In geophysical research, gravity-based inversion is essential for identifying geological anomalies, mapping rock structures, and extracting resources such as oil and minerals. Traditional gravity inversion methods, however, face challenges such as the volumetric effects of gravity fields and the management of large, complex matrices. Unsupervised learning techniques often struggle with overfitting and interpreting gravity data. This study explores the application of various U-Net-based network architectures in gravity inversion, each offering distinct challenges and advantages. Nested U-Net, although effective, requires a high parameter count, leading to extended training periods. Recurrent Residual U-Net's implicit attention mechanism restricts its dynamic adaptability, while Attention U-Net's lack of residual connections raises concerns about gradient issues. This research comprehensively analyzes the training processes, core functionalities, and module distribution of these networks, including Residual U-Net++. Our synthetic studies compare these networks with traditional focused regularized gravity inversion for reconstructing density anomalies. The results demonstrate that Nested U-Net closely approximates the actual model, despite some redundancy. Recurrent Residual U-Net shows improved alignment with minimal redundancies, and Attention U-Net is effective in density prediction but encounters difficulties in areas of low density. Notably, Residual U-Net++ excels in inversion modeling, achieving the lowest misfit percentage and accurately replicating density values. In practical applications, Residual U-Net++ impressively reconstructed the F2 salt diapir in the Nordkapp Basin with well-defined boundaries that closely match seismic data interpretations. These results underscore the capabilities of Residual U-Net++ in geophysical data analysis, structural reconstruction, and inversion, demonstrating its effectiveness in both simulated settings and real-world scenarios.
