Research on a Deep Learning‐Based Inversion Method for Surface Wave Dispersion Spectra

Abstract The near‐surface shear‐wave velocity structure is crucial for geological engineering and near‐surface exploration. Surface wave inversion is a key technique for obtaining the velocity structure, but traditional methods suffer from limitations such as heavy reliance on the initial model, high computational cost, and susceptibility to local minima. This study proposes INVNET, a deep learning model based on Convolutional Neural Networks, for efficient inversion from surface wave dispersion spectra to the Versus structure. We first constructed a large‐scale synthetic data set by randomly generating layered velocity models conforming to geological priors and performing forward modeling using the Generalized Reflection/Transmission Coefficient (GRTC) method. To enhance data realism, we introduced a multi‐depth noise source weighted stacking strategy, generating 20000 sample pairs of noise‐rich dispersion spectra and their corresponding velocity models. INVNET adopts an encoder style architecture that can automatically extract deep features from dispersion spectra and map them to velocity values. Testing on synthetic data shows that INVNET achieves high inversion accuracy, with a single inversion taking less than 1 s. In a practical application to field data from the Qademah area, INVNET rapidly inverted a reasonable three‐layer velocity structure, consistent with results from conventional methods. This result demonstrates the model's potential for real‐world applications. This research demonstrates that deep learning provides a novel solution for surface wave inversion that is independent of initial models, offers high accuracy, and is fast.