| Abstract Scope |
The accelerated development of fusion structural materials requires comprehensive, high-quality databases compiled from increasingly diverse scientific literature and online resources. We present Mat-FUS, a large language model (LLM)-driven data extraction framework specifically developed for fusion materials research. Mat-FUS performs multimodal information extraction from scientific text, tables, figures, and PDFs, allowing for automated collection of alloy compositions, processing histories, irradiation conditions, microstructures, mechanical properties, and nuclear performance metrics. To improve extraction accuracy and robustness, the framework leverages DSPy to replace manual prompt engineering with modular, optimizable LLM programs, enabling systematic optimization of prompts, model parameters, retrieval-augmented generation, and agent-based reasoning. Applied across fusion materials literature, Mat-FUS generates structured databases aiding rapid data discovery and surrogate model development for activation, transmutation, helium production, radiation damage, and materials informatics. The framework provides a scalable foundation for FAIR fusion materials databases and AI-assisted materials design for future reactors. |