Framework to completely bypass expensive DFT calculations via graph neural networks for vacancy formation energy predictions in FCC high entropy alloys

The compositional complexity and chemical randomness of high entropy alloys (HEAs) make conventional atomic-scale calculations, such as density functional theory (DFT), prohibitively expensive for property prediction. One key property of interest is the vacancy formation energy ( \({E}_{v}^{f}\) ), which plays a crucial role in diffusion and microstructure evolution. In this work, we present a machine learning (ML) framework that eliminates the need for DFT calculations by predicting \({E}_{v}^{f}{\rm{s}}\) in HEAs using models trained on binary and ternary alloys. Our approach first relaxes face-centered cubic (FCC) structures using a fine-tuned CHGNet model and then uses the resulting configurations as input into a crystal graph convolutional neural network (CGCNN) to predict both Bader charges and \({E}_{v}^{f}{\rm{s}}\) . Incorporating Bader charges as descriptors introduces DFT-informed electronic structure information into the model, significantly improving prediction accuracy compared to using elemental features alone. Furthermore, we demonstrate that the model’s performance generalizes well to other alloy systems with minimal fine-tuning, offering a robust and efficient path toward high-throughput defect property prediction in complex alloys.