For reservoir structural models with obvious nonstationary and heterogeneous characteristics, traditional geostatistical simulation methods tend to produce suboptimal results. Additionally, these methods are computationally resource-intensive in consecutive simulation processes. Thanks to the feature extraction capability of deep learning, the generative adversarial network-based method can overcome the limitations of geostatistical simulation and effectively portray the structural attributes of the reservoir models. However, the fixed receptive fields may restrict the extraction of local geospatial multiscale features, while the gradient anomalies and mode collapse during the training process can cause poor reconstruction. Moreover, the sparsely distributed conditioning data lead to possible noise and artifacts in the simulation results due to its weak constraint ability. Therefore, this paper proposes an improved conditioning spectral normalization generation adversarial network framework (CSNGAN-ASPP) to achieve efficient and automatic reconstruction of reservoir geological bodies under sparse hard data constraints. Specifically, CSNGAN-ASPP features an encoder-decoder type generator with an atrous spatial pyramid pooling (ASPP) structure, which effectively identifies and extracts multi-scale geological features. A spectral normalization strategy is integrated into the discriminator to enhance the network stability. Attention mechanisms are incorporated to focus on the critical features. In addition, a joint loss function is defined to optimize the network parameters and thereby ensure the realism and accuracy of the simulation results. Three types of reservoir model were introduced to validate the reconstruction performance of CSNGAN-ASPP. The results show that they not only accurately conform to conditioning data constraints but also closely match the reference model in terms of spatial variance, channel connectivity, and facies attribute distribution. For the trained CSNGAN-ASPP, multiple corresponding simulation results can be obtained quickly through inputting conditioning data, thus achieving efficient and automatic reservoir geological model reconstruction.
