downstream tasks or further adaptation.