Reconstruction and Software Implementation of Three-dimensional Porous Media Based on Improved InfoGAN
In the core 3D reconstruction,the existing deep learning methods will cause model collapse,resulting in low image diversity.To solve this problem,a 3D core reconstruction model based on improved InfoGAN is proposed.The model adopts unsupervised learning,the random noise in the generator is changed to Gaussian noise,and the original single discriminator is changed to double discriminator in the discriminator part.BCELOSS and MSELOSS are respectively used as the discriminators for training.The classifier classifies the core type on the basis of the discriminator,so that the subsequent reconstruction of different lithologies can be carried out.In the aspect of learning rate,dual time scale update rule is used,that is,learning rate of generator and discriminator is different.Finally,through porosity experiment and comparison with previous improvement and other 3D reconstruction algorithms,the digital core with more similar topological structure to the original core sample is constructed.
improve InfoGAN3D reconstructiondigital porous media
NETL
NSTL
万方数据
