Simulating seismic data using generative adversarial networks

In this extended abstract, we describe a method to use deep learning to sample from an unknown random variable that describes the variability of seismic data. This method is training in an unsupervised manner using available seismic data. Using this method, we are able to produce realistic, randomly generated seismic data samples that contain a rich set of geological features as well as geophysical noise. This method can be used to produce large volumes of data to train modern machine learning techniques that often require very large amounts of data. Additionally, it can be used to statistically characterize the performance of machine learning algorithms.