Deep learning-based regularization of post-stack seismic inversion

Abstract

Seismic inversion is the prime method to estimate subsurface properties from seismic data. However, such inversion is a notoriously ill-posed inverse problem due to the band-limited and noisy nature of the data. Consequently, the data misfit term must be augmented with appropriate regularization that incorporates prior information about the sought-after solution. Conventionally, model-based regularization terms are problem-dependent and hand-crafted; this can limit the modeling capability of the inverse problem. Recently, a new framework has emerged under the name of Plug-and-Play (PnP) regularization, which suggests reinterpreting the effect of the regularizer as a denoising problem. Convolutional neural networks-based denoisers are state-of-the-art methods for image denoising: their adoption in the PnP framework has led to algorithms with improved capabilities over classical regularization in computer vision and medical imaging applications. In this work, we present a comparison between standard model-based and data-driven regularization techniques in post-stack seismic inversion and give some insights into the optimization and denoiser-related parameters tuning. The results on synthetic seismic data indicate that PnP regularization using a bias-free CNN-based denoiser with an additional noise map as input can outperform standard model-based methods.