High-Throughput and Robust Materials Design Hypothesis Generation via a Large Language Model

The increasing complexity of materials design requires advanced methods to manage and transfer extensive information across various domains. This research utilizes GPT-4, a large language model (LLM), to mimic one of the human design activities by synergistically combining mechanisms from research articles to generate materials design hypotheses. First, through prompt engineering, the model accurately extracts key information from individual articles using the materials system chart framework—a structured framework comprehending the dynamic interplay among Processing, Structure, and Property. Then, it synthesizes innovative hypotheses from this information, including design ideas for high entropy alloys with superior cryogenic properties and halide solid electrolytes with high ion conductivity and formability. Notably, recent publications have validated these hypotheses, demonstrating the LLM’s ability to generate novel ideas not previously established in the literature. This LLM-assisted design approach enhances the potential for disruptive discoveries of new materials while maintaining scientific and engineering plausibility.