AEDA: an adversarial architecture for deep surrogates applied to uncertainty quantification in seismic imaging

Abstract We introduce a new deep neural network architecture as a surrogate for producing seismic images through Reverse Time Migration under uncertainty. The novelty here lies in employing an adversarial architecture, where the generator’s core is an encoder-decoder neural network that produces seismic images conditioned on the velocity fields. Such an adversarial training approach aims to extend the applicability of the encoder-decoder surrogate model for velocity fields with high dimensionality, which acts as a surrogate model to enable uncertainty quantification in Reverse Time Migration. Also, we propose an a-priori assessment of the impact of epistemic uncertainties in seismic images due to the use of deep-learning surrogate models. We demonstrate, through numerical experimentation using two geological scenarios, that the novel training approach can enhance the accuracy of seismic images at minimal cost. Most importantly, the novel training approach can replace the costly RTM imaging method, making many-query tasks like sensitivity analysis, design, and optimization viable, in addition to uncertainty quantification.