Year: 2025
Type: article
Abstract: Solving the elastic or acoustic wave equation is essential for seismic imaging and inversion techniques. Although conventional methods, like finite difference or finite element schemes, are widely used, they suffer from low computational efficiency, especially in large‐scale applications. To overcome this limitation, we propose a novel deep learning‐based framework using Fourier neural operators (FNOs), which learn mappings from geological parameters to wavefield solutions. By integrating finite difference simulations with stochastic medium modelling, we generated training datasets encompassing diverse geological conditions. The neural operator was iteratively optimized through targeted training trials to enhance its predictive capability. The resulting operator achieves high accuracy ( L 2 error: 0.05–0.30) while preserving numerical fidelity comparable to traditional methods. More notably, the operator offers significant speedups, 170‐fold for acoustic and 260‐fold for elastic wave equations. Validated through comprehensive experiments, this operator serves as an efficient and reliable input for downstream seismic processing workflows, enabling end‐to‐end acceleration in seismic waveform inversion and imaging systems.
Source: Near Surface Geophysics
DOI: 10.1002/nsg.70033
Institution: China University of Geosciences
Language: English
Subfield: Geophysics
Field: Earth and Planetary Sciences
Domain: Physical Sciences
Open Access status: bronze
Awards: 近地表粘弹介质多分量表面波多目标全速度谱反演研究, G1844950806
OpenAlex ID: W4417449398