Learning-Driven Optimization Framework for Viscoacoustic Full-Waveform Inversion

Ning Li

doi:10.1109/tgrs.2026.3728102

Learning-Driven Optimization Framework for Viscoacoustic Full-Waveform Inversion

2026-08-27

0

PRE

AI

DOI：10.1109/tgrs.2026.3728102

Original

Original request for help

Save

Abstract

En 中文

Full waveform inversion (FWI) is a high-resolution imaging technique that reconstructs subsurface medium parameters by fully exploiting seismic wavefield information and has significant applications in physical parameter inversion. However, in viscoacoustic media, multiparameter FWI suffers from strong coupling between velocity and the quality factor (Q), while the sensitivity of seismic data to Q is relatively weak. These issues often lead to severe parameter crosstalk, reduced convergence stability, and limited inversion accuracy. To address these challenges, we propose a convolutional implicit regularization-based Q full waveform inversion method (CIR-QFWI). In this approach, an unsupervised convolutional neural network driven by the deep image prior (DIP) is embedded into the FWI iterative framework. The inherent prior imposed by the network architecture acts as an implicit regularization, guiding the velocity and Q models toward geologically plausible structures and effectively mitigating parameter coupling and inversion artifacts. Meanwhile, to avoid the complexity of gradient derivation and implementation associated with the conventional adjoint-state method in viscoacoustic multiparameter inversion, we develop a unified gradient computation framework based on automatic differentiation. By leveraging computational graphs and the chain rule, gradients are obtained automatically, significantly simplifying the implementation. To further alleviate the high memory consumption of AD in long time-step wavefield propagation, a checkpointing strategy is introduced to balance computational efficiency and memory usage. Numerical experiments demonstrate that, compared with conventional FWI methods, the proposed CIR-QFWI achieves higher-resolution reconstructions of both velocity and Q models, validating its effectiveness and potential for multiparameter inversion in viscoacoustic media.

Keywords:

Full waveform inversion (FWI)

viscoacoustic media

quality factor (Q)

deep image prior (DIP)

automatic differentiation (AD)

Learning-Driven Optimization Framework for Viscoacoustic Full-Waveform Inversion

Abstract

Journal

Organization

Cited Papers