Enabling deep-learning-based uncertainty quantification at scale for post-stack UHR seismicinversion

SUMMARY

An accurate characterization of the shallow subsurface is crucial for evaluating the load-bearing capacity at offshore wind farm sites.

It is widely recognized that seismic methods, particularly inversion, can significantly enhance traditional geotechnical approaches with three-dimensional information that is otherwise unavailable. A comprehensive geotechnical analysis, however, relies on several scenarios with different confidence levels to establish robust and trustworthy geotechnical parameter ranges.

Deterministic seismic inversion is inherently ill-suited for this purpose because it provides only the most likely subsurface characterization. Here, we discuss a variational-inferencebased Bayesian framework for UHR post-stack inversion, explicitly characterizing the uncertainties in acoustic impedance estimates.

A notable feature of this approach is the ability to express these uncertainties in terms of credible intervals with arbitrary credibility. Our methodology leverages advanced deep-learning techniques, such as generative modeling via normalizing flows, to enable efficient and scalable implementations.