Normalizing flows for regularization of 3D seismic inverse problems

We present the first known exploration of a normalizing flow (NF) for generative 3D volumes. First, we tackle computational issues surrounding the high dimensionality of our desired 3D volume output. This is of particular concern in normalizing flows since their invertibility constraint implies equal dimension of output and input. Our findings show that by “freezing” expensive layers we can efficiently train a normalizing flow on 3D volumes. Using this NF architecture, we train a generative model on volume sections of the 3D BG compass model. Our method produces visually plausible generative samples which are efficient to produce. We demonstrate its practical use by using our trained generative model as an implicit prior in a Maximum A Posteriori (MAP) framework. We evaluate this MAP framework by estimating the solution of a inverse problem in seismic imaging. Our method results in higher SNR estimates than the baseline and in less iterations, importantly saving the computational cost of evaluating the expensive 3D PDE solver during optimization. Finally, through scaling analysis of training cost, we show that NF convolutional layers allow this approach to scale favorably to larger volumes.