Finite element data-driven deep learning-based tensile failure analysis of precast bridge slab joint.

• Efficient strain predictions are achieved with DeepLabv3+, enhancing the speed and accuracy of structural analysis. • Rapid and precise failure mode determinations are made possible by VGG-16, employing focal loss for efficient classification. • Accurate interpretability is ensured through Grad-CAM, aligning neural focus areas with fracture mechanics theory. • The method significantly improves efficiency, providing quick and confident results without the need for labor-intensive re-modeling during the FEA process. The joint is a critical component that facilitates force transmission across precast bridge slabs while maintaining structural integrity. Finite Element Analysis (FEA) is commonly used to examine the mechanical properties and load-bearing capacity of joints with different configurations. However, the repetitive process of "modeling-meshing-computing" consumes substantial manpower, time, and expertise. This study proposes an approach consisting of three key steps: (1) predicting strain distributions, (2) determining failure modes, and (3) analyzing interpretability of predicted results. Specifically, DeepLabv3+ is used to map the Finite Element (FE) model to strain distributions. A binary classification model based on VGG-16 predicts the failure state of the specimen. Interpretability analysis, using Gradient-weighted Class Activation Mapping (Grad-CAM), provides insights into the neural network's decision-making process. The optimized models achieve a Mean Intersection over Union (MIoU) of 0.5691 for strain distribution prediction on the cross-sectional surface and 0.6424 on the front surface. The failure detection model achieves an F1 score of 0.992 in determining failure states. The interpretability results align well with fracture mechanics theory, enhancing confidence in the model's predictions. This method serves as a surrogate to reduce the workload of finite element analyses, facilitating strain analysis and failure detection. [ABSTRACT FROM AUTHOR]