Forward constrained 3D gravity density inversion based on EdU-Net with well constraints

With the increasing complexity of exploration objects, data-driven deep learning gravity inversion methods alone can no longer meet the accuracy requirements for deep exploration of oil and gas reservoirs. In order to further improve the accuracy of gravity data processing, the use of well constraints to incorporate rich well data into the gravity data processing process has become a development trend. This article incorporates well constraints into enhanced dual U-Net (EdU-Net, Dong et al., 2023), resulting in inversion results with good depth resolution and forward fitting, as well as good agreement with logging data. Similar to EdU-Net, the network consists of two parts: first, we train a large-scale generalized pre-training network Net-I, and then on the basis of it, we take a part of the original label as well constraint, let it enter the network as an additional label, and quickly generate an enhanced Net-II for target data through fine-tuning. The optimized Net-II has more realistic inversion results at the drilling position. Then, by comparing the inversion results of the simulation model, we demonstrate that this method has better performance. Finally, we attempt to apply this method to the measured data in the Gonghe Basin of Qinghai Province, China, and obtained density inversion results that were consistent with the local logging information obtained.