Forward modeling guided deep learning for 3D gravity inversion

This paper presents a three-dimensional density inversion method for gravity anomalies utilizing deep neural networks under the constraint of the potential field equations. The potential field equation is added to the neural network loss function as a constraint term. To train the network, a labeled dataset is created by forward modeling the complex stochastic model. The outputs are rebuilt by forward modeling while minimizing the density model error and gravity data error. This improves the noise immunity and accuracy of the network with the same data. This also enhances the interpretability and physical constraints of the network. The model testing validates the correctness and effectiveness of the proposed method.