Geological model automatic reconstruction based on conditioning Wasserstein generative adversarial network with gradient penalty

Due to the structure complexity and heterogeneity of the geological models, it is difficult for traditional methods to characterize the corresponding anisotropic and structural features. Therefore, one of the generative models called Generative Adversarial Network (GAN) are introduced to the geological modeling fields, which describes the complex structural features effectively according to fitting the high-order statistical characteristics. However, the traditional GAN might cause gradient explosion or vanishment, insufficient model diversity, resulting the network cannot capture the spatial pattern and characteristics of geological models so that the reconstruction always has a bad performance. For this issue, this paper introduced the Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP), which can better measure the distribution discrepancy between the generative data and real data and provide a meaningful gradient for the training process. In addition, the gradient penalty term can make the objective function conform with Lipschitz constraints, which ensures the training process more stable and the correlation between the generative and real samples. Meanwhile, the conditioning loss function can make the reconstruction conform with the conditioning constraints. The 2D and 3D categorical facies model were introduced to perform experimental verification. The results show that the CWGAN-GP ensure the conditioning constraints and the reconstruction diversity simultaneously. In addition, for the network finished training, through inputting different kinds of conditioning data, a variety of stochastic simulation results can be generated, thereby realizing rapid and automatic geological model reconstruction.