Deep Learning Inversion of Shear Wave Velocity Profile for Pavement Using Frequency-Velocity Spectra

Summary This study explores the feasibility of using deep learning to invert shear wave velocity (Vs) profiles of pavement. Instead of relying on normal modes, the entire frequency-velocity spectrum (FVS) is used to train a deep learning model, incorporating actual receiver locations in surface wave testing. A simulated dataset for multichannel analysis of surface waves (MASW) is generated using efficient full-wavefield modeling with a fast Fourier-Bessel series and dual-variable matrix method. The model employs a residual network, a deep convolutional neural network, to map connections between input FVS and output Vs profiles using residual mapping. Once trained, the model can instantly predict Vs profiles. Performance is tested with three pavement cases: within the training data range, partially outside the range, and noisy high-frequency data. Predictions closely match ground truth for all cases, with calculated FVS aligning with true Vs profiles. The results show that deep learning inversion accurately and efficiently predicts both Vs and pavement thickness while addressing non-linearities. Additionally, it handles distorted or missing FVS data, demonstrating high flexibility and robustness for pavement inversion tasks.