Learning Full-Field System Dynamics via Digital Twin Supervised Graph Models

Learning Full-Field System Dynamics via Digital Twin Supervised Graph Models

Structural health monitoring is fundamentally challenged by the need for full-field structural assessment despite measurements being available only at sparse sensor locations. This work presents a Digital Twin-informed Graph Neural Network framework that combines physics-informed graph representations with data-driven learning to approximate high-fidelity structural simulations at low computational cost. The proposed spatiotemporal architecture integrates Graph Attention Networks, recurrent neural networks, and cross-attention mechanisms to propagate information from measured to unmeasured structural locations. Trained using Digital Twin-generated full-field responses, the framework enables efficient sparse-to-full-field prediction while preserving key structural relationships. Validation on the IASC-ASCE benchmark demonstrates promising performance across multiple damage scenarios, highlighting the potential of physics-informed GNNs for computationally efficient digital twins and real-time full-field structural monitoring using sparse sensors.