A Physics-Guided Deep Learning Workflow for Partial-Stack Seismic Inversion

Haibin Di

doi:10.1111/1365-2478.70101

A Physics-Guided Deep Learning Workflow for Partial-Stack Seismic Inversion

2025-11-04

0

PRE

AI

DOI：10.1111/1365-2478.70101

Original

Original request for help

Save

Abstract

En 中文

Pre- and post-stack seismic inversion is the primary approach for converting collected seismic data into geophysical property models particularly velocity for subsurface interpretation and reservoir characterization. The traditional workflows often start from an initial property model and iteratively revise it by minimizing the misfit between the acquired real seismic dataset and the synthetic one derived from the updated property models, here denoted as the soft constraint in seismic space. However, it heavily relies on human supervision in building a good initial model and monitoring the misfit optimization process. On the contrary, most of the recent deep learning-based workflows target at non-linearly mapping seismic patterns to properties measured at available well only, here denoted as the hard constraint in property space. Correspondingly, its accuracy greatly depends on the availability of sufficient training wells; otherwise, overfitting occurs causing the machine prediction not meeting the soft constraint throughout the target seismic survey. To resolve these limitations, this study presents a practical workflow that enables rock property inversion from partial-stack seismic via two physics-guided convolutional neural networks (CNNs), with the first one embedding the approximated AVO gradient to build initial property models that satisfy the hard constraint and the second one embedding the reflection coefficients to refine the models by enforcing the soft constraint. Between both CNNs is the use of well-established relevant physics to generate pseudo property-reflectivity-seismic pairs for training the second CNN. Its added values are validated through applications to the Volve survey in North Sea and the Exmouth survey in Western Australia. The produced property volumes not only are observed of high lateral consistency and vertical resolution but also derive synthetic seismic data that are closely correlated with actual seismic data.

Keywords:

WAVELET EXTRACTION

A Physics-Guided Deep Learning Workflow for Partial-Stack Seismic Inversion

Abstract

Journal

Organization

Cited Papers