Evaluation of the application of large language models in the entire process of battery research and development of a comprehensive database forinorganic solid electrolyte

The emergence of large language models has significantly advanced scientific research. Representative models such as ChatGPT and DeepSeek R1 have brought notable changes to the paradigm of scientific research. While these models are general-purpose, they have demonstrated strong generalization capabilities in the field of batteries, especially in solid-state battery research. In this study, we systematically screen 5309268 articles from key journals up to 2024, and accurately extract 124021 papers related to batteries. Additionally, we comprehensively search through 17559750 patent applications and granted patents from the European Patent Office and the United States Patent and Trademark Office up to 2024, identifying 125716 battery-related patents. Utilizing these extensive literature and patents, we conduct numerous experiments to evaluate the structured output capabilities of knowledge base, contextual learning, instruction adherence, and language models. Through multi-dimensional model evaluations and analyses, the following points are found. First, the model exhibits high accuracy in screening literature on inorganic solid-state electrolytes, equivalent to the level of a doctoral student in the relevant field. Based on 10604 data entries, the model demonstrates good recognition capabilities in identifying literature on in-situ polymerization/solidification technology. However, its understanding accuracy for this emerging technology is slightly lower than that for solid-state electrolytes, requiring further fine-tuning to improve accuracy. Second, through testing with 10604 data entries, the model achieves reliable accuracy in extracting inorganic ionic conductivity data. Third, based on solid-state lithium battery patents from four companies in South Korea and Japan over the past 20 years, this model proves effective in analyzing historical patent trends and conducting comparative analyses. Furthermore, the model-generated personalized literature reports based on the latest publications also show high accuracy. Fourth, by utilizing the iterative strategy of the model, we enable DeepSeek to engage in self-reflection thinking, thereby providing more comprehensive responses. The research results indicate that language models possess strong capabilities in content summarization and trend analysis. However, we also observe that the model may occasionally experience issues with numerical hallucinations. Additionally, while processing a large number of battery-related data, there is still room for optimization in engineering applications. According to the characteristics of the model and the above test results, we utilize the DeepSeek V3-0324 model to extract data on inorganic solid electrolyte materials, including 5970 ionic conductivity entries, 387 diffusion coefficient entries, and 3094 migration barrier entries. Additionally, it includes over 1000 data entries related to chemical, electrochemical, and mechanical properties, covering nearly all physical, chemical, and electrochemical properties related to inorganic solid electrolytes. This also means that the application of large language models in scientific research has shifted from auxiliary research to actively promoting its development. The datasets presented in this paper may be available at the website: <ext-link ext-link-type="uri" xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="https://cmpdc.iphy.ac.cn/literature/SSE.html">https://cmpdc.iphy.ac.cn/literature/SSE.html</ext-link> (DOI: <ext-link ext-link-type="uri" xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="https://doi.org/10.57760/sciencedb.j00213.00172">https://doi.org/10.57760/sciencedb.j00213.00172</ext-link>).