Improving Heterogeneous Aquifer Characterization Using a Deep Learning‐Based Ensemble Smoother With a Hybrid Prior Strategy

Abstract Accurate characterization of aquifer hydraulic properties is crucial for groundwater resource management and pollution control. Most traditional inverse methods for aquifer characterization rely on unbiased prior assumptions to make robust estimation; however, real‐world applications frequently involve biased priors that can significantly compromise characterization accuracy. To mitigate this issue, this study introduces a hybrid prior strategy that combines multiple plausible prior assumptions to form a comprehensive prior. This strategy is implemented within a deep learning‐based ensemble smoother, termed ESDL, enabling robust estimation of unknown, heterogeneous aquifer hydraulic properties without relying on unbiased prior assumptions. We systematically evaluate the performance of ESDL under unbiased, biased, and hybrid priors across hydraulic conductivity fields characterized by log‐Gaussian, channelized, and three‐facies distributions. Using the Kalman‐based ensemble smoother (ESK) as a benchmark, results demonstrate that the hybrid prior strategy significantly enhances the robustness of aquifer characterization under biased prior assumptions, with ESDL consistently outperforming ESK in terms of both accuracy and stability. Notably, ESDL maintains high effectiveness even when the hybrid prior ensemble excludes unbiased samples, highlighting its adaptability and reliability in complex groundwater systems.