Gravity-Data Inversion for Basement Depth Using a Distribution-Constrained Diffusion Model and U-Net: A Case Study in the Pearl River Mouth Basin

Gravity-based sedimentary basement inversion is essential for basin analysis and hydrocarbon exploration, but it remains a non-unique problem due to the coupling between basement depth and residual density contrast in gravity anomalies. Deep learning is a promising approach, yet its generalization ability depends strongly on the geological realism and diversity of the training data. In this study, we propose a generation-to-inversion framework combining a Distribution-Constrained Diffusion Model (DCDM) and a U-Net-based inversion network. The DCDM was trained on 512 real sedimentary basin samples and incorporates Wasserstein-2 distribution alignment and boundary range regularization to generate geological synthetic basement depth models, with improved statistical consistency and physical plausibility. A total of 9488 DCDM-generated synthetic samples were combined with the 512 real samples to form 10,000 basement models. By assigning 41 residual density-contrast values and performing dual-interface gravity forward modeling, 410,000 gravity basement pairs were constructed for network training, validation, and testing. Synthetic experiments show that the proposed method achieves a depth RMSE of 0.200 km, a depth MAE of 0.094 km, and an equivalent residual density contrast RMSE of 0.015 g/cm3. Application to the Pearl River Mouth Basin produces basement depths consistent with independent constraints and seismic profiles, with an RMSE of 0.53 km at constraint points. The results demonstrate that the DCDM-generated basement depth samples improve the geological representativeness of the training data and support U-Net-based inversion of basement depth and equivalent residual density contrast.