Seismic wavefield simulation and inversion for variable velocity models via Gabor-Filtered FNO

Seismic wavefield simulation is fundamental to exploration geophysics, but it remains computationally demanding. Conventional numerical solvers face inherent trade-offs between accuracy and efficiency. Deep learning approaches has been effectively adopted in seismic wavefield simulation, but they are often limited by poor generalization to unseen geological configurations and inadequate representation of fine-scale wavefield details. We develop a modified Gabor-filtered Fourier Neural Operator (GFNO) for time-domain seismic modeling that addresses these limitations through three key innovations: learning mappings in infinite-dimensional function space to enable generalization across diverse velocity models, a causally consistent time-stepping scheme that respects the temporal evolution of wave propagation, and dense connections between Gabor Fourier layers that preserve local details during wave propagation. We evaluate the proposed framework on elastic wave propagation using multiple velocity families from the OpenFWI benchmark, as well as the 2D modified Marmousi models. For forward modeling, the GFNO accurately predicts wavefield evolution and generalizes effectively to unseen velocity models. Comparative experiments demonstrate that the modified GFNO outperforms the standard Fourier Neural Operator (FNO). For seismic inversion, the network establishes nonlinear relationships between seismic data and velocity models, achieving accurate velocity reconstruction. The successful transfer to the Marmousi models underscores the framework’s potential for real-world exploration scenarios. Our researches establish neural operator-based methods as scalable and generalizable tools for time-domain seismic modeling and inversion, with potential applications in rapid velocity model building.