Synergizing Deep Learning and Full-Waveform Inversion: Bridging Data-Driven and Theory-Guided Approaches for Enhanced Seismic Imaging

TL;DR

This paper addresses the challenge of seismic inversion by integrating data-driven deep learning with physics-based full-waveform inversion (FWI). It reviews how DL can serve as data-driven predictors, initializers, and theory-guided operators to complement traditional FWI, where the objective is to minimize the misfit $J(m)=|d-F(m)|^2$ between observed data $d$ and forward-modelled data $F(m)$. The authors categorize approaches into data-driven velocity estimation, end-to-end learning, and physics-informed or hybrid methods, illustrating 3D real-data extensions and anisotropic settings. They outline key challenges—data quality, generalization, interpretability, and computational cost—and propose future directions such as uncertainty quantification, transfer learning, generative data synthesis, and multi-modal data fusion. Overall, the review suggests that the synergy between DL and FWI can yield faster, more robust, and higher-resolution subsurface characterizations with broad geophysical impact.

Abstract

This review explores the integration of deep learning (DL) with full-waveform inversion (FWI) for enhanced seismic imaging and subsurface characterization. It covers FWI and DL fundamentals, geophysical applications (velocity estimation, deconvolution, tomography), and challenges (model complexity, data quality). The review also outlines future research directions, including hybrid, generative, and physics-informed models for improved accuracy, efficiency, and reliability in subsurface property estimation. The synergy between DL and FWI has the potential to transform geophysics, providing new insights into Earth's subsurface.

Figures (11)

• Figure 1: First practical application of FWI using the Marmousi model. This shows significant improvements for the FWI results as presented by Bunks1995.
• Figure 2: Improvements in velocity model and pre-stack depth migrated images obtained through FWI over the Valhall field. The FWI updated velocity model demonstrated a network of shallow high-velocity channels and a gas-filled fracture extension from a gas cloud which was not previously identifiable in conventional tomography. The impact is evident in the migrated sections, which show more continues events in otherwise poorly illuminated area. Adapted from Sirgue2009 and Sirgue2010.
• Figure 3: Imaging improvements obtained through orthorhombic imaging. This produces sharp truncations and clearer faults as highlighted by the red dashed ovals, as well as better focussed gathers. Adapted from Xie2017.
• Figure 4: The Single Neuron Perceptron. The input values are multiplied by the weights. If the weighted sum of the product satisfies the activation function, the perceptron is activated and "fires" a signal. Adapted from Rosenblatt1958.
• Figure 5: The Mark I Perceptron Machine was the first machine used to implement the Perceptron algorithm. The machine was connected to a camera that used 20×20 array of cadmium sulphide photocells to produce a 400-pixel image. To the right is a patch-board that allowed experimentation with different combinations of input features. This was usually wired up randomly to demonstrate the ability of the perceptron to learn. Adapted from Hecht-Nielsen1990Bishop2006.