Deep Learning in Seismic Inverse Problems with Recurrent Inference Machines

Machine learning approaches are rapidly finding their way into many applications in processing and imaging seismic data. More specifically, various convolutional deep-learning architectures are currently being explored for seismic data processing tasks from denoising to imaging. Here, we present Recurrent Inference Machines (RIMs): a recurrent network architecture designed specifically for inverse problems, where a known forward operator is known and used as a constraint. We describe how both the original RIM and its invertible counterpart (iRIM) are designed to mimic gradient-based optimisation methods, and thus learn to perform data-driven regularisation and implicit model shaping due to their deep learning nature. We show examples of using RIMs to perform seismic data interpolation and image-domain inversion by deblurring, benchmarking them against UNets as a more widely-used deep learning architecture. Our examples show that RIMs outperform UNets particularly in dealing with features not necessarily present in the training data, due to the role of the forward operator as an additional constraint in training.