Presentation Information
[2Yin-A-22]Accelerating Materials and Process Development via Symbolic Regression and Literature-Search Agents
〇Masafumi Tsuyuki1, Shotaro agatsuma1, Ryota Yagi2 (1. R&D Group, Hitachi, Ltd., 2. Government & Public Corporation Information Systems Division, Hitachi, Ltd.)
Keywords:
Symbolic Regression,Genetic Algorithm,AI Agent,Materials Informatics,Data-driven Materials Design
Accurate physical interpretation of experimental data is essential for accelerating materials and manufacturing process development. However, this process remains a bottleneck due to its reliance on expert intuition for handling complex theoretical models and extensive literature reviews. To address this problem, we propose a novel data interpretation framework integrating symbolic regression (SR) and a literature-retrieval agent. Our approach employs a hybrid SR method combining genetic programming and Large Language Model (LLM) to derive mathematical expressions that achieve both numerical precision and physical interpretability. Subsequently, an agent utilizes these expressions as search queries to verify consistency with prior studies and synthesize explanatory insights into the underlying physics. Validated on the LLM-SR benchmark (chemical kinetics) and electrode manufacturing data, our method outperforms existing baselines by improving formula structural accuracy by 18% and reducing mean squared error by an order of magnitude, effectively automating the interpretation of complex physical phenomena.