Reconstruction of 3D digital rocks with controllable porosity using CVAE-GAN

Digital rock technology provides an effective approach for analyzing the pore structure and physical properties of rocks in geophysics and petroleum science. Although deep learning techniques have significantly improved the efficiency of digital rock reconstruction, existing works within this field often lack necessary constraints. Therefore, our study proposes a new method for digital rock reconstruction that employs a conditional variational auto-encoder generative adversarial network (CVAE-GAN). This approach integrates the strengths of a conditional variational auto-encoder (CVAE) and a conditional generative adversarial network (CGAN) to produce high-quality digital rocks. To enable the controllable porosity reconstruction of digital rocks, the proposed method includes porosity information within the encoder, decoder, and discriminator. Additionally, we used the Wasserstein GAN with gradient penalty (WGAN-GP) during the training process to enhance the stability of the neural networks. We conducted experimental evaluations using different types of samples to validate the effectiveness of our reconstruction approach. Calculations of pore structure parameters and simulations of rock physical properties were performed on the reconstructed digital rocks, indicating a robust correspondence with both the real samples and theoretical models. These results provide compelling evidence for the accuracy of the proposed method in digital rock reconstruction and suggest its promising prospects for investigating petrophysical properties.