Abstract: Full waveform inversion (FWI) is an imaging technique that uses the full-waveform information contained in seismic records to characterize subsurface structures with high resolution. Conventional FWI commonly adopts the L1 or L2 norm as the objective function, in which the data misfit is measured by pointwise differences between observed and simulated seismic records. However, such objective functions are highly dependent on the initial model and are prone to convergence toward local minima, thereby limiting the accuracy and stability of inversion results. To enhance the global optimization capability and robustness of FWI, this study introduces an objective function based on the Wasserstein distance. Unlike traditional norm-based measures, the Wasserstein distance characterizes the global discrepancy between data distributions rather than pointwise errors, leading to improved global convergence behavior and noise robustness during inversion. To further improve computational efficiency, Sinkhorn regularization, i.e., entropy-regularized optimal transport, is incorporated into the Wasserstein framework, reducing computational complexity while preserving its physical meaning. In addition, automatic differentiation within a deep learning framework is employed to achieve efficient and accurate gradient computation, thereby improving the computational efficiency and numerical stability of the proposed algorithm. Theoretical analysis and numerical experiments demonstrate that the proposed method can effectively reduce inversion errors caused by inaccuracies in the initial model for complex geological models. Compared with conventional L1 and L2 norm-based methods, the proposed method exhibits significant advantages in terms of inversion accuracy and stability.