ResDM-Net++: An Enhanced Diffusion Model for Gravity Inversion

In recent years, deep learning techniques have increasingly been applied to gravity inverse problems, and ResU-Net++ has demonstrated significant success in density-detail preservation during image segmentation and feature extraction. However, its large parameter space and high computational demand make it prone to overfitting. Diffusion models, conversely, excel at generating detailed samples and handling noisy data but generally underperform in segmentation tasks compared to ResU-Net++. To address these challenges, we propose ResDM-Net++, a novel framework that integrates the essential modules of ResU-Net++ into a diffusion model while embedding geophysical insights throughout. Specifically, we employ multi-channel gravity inputs obtained from physically forward-modeled density distributions to ensure that both localized and global subsurface features are learned. Moreover, rather than applying a generic denoising approach, the diffusion component is carefully adapted to retain the inherent spatial correlations of geophysical data, thus improving inversion stability and mitigating overfitting risks. Finally, ResDM-Net++ further embeds physics-based constraints into both the encoder-decoder path and the diffusion steps, forging a synergy between robust denoising and domain-focused feature extraction. Numerical study shows that ResDM-Net++ accurately recovers subsurface density anomalies, exhibiting clear model boundaries and minimal fitting errors. In field applications, it successfully delineates the F2 salt dome in Norway’s Nordkapp Basin with boundaries closely matching seismic interpretations, underscoring ResDM-Net++’s effectiveness in gravity data analysis, structural reconstruction, and inversion in both synthetic and real-world scenarios.