| Abstract Scope |
In this study, we employ three machine learning (ML) models - XGBoost, graph neural networks (GNN), and graph attention networks (GAT) - to predict unstable stacking fault energies (USFEs) across more than 1,000 refractory non-dilute random alloys, encompassing mono-, binary, ternary, quaternary, and quinary systems. Training data are generated via atomistic simulations, with each alloy composition encoded as a five-dimensional vector and normalized to ensure balanced input features. While explicit inter-composition relationships are not encoded, GNN- and GAT-based models capture compositional trends through node features and internal attention mechanisms. Additionally, we develop corresponding models to predict lattice parameters, a secondary output of the simulations. Among all approaches, the GAT model demonstrates the best performance, surpassing both GNN and XGBoost. These results underscore the promise of graph-based ML frameworks in accurately predicting materials properties from simulation-derived data, potentially advancing computational materials design. |