Predicting Groundwater Flow and Solute Transport in Tidally Forced Coastal Aquifers Using DeepONet

Coastal aquifer responses to tides can be simulated using coupled groundwater flow and solute transport models, but repeated numerical simulations are computationally costly. We developed a Deep Operator Network (DeepONet) surrogate for an idealized two-dimensional coastal aquifer simulated with SUTRA. The model maps tidal amplitude and landward hydraulic head to transient pressure head and concentration fields. Supervised, physics-regularized, and physics-only training strategies were compared, and the supervised model was evaluated at higher tidal amplitudes using few shot adaptation. On test simulations within source-range, the supervised model achieved pooled RMSEs of 5.87×10⁻³ m for pressure head and 5.75×10⁻⁴ for concentration. Adding continuous physics constraints did not improve agreement with the numerical reference, and lower physics residuals did not necessarily correspond to smaller field errors. Prediction accuracy decreased outside the source range, whereas adaptation with two higher-amplitude simulations reduced pressure-head and concentration RMSEs by 36.8% and 68.5%, respectively. The surrogate reconstructed a complete 72-h response in approximately 0.018 s, compared with approximately 15 s for the numerical simulation. These results demonstrate the potential of DeepONet as a computationally efficient surrogate for tide-driven coastal aquifer simulations while showing the dependence of physics-based regularization and adaptation on training data and forcing-space coverage.