A Deep Learning-Based Forward Surrogate Model for Accelerating Surface-Wave Dispersion Inversion

Abstract We present a deep learning-based surrogate forward operator to accelerate 1D shear-wave velocity (VS) inversion from Rayleigh-wave dispersion. A gated recurrent unit-based surrogate forward model (FGRU) is trained to predict phase-velocity dispersion curves from the VS profiles. Training data are generated using a cubic B-spline crust–mantle parameterization with sufficiently broad variability to ensure accurate learning of the mapping. Benchmarks on a consumer-grade workstation show that FGRU reproduces the reference dispersion curves computed by a physics-based surface-wave dispersion solver accelerated using numba (DISBA) with high fidelity. In contrast, graphic processing unit (GPU)-accelerated batch inference enables forward evaluations that are orders of magnitude faster. When embedded in particle swarm optimization (FGRU–PSO), synthetic tests demonstrate that the surrogate forward operator accelerates the inversion by reducing the computational cost of large forward-model ensembles, while producing results comparable to those obtained with the conventional solver (DISBA–PSO). Application to South China phase-velocity data allows up to one million model evaluations per grid point at a manageable cost. Compared with previous tomographic models, the resulting 3D VS model not only recovers the major tectonic features but also reveals clearer shallow basin-related low-velocity anomalies, owing to a more thorough exploration of model space. These results demonstrate that the GPU-parallelized surrogate forward operator provides an efficient and practical route for accelerating regional surface-wave inversion.