Shifting Paradigms: Data-Centric Approach for Marine Statics Correction using Symmetric Autoencoding

Abstract

Deep learning has demonstrated remarkable performance in a wide variety of domains and is often leveraged for making high-stakes decisions. Parallel to its growing and beneficial presence in other domains, deep learning is gaining a notable reputation for solving challenging problems in geophysics. A key problem - given the escalating energy and geosequestration demands in present times - is marine statics correction. The traditional workflow for correcting marine statics has been based on a model-centric paradigm. This paradigm involves a series of transformations between non-commensurate spaces: first, inversion from seismic data space to velocity model space and second, forward modeling from velocity model space to seismic data space. Statics correction within this paradigm has severe drawbacks, mainly the high compute, time and labor cost, and inaccuracies stemming from errors in velocity model inversion or from unmet assumptions about subsurface structure. Overcoming these drawbacks was thus, the prime motivation for our study - where we chose to leverage deep learning as the core algorithmic tool to understand the limits of the model-centric paradigm and explore the performance horizons of a different, data-centric, paradigm to statics correction. The main feature of the data-centric paradigm is the direct mapping between commensurate data spaces, eliminating the need for intermediary transformations to and from velocity model space. Initial benchmark tests on the model-centric approach revealed the impact of inaccuracies in velocity model inversion as substantial nonzero timeshifts - exceeding 0.01s, and reaching values as large as 0.04s - for most arrivals in seismic data. These arrival time precision levels are unacceptable for good seismic imaging and time-lapse analysis; underscoring the need for an improved approach to marine statics correction. Consequently, we began our investigations into the data-centric paradigm. With the focus of disentangling the effects of varying seawater velocity from coherent subsurface geology in seismic records, we implemented an autoencoder algorithm, named SymAE. Notably, SymAE leverages the permutation symmetry of coherent subsurface information to perform the separation of information from nuisance variations. Once trained, SymAE is able to redatum selected subsurface and water velocity information in its latent space to produce statics-corrected seismic records. Our results show that for training datasets of increasing subsurface complexity, SymAE strongly converges all dynamic timeshifts to zero, aligning perturbed traces to reference traces. Crucially, SymAE delivers the required timeshift precision of 0.01 seconds for all arrivals - an achievement that the model-centric approach falls short of. This notable precision improvement using SymAE highlights how a streamlined data-centric paradigm outperforms the traditional model-centric paradigm of marine statics correction. This finding is pivotal as it is the foundation that lays the groundwork and opens the path towards the real-world deployment of SymAE for statics correction in challenging deepwater environments.