The intelligent inversion with model reparameterization of borehole and surface gravity data

Three-dimensional inversion of gravity data is the main approach for obtaining the density structure on the subsurface. Traditional inversion methods are relatively mature and reliable in this regard. Nowadays, the use of neural networks in the field of geophysics is becoming increasingly common. Differing from the conventional “data-driven” intelligent inversion, which establishes a direct mapping relationship from data to model by dataset training, we propose a “model reparameterization” inversion for gravity data. Outputted by a neural network, the density model is forward calculated to fit the observed gravity data, iteratively updating the network parameters by back-propagation of the data loss, without dataset training process. Essentially, this approach reparameterizes the model as network parameters, using the network to impose implicit constraints for the recovered model. The implicit constraints are incorporated by adding the network structures, including convolutional layers, activation functions, and up-sampling feature fusion. The effectiveness of this approach is demonstrated through inversion experiments on synthetic and real data sets.