<p>Accurate inversion of heterogeneous hydrogeological parameter fields is essential for reliable groundwater numerical modeling, as it directly determines the credibility and predictive capability of simulation results. However, the strong spatial variability of such parameter fields often leads to highly nonlinear and high-dimensional inversion problems, resulting in extremely complex computations and substantial computational burden. To address this issue, we propose a novel inversion framework that integrates a Convolutional Adversarial Autoencoder (CAAE), a deep Capsule Encoder-Decoder (CapsED) network, and the Iterative Local Updating Ensemble Smoother (ILUES) algorithm for efficient and accurate estimation of heterogeneous hydrogeological parameter fields. In this study, we apply the proposed inversion framework to estimate the heterogeneous hydrogeological parameter fields in the groundwater source area of Tongliao, Inner Mongolia. The results show that the CAAE reduces the dimensionality of hydraulic conductivity fields from 6710 to 448 parameters, effectively alleviating the curse of dimensionality in the inversion process. Compared with the U-Net surrogate model, the CapsED surrogate model performed better in terms of R², error distribution, and visual similarity of predicted hydraulic heads, demonstrating a stronger capability to capture the complex nonlinear mapping between inputs and outputs, as well as higher training efficiency. In addition, the framework reduced computational cost by approximately 80.5% while maintaining satisfactory inversion accuracy, thereby markedly improving computational efficiency. These results demonstrate that the proposed inversion framework can effectively and efficiently estimate heterogeneous hydrogeological parameter fields, and that it exhibits good reliability and applicability.</p>