Deep‐learning inversion of near‐surface S‐wave velocity using fundamental and higher mode Love waves

Surface‐wave dispersion inversion has found broad applications in near‐surface shear‐wave velocity ( V S ) reconstruction. However, the main research focuses on Rayleigh waves. Moreover, traditional inversion methods are often hindered by non‐uniqueness, dependence on prior assumptions and the difficulty of effectively incorporating multimode dispersion information. Recent advances in artificial intelligence have opened new possibilities for data‐driven surface‐wave inversion, but previous deep‐learning studies have largely focused on fundamental‐mode Rayleigh‐wave dispersion. We represent the first attempt to incorporate fundamental and higher mode Love‐wave dispersion information into a neural‐network‐based inversion framework for near‐surface V S reconstruction. A total of 30,100 synthetic velocity models were generated on the basis of statistically constrained layering patterns derived from field observations, and their corresponding dispersion curves were computed using the propagator‐matrix method. These datasets were used to train two one‐dimensional (1D) U‐Net models: a fundamental‐mode model and a multi‐mode model that incorporates higher mode dispersion as an additional input channel. Quantitative evaluations using the root‐mean‐square error, mean absolute error and Pearson correlation coefficient show that the multi‐mode model achieves substantially lower errors and a notably improved recovery of deeper velocity variations. This result demonstrates the added value of incorporating higher mode information. Model applicability was further examined using Love‐wave data acquired at the Boise Hydrogeophysical Research Site. After dispersion extraction, interpolation and normalization, both models successfully reproduced stratigraphic features consistent with borehole constraints, whereas the multi‐mode model captured subtle deep‐layer variations that were not resolved by the fundamental‐mode model. Moreover, theoretical dispersion curves derived from the predicted S‐wave velocity profiles show good agreement with the measured dispersion curves, confirming the physical plausibility of the inversion results. Overall, the results demonstrate that integrating higher mode Love‐wave information enhanced constraints on subsurface structures. As a first neural‐network‐based implementation of multimode Love‐wave inversion, the study offers a robust and automated approach for high‐resolution near‐surface velocity reconstruction.