Abstract: Full waveform inversion (FWI) is a high-resolution seismic imaging technique that utilizes the full wavefield information contained in seismic records to reconstruct subsurface structures with high fidelity. Conventional FWI typically adopts L1 or L2​ norm–based objective functions, in which data misfit is measured by pointwise amplitude differences. However, such objectives are highly sensitive to the initial model and prone to local minima, thereby limiting inversion accuracy and stability. To enhance global convergence and robustness, this study introduces a Wasserstein-distance-based objective function. Unlike traditional norms, the Wasserstein distance measures discrepancies between distributions rather than individual samples, enabling improved resistance to noise and better handling of phase shifts and cycle-skipping. To further improve computational efficiency, Sinkhorn entropy regularization is incorporated into the Wasserstein formulation, significantly reducing the computational burden while preserving its essential physical meaning. In addition, automatic differentiation within the deep learning framework is employed to compute gradients efficiently and accurately, greatly enhancing the numerical stability and overall performance of the inversion process. Both theoretical analysis and numerical experiments demonstrate that the proposed Wasserstein–Sinkhorn FWI method can effectively mitigate inversion errors caused by inaccurate initial models. Compared with traditional L1/L2 ​-norm objectives, the proposed approach exhibits superior convergence behavior, robustness, and imaging quality, particularly in complex geological settings.