Graph neural network surrogate enables multi-zone elastic modulus inversion in ultra-high arch dams

Elastic modulus controls hydrostatic deformation and stiffness identification in ultra-high arch dams. Multi-zone inversion remains difficult because finite element forward analyses are costly, conventional neural networks do not naturally represent unstructured finite-element topology, and monitoring-error minimisation alone can yield mechanically implausible parameters. Here we couple a finite-element-mesh-based graph neural network surrogate with physics-constrained optimisation for multi-zone elastic modulus inversion. An adaptive hydrostatic-temperature-time model separates hydrostatic deformation from long-term monitoring records. Latin hypercube sampling and finite element analyses build a modulus-response dataset. Finite element nodes are represented as graph nodes, element-edge connectivity as graph edges, and material zones, elastic moduli and monitoring-point information as graph features. A GraphSAGE network learns the nonlinear mapping from zoned moduli to monitored hydrostatic displacements. The objective combines surrogate prediction error with bounds, engineering priors, material-grade ordering and within-group dispersion constraints, and is solved by particle swarm optimisation. In the engineering case, PSO achieved the lowest single objective value of 0.8575 and an average value of 0.9330 over ten runs. Finite element verification gave a maximum absolute hydrostatic-deformation error of 0.9550 mm at the four selected monitoring points under the target operating condition, supporting the feasibility of the proposed framework for the selected engineering case.