A transformer-based deep learning approach for mechanical response prediction and failure analysis of precast bridge slab joints using strain field

The integration of artificial intelligence (AI) with finite element analysis (FEA) offers a promising approach for predicting structural mechanical and displacement responses. However, existing AI-FEA studies predominantly focus on single-material structures or simplified loading conditions, lacking a systematic framework for evaluating macro-scale responses of composite structures. To address this gap, this study proposes a failure analysis framework based on an enhanced U-shaped Transformer network, which combines the global dependency modeling capacity of attention mechanisms with a U-shaped encoder–decoder design well-suited for extracting strain-field features. The framework is applied to a novel ultra-high performance concrete (UHPC)-headed bar joint within accelerated bridge construction (ABC) systems. The enhanced Transformer-based model predicts internal rebar stresses and pull-out displacements by extracting critical features from strain field distributions. To evaluate real-world generalizability, artificial downsampling strategies were applied to simulate resolution limitations and systematic errors inherent in experimental strain measurements. Results demonstrate that the proposed method achieves high prediction accuracy for rebar stresses (MAE < 2.7 MPa) and pull-out displacements (MAE < 0.05 mm). Furthermore, discrete stress–displacement states were fitted into a logarithmic model with low weighted orthogonal distance squared sum (W-ODSS), establishing an empirically derived failure model guided by physical observations. This criterion achieved 97.2 % accuracy in failure assessment under diverse adverse conditions. Additional FEA simulations extending beyond the original dataset parameter ranges confirmed the model’s robust predictive performance. While the present study focuses on slab-to-slab connection joints, the proposed methodology shows potential for future extension to other critical engineering scenarios, such as beam–column connections and composite slab joints. More broadly, with the integration of digital image correlation (DIC) or embedded sensing systems, this framework could contribute to advancing digital twin approaches for infrastructure lifecycle management.