Multi-Hazard Graph Artificial Intelligence for Infrastructure Resilience Screening in Romania: Integrating Ground Motion, Earth Observation, Hazard Exposure and Network Topology

Natural hazards can disrupt transport infrastructure through spatially heterogeneous flooding, seismic shaking, terrain conditions, and observed ground motion, while local exposure alone does not determine system-wide consequences. This study develops a reproducible multi-hazard Graph Artificial Intelligence framework for infrastructure resilience screening on the Romanian strategic road network. A frozen attributed line graph combines network topology with European land-cover, terrain, seismic, flood, and ground-motion evidence, and an independent disruption simulator provides continuous resilience-loss targets. GCN, GraphSAGE, and GAT were trained under a leakage-controlled three-seed protocol. On the blind 194-scenario PRIMARY TEST, ensemble RMSE/R2 were 0.013343/0.997104 for GCN, 0.013978/0.996822 for GraphSAGE, and 0.015830/0.995924 for GAT. A fresh GCN branch achieved RMSE/R2 = 0.018230/0.995852 on 111 component-disjoint cases and 0.010468/0.991654 on 256 severity-specific held-out cases. On 208 predefined external anchors held out from Graph-AI fitting and selection, GAT attained the lowest ensemble RMSE (0.021620; R2 = 0.975065), while seismic-only transfer remained weak (n = 10; R2 = 0.251–0.386). A frozen six-family GCN ablation (18 checkpoints; 194 matched test scenarios) found RMSE = 0.026175 after seismic feature omission versus 0.013343 for the full-input GCN; this is simulator-specific sensitivity, not causal hazard attribution. The results support accurate emulation of the frozen simulator and transfer-aware screening, not structural safety certification. The model outputs are simulator-based rather than observed disruption outcomes; neither an end-to-end simulator-to-GNN speedup nor calibrated prediction intervals are established by the present evaluation.