Simulation-supervised ViT-CNN for peak-stress classification and attribution assessment in concrete microstructures

Image-based deep learning offers efficient prediction of complex material responses, but its integration with computational mechanics requires a clear distinction between predictive performance and physical interpretability. This study presents a simulation-supervised framework combining finite element damage mechanics with a hybrid Vision Transformer–convolutional neural network for peak-stress classification and attribution assessment in heterogeneous concrete microstructures. Three-dimensional aggregate geometries were reconstructed from micro-computed-tomography images segmented using a U-Net architecture. Monte Carlo sampling was then used to generate 330 two-dimensional microstructures with controlled nominal pore area fractions between 1 % and 11 %. Nonlinear finite element analyses provided peak-stress responses, class labels, and tension-damage fields. The samples were divided into higher- and lower-peak-stress response groups using the dataset mean, which represents a reproducible statistical partition rather than a physical damage threshold. On 100 held-out images, the proposed model achieved 92.0 % accuracy and a weighted F1-score of 91.9 %, compared with 85.0 % accuracy for the ResNet-50 baseline. Dataset-level comparison showed that the mean Spearman correlation between layer-wise relevance propagation maps and matched finite element damage fields was 0.023, compared with − 0.008 for mismatched fields and 0.122 for a pore-proximity baseline. These findings show that the model provides accurate simulation-based response classification, while its attribution maps should not be interpreted as substitutes for mechanically calculated damage fields. External validation using experimentally representative pore populations remains necessary.