Data Space Inversion for efficient and uncertainty-informed contaminant source identification in aquifers

Groundwater supplies drinking water to more than one-third of the global population and is a critical resource for urban water security. Yet it is increasingly threatened by subsurface contamination, imposing substantial environmental, societal, and economic costs. In urban aquifers, identifying active contaminant sources and quantifying their mass-discharge rates is particularly challenging because plumes often overlap, subsurface properties are highly heterogeneous and poorly known, and monitoring data are sparse due to logistical and economic constraints. Under these conditions, conventional inverse modeling approaches are often computationally demanding and difficult to implement in high-dimensional systems for uncertainty-aware decision support. Here we address the challenge of efficient and reliable contaminant source identification under hydrogeological uncertainty. We show that integrating Data Space Inversion with an iterative Ensemble Smoother (DSI+iES) enables probabilistic discrimination of active and inactive contaminant sources and estimation of source mass discharge distributions, while achieving a computational gain of up to 15 relative to conventional parameter-space inversions. Using a synthetic heterogeneous aquifer with scarce observational data representative of an urban worst-case scenario, DSI+iES reproduces posterior source-plume behavior and uncertainty ranges consistent with those obtained with physics-based transport inversions. Its computational efficiency allows subsurface heterogeneity to be retained, uncertainty to be propagated through large ensembles, and systematic data-worth analyses to be performed. These analyses show that a small number of strategically placed monitoring wells in regions of high predictive uncertainty yield disproportionate information gains, whereas many intuitively placed wells contribute little. By coupling uncertainty-aware inference with high-dimensional system representation, the proposed workflow enhances the practical applicability of contaminant source identification. Although the current test case relies on simplifying assumptions relative to real-world systems, the proposed workflow provides a basis for future application to field-scale source identification and adaptive monitoring design in data-limited settings, where timely and cost-efficient decisions are required.