A crystal graph convolutional neural network framework for predicting stacking fault energy in concentrated alloys

The stacking fault energy (SFE) in concentrated alloys is highly sensitive to the local atomic environment. Understanding this relationship necessitates extensive sampling of configurations, which significantly increases computational demands. To address this challenge, we propose a graph neural network (GNN)-based framework utilizing unrelaxed bulk and stacking fault structures as inputs to directly predict SFE. We first investigate two key extrapolation capabilities for bulk formation energy prediction: scale extrapolation (predicting formation energy in larger supercells) and compositional extrapolation (predicting formation energy for compositions beyond the training set). Leveraging the validated model, we concurrently predict formation energies of both bulk and stacking fault configurations to compute the SFE. The structural similarity between these configurations enables efficient parameter sharing, accelerating model convergence. The framework demonstrates excellent interpretability and robust compositional extrapolation capabilities in predicting SFEs. Furthermore, leveraging its exceptional compositional extrapolation, we integrate the model with Monte Carlo simulations to successfully predict ordering behavior and solute segregation at stacking faults. Finally, we introduce a hierarchical training strategy that further reduces data requirements. Collectively, our work establishes a unified and efficient framework for robust prediction of planar fault energies in complex concentrated alloys.