Interpretable and Physics-Informed Modeling of Solidification in Alloy Systems: A Generalized Framework for Multi-Component Prediction

Abstract

This study introduces AlloyGCN, a graph-based neural network designed to predict the solidus and liquidus temperatures of multi-element alloys with high accuracy and interpretability. Leveraging a graph representation where nodes and edges capture elemental properties and interactions, AlloyGCN effectively generalizes across diverse alloy compositions, including unseen systems, while requiring less extensive data than traditional thermodynamic methods. Experimental validation across various alloy systems demonstrated the model's reliability, achieving predictive accuracy comparable to CALPHAD calculations. Integration of explainable AI techniques provided valuable insights into the physical factors influencing predictions, and derived empirical equations offer practical tools for alloy design. By combining advanced machine learning with interpretable frameworks, this work contributes to the development of efficient and reliable methods for alloy property prediction and materials discovery.