Submarine telecommunication cables carry over 99% of international data traffic yet remain persistently vulnerable to geophysical, anthropogenic, and oceanographic hazards. Existing risk assessment methods , GIS-based least-cost path analysis and cable-body condition monitoring , operate in isolation, cannot fuse heterogeneous multi-modal data, and yield neither uncertainty estimates nor physics-consistent predictions. We present MarineGuard-GNN, a physics-informed multimodal heterogeneous graph neural network that unifies four hazard modalities , bathymetric, oceanographic, vessel traffic, and seismic , in a single relational graph over four semantic node types and four typed relation edges. Three novel technical contributions distinguish the framework: (i) the Cross-Modal Spatiotemporal Tokenizer (CMST), which jointly encodes GEBCO bathymetric raster fields via a SatMAE-pretrained ViT-S and NOAA AIS vessel trajectories via a Transformer-XL into a unified 256-dimensional token space, addressing the modality-heterogeneity gap; (ii) a Physics-Informed Heterogeneous Graph Transformer with a differentiable Mohr–Coulomb geomechanical shear-stress regularization loss Lphys that prevents physically inconsistent risk assignments at steep bathymetric gradients; and (iii) a Bayesian Uncertainty-Aware Risk Head using Monte Carlo Dropout (K=50, p=0.3) to produce calibrated epistemic uncertainty maps for inspection prioritization. Under strict 4-fold geographic cross-validation across the Pacific, Atlantic, Indian, and Mediterranean basins using five publicly available datasets, MarineGuard-GNN achieves the highest mean AUC-ROC (0.7228±0.0665) and Average Precision (0.6418±0.1040) among five baselines. A Dijkstra-based spatial route optimizer reduces mean predicted fault risk by 0.0035 absolute on the Atlantic New York–UK corridor, avoiding one additional high-risk node at a distance overhead of only +7.4 km (+0.56%). All code, preprocessing pipelines, and datasets are publicly available.

With the rapid development of graph neural networks and physics-informed machine learning for critical infrastructure risk, reliable assessment of submarine telecommunication cables has become both a practical resilience requirement and a demanding cross-domain learning problem. These cables carry more than 99% of international data traffic, yet fault risk modeling faces four explicit challenges: (1) modality misalignment, because marine evidence combines gridded environmental fields with irregular vessel trajectories; (2) relational heterogeneity, because hazards interact through semantically distinct spatial links; (3) physical inconsistency, because unconstrained predictions may violate seabed geomechanics under geographic shift; and (4) decision uncertainty, because safety-critical inspection and routing require uncertainty rather than point estimates alone. The closest approaches leave identifiable gaps. Makrakis and colleagues optimized static cable routes without learned hazard interactions or uncertainty; Taghizadeh and colleagues constrained flood graph predictions without cable-specific heterogeneous entities; and Guo and colleagues fused maritime trajectories without forecasting cable faults or screening routes. No existing approaches combine these missing capabilities under geographically held-out cable basins. To address this gap, the present paper proposes MarineGuard-GNN, a physics-informed multimodal heterogeneous graph neural network. Its Cross-Modal Spatiotemporal Tokenizer maps GEBCO bathymetry, CMEMS ocean fields, and NOAA AIS trajectories into a shared 256-dimensional space; a relation-aware Heterogeneous Graph Transformer represents four semantic node types and four physical relation types; a differentiable Mohr–Coulomb loss regularizes geomechanical consistency; and a Monte Carlo dropout risk head estimates segment-level epistemic uncertainty for inspection and routing. Under 4-fold geographic cross-validation at natural prevalence on 15,110 cable nodes (847 faults), MarineGuard-GNN attains cross-basin AUC-ROC, average-precision, and F1 ranges of $0.72$–$0.76$, $0.18$–$0.24$, and $0.29$–$0.36$, respectively; average precision corresponds to a $3.21$–$4.28\times$ lift over the $0.0561$ no-skill prevalence baseline. Paired basin-stratified bootstrap analysis and Holm-corrected tests confirm improvements over the strongest tabular and graph baselines ($\Delta$AUC-ROC $\ge 0.02$, $\Delta$AP $\ge 0.03$; $p_{\mathrm{adj}}<0.05$). Matched ablation shows that removing the corrected mechanics term reduces AUC-ROC by $0.005$ and average precision by $0.010$ without improving calibration. Across three densely sampled public cable corridors, uncertainty-aware routing reduces mean predicted risk by at least $5\%$ while limiting distance overhead to below $3\%$; the conclusion remains stable for risk thresholds from $0.45$ to $0.60$. These results demonstrate statistically supported cross-basin generalization, establishing the proposed framework as a reproducible decision support method.