Michael Warner

Many geophysical tasks are hindered in practice by the high costs of generating and processing data. We have developed a potential solution, mitigating the cost of certain data acquisition and generation processes by using data-domain-translation deep neural networks. Generative adversarial networks have demonstrated success in data translation in a wide variety of applications. By providing training data from domain A and domain B, networks can be trained to estimate the distributions of both domains, and hence establish a mapping from one to the other. We apply such data-translation neural networks to 3D geophysical field data examples and determine that they can be used as cost-reduction tools, providing an efficient mapping between different data types of interest in expensive data-processing workflows. Our approach is especially relevant for the translation between acoustic and elastic data sets during full-waveform inversion, which mitigates the elastic effect in the acoustic inversion.