Geological Modeling Method of Generative Adversarial Networks Based on Multi-Scale Feature Fusion and Depthwise Separable Convolutions

Complex geological structure modeling is of significant importance in fields such as resources exploration, underground engineering design, and geological hazard prediction. Generative Adversarial Networks (GANs) have demonstrated strong nonlinear modeling capabilities and pattern transfer abilities in geological modeling. However, when dealing with complex geological constraints and the reconstruction of fine structures, they still face challenges in modeling accuracy, structural connectivity, and modeling efficiency. To address these issues, this paper proposes a GAN-based geological modeling method incorporating multi-scale feature fusion and deep separable convolutions. A multi-scale feature fusion module enhances the expression of geological structure details and overall consistency, while deep separable convolutions reduce model parameters and computational costs, improving modeling efficiency. Additionally, a conditional feature adaptive fusion and progressive resolution generation strategy enhances the model's sensitivity to conditional data. To validate the method's effectiveness, typical models including two-dimensional river phases, multi-attribute ice wedges, and three-dimensional fold structures were selected. Systematic evaluations were conducted across spatial variability, connectivity, attribute consistency, and conditional point reconstruction accuracy. Comparative analyses were performed against multi-point statistical methods (e.g., QS) and an improved generative adversarial network (e.g., CWGAN-GP). The results show that at resolutions of 64×64 and 64×64×64, the MS-SWD indicators of the generated models for the two-dimensional and three-dimensional datasets are 0.016, 0.025, 0.007 9, and 0.008 7 respectively, which are significantly lower than those of the comparison methods. At the same time, the average connected region size of the generated models is closest to that of the reference model (300.59 pixels for the two-dimensional river data and 17 814.17 pixels for the three-dimensional fold data). In terms of overall accuracy, the accuracy rate and MSE indicators of the proposed method are superior to those of the comparison method (73.24%, 69.48% and 0.024, 0.047 respectively), and the advantages in efficiency and parameter quantity are proved through efficiency analysis and ablation experiments. The experiments show that the proposed method is suitable for efficient modeling tasks of complex non-stationary geological bodies since it significantly improves the modeling efficiency while ensuring reasonable and high fidelity, endowed with broad engineering application prospects.