Experiments on constructing seismic using generative adversarial network

Zhan Niu, Daniel O. Trad

Supervised machine learning attracts great attention in all areas of science. In Geophysics, however, supervised learning has the problem that available labelled data is often insufficient, limiting the chance of converging during training and harming model generality. As a solution, researchers explore ways to generate synthetic data for use in training. In this report, we explore the methodology of generating 1D data with a generative adversarial model. Both the generator and discriminator are convolutional, and the noise vectors are fed along the channel dimension to the generator. The networks are successfully trained via Wasserstein loss with gradient penalty and careful hypermeter tuning. We evaluate the trained networks quantitatively and qualitatively. We attempt to find the optimal stopping point for the training, however, the conclusion cannot be made during the training and part of it remains subjective.