Rapid assessment of contaminant migration in complex heterogeneous aquifers via transferable attention-enhanced graph neural networks.

Highly heterogeneous aquifer structures and complex boundary disturbances pose persistent challenges for predicting the fate and transport of hazardous contaminants, limiting the feasibility of early-warning systems and rapid risk assessment. While AI-based surrogate models can substantially accelerate numerical simulations, most existing models are developed for individual aquifer settings with simplified boundary conditions, limiting their applicability to new hydrogeological realizations and boundary-condition scenarios. Here, we develop a transferable surrogate modeling framework based on graph neural networks (GNNs) and attention-enhanced graph neural networks (aGNNs) to enable instantaneous and accurate prediction of groundwater heads and contaminant plumes under highly heterogeneous conditions. The proposed framework explicitly encodes spatial topology and complex boundary constraints, enabling direct prediction for previously unseen heterogeneous aquifer realizations, pumping and injection scenarios, and localized constant-head boundary configurations without retraining, provided that they remain within the predefined hydrogeological parameter space represented during training. Numerical experiments demonstrate that the GNN model provides stable hydraulic head predictions, while the aGNN accurately captures contaminant plume morphology and preferential migration pathways, achieving R2 > 0.98 across diverse synthetic heterogeneous aquifer realizations. The proposed framework provides a computationally efficient surrogate for rapid scenario evaluation and offers a foundation for future groundwater contamination assessment and remediation planning under complex hydrogeological conditions.