3-D Gravity Data Inversion Based on Enhanced Dual U-Net Framework

Three-dimensional gravity inversion is an effective method for restoring underground density distribution from gravity anomaly data. Conventional regularization inversion has good data fitting, but its inversion model has insufficient model fitting capabilities due to its low-depth resolution. Although data-driven deep learning (DL)-based gravity inversion results significantly improve depth resolution and physical property distribution, it is difficult to ensure the data fitting of the inversion results. Accordingly, this study proposes a 3-D gravity data inversion based on enhanced dual U-Net (EdU-Net) framework to solve the above problems, making the inversion results have good model and data fitting performance. The proposed EdU-Net consists of two parts: first, training a large generalization pretrained network Net I and then quickly generating an enhanced Net II for the target data through fine-tuning. Additionally, this study adds forward-fitting constraints in the framework's loss function to reduce the problem of large data-fitting errors in traditional data-driven DL inversion. The trained Net II inversion result has better model and data fitting accuracy than Net I. Moreover, by comparing the inversion results of synthetic models, this study demonstrates that the EdU-Net method performs better than the traditional DL. Finally, this method is applied to the measured data of the Gonghe Basin in Qinghai Province, China, and provides a reasonable explanation for the distribution of hot dry rocks.