Wasserstein cycle-consistent generative adversarial network for improved seismic impedance inversion: Example on 3D SEAM model
Date:
The convolutional neural networks (CNNs) have attracted great attentions in seismic exploration applications by their capability of learning the representations of data with multiple level of abstractions, given an adequate amount of labeled data. In seismic impedance inversion, however, the availability of labeled data-impedance pairs is often limited. The cycle-consistent generative adversarial networks (Cycle-GAN) is proven as a powerful semi-supervised learning solution by incorporating these unpaired data into its training, but it suffers from the training instability like most GAN algorithms. This study, starting from the recent progress in tackling such a challenge by using the Wasserstein GAN, proposes improving the Cycle-GAN algorithm by integrating the Wasserstein loss with gradient penalty loss as the target loss function, here denoted as the Wasserstein cycle-consistent GAN (WCycle-GAN). Correspondingly, the new algorithm benefits from both weaker topology in Wasserstein distance and better data regularizations in cycle-consistent loss, leading to enhanced training robustness and generalization abilities. We validate its performance through an example of impedance inversion on a subset of the 3D Seismic Advanced Modeling (SEAM) data. The results not only demonstrate the outperformance of the new algorithm over the conventional Cycle-GAN but also suggest its great potential for assisting other semi-supervised learning applications.