Research on solid-state battery information mining based on large language models and chain-of-thought reasoning

Abstract

Solid-state batteries (SSBs) represent a next-generation energy storage technology. However, the wide diversity of solid-state electrolyte (SSE) chemistries, coupled with the complexity of their synthesis and processing routes, makes systematic investigation, knowledge reuse, and database construction particularly challenging. In this work, we propose SSBExtractor, a novel information extraction framework designed to mine synthesis protocols and performance metrics from scientific literature using large language models (LLMs) enhanced with chain-of-thought (CoT) reasoning. Unlike conventional rule-based or direct prompting methods, SSBExtractor decomposes the extraction task into four structured stages, localization, extraction, matching, and formatting, with each stage utilizing a corresponding dedicated prompt. We further introduce a multi-agent interactive prompt refinement approach to improve extraction accuracy through iterative feedback from labeled data and domain experts. Applied to a dataset of 800 SSB-related papers, SSBExtractor successfully established the largest SSE synthesis-process database to date, covering 851 electrolyte materials and 15,337 process parameters. Relying on CoT-based framework rather than resource-intensive fine-tuning, the proposed framework achieved over 98.5% precision and recall, significantly outperforming standard prompting methods. Notably, the CoT-based framework proved particularly effective in handling complex cases involving multiple materials and synthesis routes within a single article, where traditional methods often failed. This work demonstrates the potential of CoT reasoning in LLMs for accurate, scalable scientific information extraction and offers a foundational methodology with prospective applicability for knowledge mining across broader materials science domains.