About this Abstract |
Meeting |
2026 TMS Annual Meeting & Exhibition
|
Symposium
|
Thermodynamics and Kinetics of Alloys IV
|
Presentation Title |
Integration of Large-Language Model and CALPHAD for Alloy Design Hypothesis Refinement |
Author(s) |
Quanliang Liu, Hyunseok Oh |
On-Site Speaker (Planned) |
Hyunseok Oh |
Abstract Scope |
Large-language models (LLMs) present a new opportunity for synthesizing novel hypotheses in materials design. However, the text generated by LLMs typically lacks constraints from physical laws unless explicitly trained on domain-specific databases. Therefore, additional steps are needed to refine these hypotheses to ensure physical plausibility. In this study, we develop an integrated LLM-CALPHAD framework designed to bridge the knowledge gap inherent in LLM-generated hypotheses. To refine a given hypothesis, the framework enables the LLM to reference CALPHAD handbooks and related research literature for similar calculations. Consequently, the LLM suggests appropriate CALPHAD modules and databases and automatically generates input files, including calculation parameters, for automated computations. Feedback from CALPHAD calculations iteratively refines the hypotheses generated by the LLM. This CALPHAD-guided approach enhances the potential for disruptive discovery of novel materials while maintaining scientific and engineering plausibility. |
Proceedings Inclusion? |
Planned: |
Keywords |
Computational Materials Science & Engineering, Machine Learning, Phase Transformations |