Data-driven Early Cycle Prediction Using Large Language Model and Graph Mining for Lithium Metal Battery

Lithium metal batteries (LMBs) represent a promising next-generation energy storage technology, characterized by their high capacity, low density lithium metal anodes, and low electric potential. Despite these advantages, challenges such as low Coulombic efficiency (CE) and limited lifespan due to undesirable side reactions on the solid-electrolyte interphase (SEI) and uneven lithium stripping/plating persist. Addressing these issues necessitates the development of novel materials for battery components.<br/><br/>Battery performance is heavily influenced by the structure and composition of materials, such as high nickel cathodes, high stability electrolytes, and high entropy separators. While understanding individual material characteristics is important, comprehensively understanding their interactions is essential for optimizing battery performance, lifespan, and stability. To fully harness these innovative battery technologies, a data-driven approach to developing materials based on a holistic understanding of the entire battery cell is necessary.<br/><br/>This like development tackles two critical challenges: (1) how to efficiently collect high-quality data on battery materials and performance, and (2) how to effectively gather information on various material combinations. In this study, we present a data-driven approach to accelerate the discovery and optimization of materials for LMBs. We developed a comprehensive dataset by integrating extensive battery material information with performance metrics, utilizing a large language model (LLM) and the Material Graph Digitizer (MatGD) tool. The LLM, an advanced artificial intelligence (AI) system trained in natural language processing, extracted high-accuracy data on material structure and composition from scientific literature, experimentally validated, through prompt engineering and few-shot learning. Concurrently, MatGD, a graph mining tool, extracted quantitative data from cycle performance and electrochemical impedance spectroscopy (EIS) graphs, consolidating this information into a unified dataset.<br/><br/>This dataset enable the training of supervised machine learning models to predict and design materials with high capacity and long-term lifespan for LMBs. Our models identify material combinations that significantly enhance battery cycle life and stability. The promising LMB compositions discovered through this approach are subsequently validated through experimental testing.<br/><br/>Our research proposes a novel data-driven methodology that not only enhances the understanding of material interactions within battery cells but also expedites the development of high-performance LMBs, potentially revolutionizing the landscape of energy storage technologies.