Cross‐Dimensional Generative Adversarial Networks (CDGAN) Geomodeling: Bridging 2D Geological Figures and 3D Reservoir Modeling

Generative adversarial networks (GANs) have proven effective in simulating complex reservoir environments, such as meandering channels and deltas. In classic GANs, the dimensionality of training data determines that of generated data: a 2D (or 3D) reservoir facies simulator (generator) requires training with corresponding 2D (or 3D) data sets. However, most available 3D geological data are scarce, limiting their practical application. In contrast, 2D geological data, such as outcrop profiles and satellite images, which are relatively accessible, offer rich details and wide coverage. Thus, we propose CDGAN, a cross‐dimensional GAN‐based framework for generating 3D geological models from 2D geological figures. Its architecture integrates a 3D generator, a slicer, and multiple 2D discriminators to resolve the dimensional incompatibility between 2D training data and 3D generated models. Specifically, to ensure training robustness, a multi‐discriminator strategy mitigates directional pattern interference, while progressive growing technology enhances training stability and multi‐scale feature representation. When additional data (e.g., well‐interpreted facies data and probability volumes) are incorporated into the CDGAN architecture for training, the generation results can be controlled to achieve data conditioning. CDGAN was validated on two synthetic cases and two real‐field cases. In theoretical tests, CDGAN results more closely resemble the reference data than those generated by SliceGAN, as demonstrated by comparisons of geological morphology, proportions of different facies types, multi‐scale sliced Wasserstein distance (SWD) and multi‐dimensional scaling (MDS). Conditional CDGAN achieved a match rate exceeding 85% with 1D hard well data and was effectively constrained by low‐resolution 3D sand probability volumes. In real applications, 2D expert‐interpreted facies planes and profiles enabled the pre‐trained CDGAN generator to generate diverse 3D reservoir types that align well with known geological patterns, well data, and sand probability distributions across different layers. This approach bridges the longstanding gap between traditional 2D reservoir architecture analysis and 3D reservoir modeling. By enabling reliable 3D geological reconstruction from widely available 2D data, and supporting conditional constraints from well data and probability volumes, CDGAN provides a robust, convenient, and trusted data‐driven solution for practical deep learning‐based geological modeling.