Data-augmented crystal graph convolutional neural network for uncertainty-aware thermodynamic stability prediction of two-dimensional MXene/MoS2 materials

초록

Two-dimensional (2D) MXenes and molybdenum disulfide (MoS2) are promising functional materials, but density functional theory (DFT) evaluation of thermodynamic stability is computationally expensive. This study develops a crystal graph convolutional neural network (CGCNN) framework to predict formation energy ( E f) and energy above the convex hull ( E hull) from a limited dataset of 414 Materials Project structures. The limited dataset represents epistemic uncertainty, while controlled atom-wise perturbations represent local structural variability, whereas rigid translation is used as an invariance control. CGCNN hyperparameters were selected using an L9(34) orthogonal array with analysis of means and range analysis. Among the evaluated augmentation strategies, the combined shift case achieved the lowest mean absolute errors of 0.0052 and 0.0058 eV atom−1 for E f and E hull, respectively. Separately, the three repeated baseline models were used to construct validation-calibrated 95% prediction intervals for uncertainty-aware screening of held-out structures, with 97.6% empirical coverage for both targets. The results demonstrate a practical small-data workflow for thermodynamic stability prediction while highlighting reduced reliability in sparsely represented unstable regions. © 2026 The Authors.

키워드