Estimation of non-point contaminant source location and release history in an aquifer via restart ensemble Kalman...

Identifying sources of groundwater contamination and their corresponding release histories is crucial for effective contaminant control and remediation strategies. Although point sources have been extensively studied, non-point sources-characterized by irregular geometries and dynamic release patterns-remain difficult to identify due to significant uncertainties and the high computational demands of high-fidelity models. To address these challenges, we propose an inversion framework (ODRDCN-rEnKF) that integrates an optimized deep residual dense convolutional network (ODRDCN) with a restart ensemble Kalman filter (r-EnKF). ODRDCN serves as a surrogate for the forward model, efficiently capturing the complex relationship between source parameters and concentration distributions. Combined r-EnKF, the framework simultaneously identifies the spatial configuration of non-point sources-represented by rotated ellipsoids-and their time-varying release history-modeled using parameterized exponential functions. Testing in a synthetic aquifer shows that ODRDCN achieves accurate surrogate modeling with minimal training data, and ODRDCN-rEnKF reliably reconstructs both spatial and temporal source features. Compared to r-EnKF alone, the proposed framework significantly lowers computational costs while preserving high inversion accuracy, offering a promising approach for identifying non-point sources in groundwater contamination studies.