Recent advances in data-driven research have shown great potential in understanding the intricate relationships between materials and their performances. Herein, we introduce LLMB, an AI agent for lithium metal battery research that integrates a large language model (LLM) for hierarchical text mining with an automatic graph mining tool, Material Graph Digitizer (MatGD). This agent enables state-of-the-art accurate extraction of battery material data and cyclability performance metrics from diverse sources. Through text mining, we extracted composition and operating condition information for 15,398 battery cells, and graph mining yielded cyclability data for 10,242 cells. By aligning and merging these, we constructed a comprehensive database of 8,074 cells, containing component specifics and capacity. Utilizing the comprehensive database constructed through the LLMB agent, we developed the first machine learning model to predict capacities of LMBs using material information from battery components. Furthermore, molecular simulations and material analyses were performed to elucidate how the identified predictive features influence the physicochemical properties governing the battery performance. Based on these models and material analysis, we experimentally validated that the weakly solvating electrolyte, induced from low EState VSA6 solvents, facilitates the formation of an anion-derived solid-electrolyte interphase (SEI) and promotes highly crystallized Li plating, thereby confirming the reliability of our models.
Lee et al. (Tue,) studied this question.