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

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Machine-learned interatomic potentials (MLIPs) are increasingly used in materials simulations.
• Existing benchmarking methods for MLIPs lack robustness, automation, and chemical insight.
• Standardized evaluation is crucial for advancing MLIP development and application.

Purpose of the Study:

• To introduce LiPS-25, a curated benchmark dataset for Li-P-S solid-state electrolytes.
• To present a comprehensive suite of performance tests for MLIPs.
• To facilitate systematic numerical experiments for assessing MLIP performance and fine-tuning.

Main Methods:

• Development of the LiPS-25 dataset, including crystalline and amorphous configurations.
• Design of performance tests encompassing numerical error metrics and physically motivated tasks.
• Application of the dataset and tests to graph-based MLIP architectures for numerical experiments.

Main Results:

• Demonstration of systematic assessment of hyperparameter effects on MLIP task-level performance.
• Analysis of fine-tuning behavior for pretrained MLIP models using the benchmark.
• Validation of the benchmark's utility for evaluating MLIPs in Li-P-S electrolytes.

Conclusions:

• The LiPS-25 benchmark provides a robust framework for evaluating MLIPs in solid-state electrolytes.
• The methodology and code are adaptable to other material systems, promoting broader MLIP development.
• This work advances automated, chemically informed benchmarking for materials simulations.