A Random Medium Modeling Method Based on Wasserstein Convolutional Generative Adversarial Networks and Velocity Mask

Seismic forward modeling is crucial for exploration geophysics, especially in seismic exploration. As research has progressed, the limitations of forward models assuming homogeneous media have become increasingly evident. These models no longer adequately address the complexities of real-world scenarios. Furthermore, the multiscale heterogeneity of subsurface structures significantly influences exploration outcomes and geological interpretations. Therefore, constructing a random medium model that accurately reflects heterogeneity is necessary. However, the current modeling methods face issues of low efficiency and insufficient flexibility, making it difficult to construct complex random media models. Therefore, this article proposes a random medium modeling method that combines Wasserstein convolutional generative adversarial networks (WCGANs) and velocity masking. A training dataset is constructed through spectral decomposition, and the random media parameters are constrained into a composite conditional vector. The Wasserstein distance and gradient penalty are added to the loss function to train the networks. Also, the generator and discriminator are improved using convolutional network structures. After comprehensive evaluation through visual inspection, principal component analysis (PCA), and multiscale sliced Wasserstein distance (MS-SWD), the results show that this method improves training stability and the diversity of generated random perturbation models. Consequently, using WCGANs can improve modeling efficiency, while masking techniques can control the characteristics of different regions in the model, simulating complex geological structures. Therefore, by combining these two methods, it is possible to establish a regional multiscale random media model. This approach efficiently simulates subsurface heterogeneity, addressing issues of long modeling times and inflexibility.