Machine learning evaluation demonstrates rapid shear-wave velocity inversion from surface wave dispersion spectra, indicating that neural networks bypass initial model dependencies.
Key Points
Develop and evaluate INVNET, a convolutional neural network designed to invert near-surface shear-wave velocity structures directly from surface wave dispersion spectra without initial model dependencies.
Generated a synthetic training set of 20,000 sample pairs using layered velocity models with geological priors, Generalized Reflection/Transmission Coefficient forward modeling, and multi-depth noise source stacking.
Designed an encoder-style convolutional neural network architecture (INVNET) to extract deep features from dispersion spectra and map them to shear-wave velocity values.
Validated performance using synthetic test datasets and real-world seismic field data collected from the Qademah area.
INVNET achieved high inversion accuracy on synthetic data, executing each individual inversion in less than 1 s.
Field application in the Qademah area resolved a geologically consistent three-layer velocity structure that matched results obtained from traditional inversion techniques.