Prediction of Oxygen Vacancy Formation Energy in Oxides Based on the Graph Neural Network Approach

Oxygen vacancy formation energy (Eovf) prediction plays an important role in addressing the issue of performance evaluation in various applications of materials. In this study, a convolutional and attentional graph neural network (CAGNN) was proposed for the prediction of charge-neutral Eovf at arbitrary oxygen sites of oxide crystals. The crystal structure and atomic properties were the only needed input. For the elements in the training set distribution, the mean absolute error (MAE) of predicting Eovf with CAGNN was 0.50 eV. The computational cost of CAGNN was significantly reduced compared to density functional theory (DFT). The prediction of Eovf (oxides containing n cationic elements, n = 2, 3, 4) with the CAGNN model trained by the data of oxides containing m (m < n) cationic elements were carried out, which validated the model’s generalization and extrapolation performance. Finally, through the materials rediscovery tests, CAGNN shows promising application prospects in material discovery for fields with a strong correlation to the Eovf.