Efficient damage identification in hollow slab girder bridges using a hybrid graph-attention network and differentiable inversion

In hollow slab girder bridges, which are prevalent within the highway network, hinge joint degradation critically disrupts transverse load distribution, threatening overall structural safety when coupled with girder deterioration. However, traditional finite element model updating (FEMU) for damage identification in such multi-component systems suffers from the curse of dimensionality and high computational costs, while conventional surrogate models lack topological generalizability. This paper proposes an automated damage identification method driven by a Hybrid Graph-Attention Network (HG-AttnNet) and differentiable optimization inversion. By modeling bridges as chain graphs, HG-AttnNet acts as a topology-aware surrogate that maps structural parameters to deflections across diverse configurations. Leveraging automatic differentiation, the framework enables exact gradient-based inversion of multi-component damage, bypassing inefficient numerical gradient evaluations. Validated on a field-tested bridge, HG-AttnNet predicted deflections with an RMSE of 0.005 mm. During inversion, the method accurately isolated degraded components and reconstructed measured deflections (RMSE = 0.040 mm) in just 13.46 s, reducing computational time by over 99.5% compared to FEMU. This framework provides a transformative paradigm for rapid, component-level evaluation of extensive bridge inventories.