One-dimensional shear-wave velocity profile inversion using deep learning guided by wave physics

Obtaining near-surface shear-wave velocity (V_s) profiles is essential for advancing the understanding of site effects, thereby playing a pivotal role in both the assessment and mitigation of seismic hazards induced by these effects. Numerous inversion methods have been proposed for near-surface V_s profile inversion, utilizing measurements from either single station or multiple stations. However, these methods are often sensitive to initial profiles and exhibit slow convergence in the absence of appropriate initial profiles. To address these issues, we propose a novel inversion method based on physics-guided neural network. The network structure is designed according to the theory of the frequency domain method, and a physics-constrained loss function is introduced to avoid solutions that violate physical laws or empirical constraints, thereby enhancing the well-posedness of inversion problems. Both synthetic and real downhole array signals are employed to evaluate the performance of the proposed method. The results demonstrate that the proposed method exhibits robustness to noise and initial V_s profiles. Furthermore, comparative experiments with established techniques demonstrate that the proposed method not only produces more reliable V_s profiles by utilizing downhole array signals but also achieves higher computational efficiency.