Data-driven prediction of ionic conductivity in solid-state electrolytes with machine learning and large language...

Area of Science:

• Materials Science
• Electrochemistry
• Computational Chemistry

Background:

• Solid-state electrolytes (SSEs) are crucial for developing safer and more stable next-generation lithium-ion batteries.
• Current experimental methods for discovering new SSEs are slow and costly.
• Existing machine learning models often overlook critical structural information or struggle with limited, disordered crystallographic data.

Purpose of the Study:

• To investigate the effectiveness of machine learning models in accelerating the discovery of solid-state electrolytes (SSEs).
• To compare the performance of traditional machine learning models with descriptors against large language models (LLMs) for predicting SSE ionic conductivity.
• To evaluate the impact of incorporating structural features into machine learning models for SSE property prediction.

Main Methods:

• Trained a gradient-boosted tree regressor model using stoichiometric and geometric descriptors on a dataset of 499 room-temperature, structure-labeled SSEs.
• Fine-tuned large language models (LLMs) using metadata from crystal structure files (CIFs), such as chemical formula and symmetry information.
• Utilized Shapley Additive exPlanations (SHAP) to interpret the feature importance in the tree-based model.

Main Results:

• The stoichiometric descriptor model achieved a test Mean Absolute Error (MAE) of 1.108 in log(S/cm).
• Incorporating geometric descriptors did not significantly reduce the MAE but provided complementary structural insights, highlighting the importance of features like density and L-values.
• Mistral-7B LLM achieved the lowest MAE (0.798), while Qwen3-8B demonstrated superior ranking performance (SRCC = 0.849) using only formula and disorder tags from CIF metadata.

Conclusions:

• Global geometric descriptors enhance the predictive power and interpretability of tree-based models for SSEs.
• Large language models offer a promising, low-preprocessing alternative for rapid screening of SSE materials, rivaling traditional methods.
• These machine learning approaches significantly accelerate the search for high-performance solid-state electrolytes for advanced battery applications.