| Abstract Scope |
The fluctuating energy landscape due to chemical disorder in refractory multi-principal element alloys (RMPEAs) can mitigate strength–ductility trade-off. While attention-based graph neural networks (AGNN) have shown success in prediction of various materials properties, the lack of disorder-informed descriptors can lead to incorrect alloy mechanical property predictions. This research focuses on developing novel descriptors that accurately capture the effect of chemical disorder on electronic properties. By utilizing AGNN, combined predictions of strength and ductility will result in a better RMPEA predictions. In this study, a AGNN is trained on unary, binary and few ternary disordered alloys composed of Mo, Nb, Ta, W, V and Ti across systematically-varied compositions. After training, the network extrapolates disorder-electronic property relationships to multi-component RMPEAs. By integrating these physics-grounded features with an attention-based learning architecture, the ML model can produce more reliable, disorder-aware strength-ductility predictions for design of novel structural RMPEAs. |