Materials Research Agent

Materials discovery pipelines rely on the availability of high-quality datasets, accurate predictive models, and exploration tools for searching through published experimental and simulation-based findings. The number of documents containing valuable data (e.g., peer-reviewed research papers, patents, and handbooks) is quite high, making manual compilation of datasets a challenging task. Further, developing reliable machine learning models for predicting material properties with uncertainty estimates and explanations is highly desirable for rational experimental planning. Therefore, we propose an agentic approach for materials discovery that combines (a) information extraction from research papers to create machine-readable materials composition--property datasets, (b) training uncertainty-aware property prediction models, (c) delineating the effect of input features on material properties through explainable AI techniques, and (d) materials selection charts to assist in identifying potential compositions of interest required to push existing material--property frontiers.