Physics-informed spatiotemporal graph neural network models for groundwater contaminant prediction

Area of Science:

• Environmental Science
• Hydrogeology
• Artificial Intelligence

Background:

• Groundwater contamination prediction faces challenges due to physical model complexity and data-driven model limitations in spatial representation.
• Existing methods struggle with the spatiotemporal coupling and parameter dependence inherent in groundwater systems.

Purpose of the Study:

• To develop and compare three physics-driven spatiotemporal graph neural network (STGNN) architectures for enhanced groundwater contamination prediction.
• To integrate physical constraints (Darcy flow) into data-driven models using a dynamic graph construction strategy.

Main Methods:

• Three STGNN architectures were developed: recurrent (RGNN-ST), convolutional (ConvGNN-ST), and attention-based (AttentionGNN-ST).
• A dynamic graph construction strategy embedded hydraulic head and conductivity into edge weights, linking to Darcy-flow.
• Models were trained and tested using MODFLOW-MT3DMS simulations on homogeneous and heterogeneous aquifers.

Main Results:

• RGNN-ST achieved high accuracy (R² up to 99.95% in homogeneous, 99.60% in heterogeneous conditions) with low NRMSE (1.68%).
• All models showed robustness to monitoring density, time window, and aquifer heterogeneity, accurately capturing contaminant migration.
• RGNN-ST offered the best stability and accuracy-efficiency balance, with interpretable predictions aligning with groundwater physics.

Conclusions:

• Physics-informed STGNNs offer a transferable framework for groundwater contamination forecasting and risk management.
• RGNN-ST provides a reliable and interpretable approach, balancing multi-source physical features for accurate predictions.
• The study highlights the potential of integrating physics with AI for optimizing monitoring networks and improving groundwater management.