Geophysical-guided Wasserstein cycle-consistent generative adversarial networks for seismic impedance inversion

Abstract

Deep learning has shown great ability to solve the nonlinear inversion problem in geophysical fields. Insufficient-labeled data and a lack of geophysical constraints become the main challenges in training the networks. In seismic impedance inversion, combined semisupervised learning and generative adversarial networks (GANs) named cycle-consistent GAN (cyc-GAN) are proven to achieve better inversion accuracy with insufficient labeled data. The next improvement of cyc-GAN is Geo-cyc-GAN, which imposes the convolutional model as a geophysical constraint. This improvement can speed up the training process and achieve better accuracy. However, like most GAN algorithms, the cyc-GAN and Geo-cyc-GAN suffer from training instability. Therefore, we proposed geophysical-guided Wasserstein cycle-consistent generative adversarial networks (Geo-cyc-WGANs) to overcome the training instability of GAN and increase its accuracy. Geo-cyc-WGAN uses Wasserstein distance as a loss function instead of cross-entropy to improve the training stability. The experiment results of synthetic data using small labeled traces show that Geo-cyc-WGAN achieves the highest accuracy, better lateral continuity, and a more stable training process than another geophysical guide-based method. The experiment results of real data also show that Geo-cyc-WGAN can obtain better accurate impedance results than other methods.