Physics audited and uncertainty aware surrogate modeling for reactive nitrate transport in groundwater

Nitrate ( ) contamination in groundwater often persists due to donor limitation, redox competition, and long residence times that constrain natural attenuation. This study presents a physics audited and uncertainty aware surrogate modeling framework for reactive nitrate transport in porous media. High-fidelity PHREEQC simulations were generated for four one-dimensional benchmarks of increasing geochemical complexity, ranging from linear heterotrophic denitrification to dual-substrate Monod kinetics with explicit - competition. The surrogate models were trained directly on space-time concentration fields using the CatBoost gradient boosting algorithm, with hyperparameters tuned via Bayesian optimization and simulation level data partitioning to prevent leakage. The surrogate takes hydrodynamic and geochemical inputs along with spatial and temporal coordinates, and predicts nitrate concentration fields across the domain. Uncertainty quantification was performed using a subsampling ensemble to characterize epistemic variability and conformalized quantile regression to provide calibrated prediction intervals. A physics audit was applied post hoc to verify consistency with the governing advection-dispersion-reaction balance, solute mass conservation, and non-negativity. Results show that the surrogates accurately reproduce nonlinear reaction fronts and donor-acceptor competition, achieving test R² up to 0.997 with mass balance errors typically below 0.5%. Epistemic uncertainty remains small relative to aleatoric variability, which is concentrated along reactive transition zones. The presented framework is a reliable and computationally efficient tool for scenario analysis and risk-informed groundwater quality management in nitrate-impacted aquifers.