Data-driven Prediction of Ionic Conductivity in Solid-State Electrolytes with Machine Learning and Large Language Models

TL;DR

This paper demonstrates that combining geometric descriptors with machine learning and large language models enables efficient, accurate prediction of ionic conductivity in solid-state electrolytes, facilitating accelerated materials discovery.

Contribution

It introduces a dual approach using gradient-boosted trees with geometric descriptors and fine-tuned LLMs on CIF metadata for predicting ionic conductivity.

Findings

Gradient-boosted tree regressor achieved MAE of 0.543 in log(S cm-1).

SHAP analysis identified probe-occupiable volume and lattice parameters as key factors.

LLama-3.1-8B-Instruct achieved MAE of 0.657 using only CIF metadata.

Abstract

Solid-state electrolytes (SSEs) are attractive for next-generation lithium-ion batteries due to improved safety and stability but their low room-temperature ionic conductivity hinders practical application. Experimental synthesis and testing of new SSEs remain time-consuming and resource intensive. Machine learning (ML) offers an accelerated route for SSE discovery; however, composition-only models neglect structural factors important for ion transport while graph neural networks (GNNs) are challenged by the scarcity of structure-labeled conductivity data and the prevalence of crystallographic disorder in CIFs. Here, we train two complementary predictors on the same room-temperature, structure-labeled dataset (n = 499). A gradient-boosted tree regressor (GBR) combining stoichiometric and geometric descriptors achieves best performance (MAE = 0.543 in log(S cm-1)), and Shapley Additive exPlanations (SHAP) identifies probe-occupiable volume (POAV) and lattice parameters as key correlations for conductivity. In parallel, we fine-tune large language models (LLMs) using compact text prompts derived from CIF metadata (formula with optional symmetry and disorder tags), avoiding direct use of raw atomic coordinates. Notably, Llama-3.1-8B-Instruct achieves high accuracy (MAE = 0.657 in log(S cm-1)) using formula and symmetry information, eliminating the need for numerical feature extraction from CIF files. Together, these results show that global geometric descriptors improve tree-based predictions and enable interpretable structure-property analysis, while LLMs provide a competitive low-preprocessing alternative for rapid SSE screening.