Real-time seismic attributes computation with conditional GANs

In this work, we propose a deep learning pipeline for training neural networks that can accurately approximate attributes in three-dimensional seismic datasets. Using generative adversarial modeling, we train specialized networks that learn to map attributes, given an amplitude seismic volume as input. The trained network is used to compute transformations in original data much faster than its exact formulation since inference time is rapid in modern GPU architectures. Initial results show that deep neural networks are robust to learn different attributes with distinct data distributions. Via model inference, attribute computations is up to 80x faster than classical formulation.