<p>Pinpointing the source of groundwater pollution and determining the hydraulic conductivity field present considerable challenges, as these can only be inferred through the analysis of limited observation data. This necessitates repeated forward numerical simulations during inversion, which is markedly time-intensive. In this study, a deep convolutional neural network (DCNN), bidirectional long short-term memory (BiLSTM), and an attention mechanism are leveraged to develop a surrogate model that approximates the numerical simulation model with high accuracy. The DCNN is employed to capture spatial features of both model and pollution source parameters (coordinates and release intensity), the outcomes of which are integrated into feature vectors. These vectors then serve as inputs to the BiLSTM for predicting corresponding groundwater concentration values, while the attention mechanism selectively emphasizes features with significant predictive influence, thereby refining accuracy. Subsequently, this surrogate model is integrated with the ensemble smoother with multiple data assimilation (ESMDA) to update the uncertain parameters. This methodology allows for the adjustment of individual grid values of the model and the parameters describing the pollution source, simultaneously providing a blueprint for constructing surrogate models of high-dimensional aquifer systems. Validation against both two- and three-dimensional case studies reveals the surrogate model’s high fidelity, notably maintaining precise predictions in three-dimensional scenarios. Ultimately, the application of this integrated inversion framework successfully identifies the hydraulic conductivity field and pollution source parameters with a high degree of congruence with observations. Moreover, adoption of the surrogate model reduces computational time by more than 30%, underlining the efficacy of the approach.</p>