Enhancing Groundwater Contaminant Transport Forecasts Through Synergy of Machine Learning and Physics‐Based Models

Abstract Groundwater is vital for water supply, irrigation, and environmental sustainability, yet contamination remains a global challenge. Remediation of contaminated aquifers is hindered by difficulties in tracking contaminant spread and identifying pollutant sources. Although advances in aquifer characterization have improved groundwater modeling, limitations persist in capturing spatial heterogeneity and accurately predicting contaminant transport. This work presents a hybrid modeling framework integrating a physics‐based travel‐time model with a multi‐output feed‐forward Deep Neural Network (DNN) to parametrize local BreakThrough Curves (BTCs) at sampling and compliance locations. Focusing on local concentration breakthrough curves rather than hydraulic head or distributed aquifer properties, the method better captures the effects of spatial variations in hydraulic conductivity on the dynamics of the contaminant plume. Model performance was evaluated at different levels of information, and extended to ungauged locations to assess generalization capability. The DNN accurately infers spatially variable transport parameters, with Wasserstein distances consistently lower than those of the moments‐based benchmark when concentration data are sufficient. Performance degrades gradually with reduced data, yet our hybrid approach is less impacted by partial sampling of the BTCs than the traditional moments‐based methods. DNN‐inferred parameters performed better than ensemble‐mean BTC predictions, demonstrating the added value of site‐specific inference in capturing local concentration dynamics that ensemble statistics fail to reproduce. This robustness is attributed to the physical constraints embedded in the transport model and to the cross‐regime learning capability of the DNN, establishing the proposed framework as a practical tool for subsurface characterization under typical field monitoring conditions.