A Deep Learning Framework for Joint Inversion of Gravity and Magnetic Data

Data-driven deep learning inversion of gravity and magnetic data is an emerging technique in obtaining subsurface density and magnetization source distributions. However, the absence of geophysical constraints, inflexibility of structural coupling and oversimplified features of synthetic data restrict the data-driven deep learning joint inversion, outputting models inconsistent with geophysical observations and geological priors. We propose a physics-informed deep learning framework for joint inversion of gravity and magnetic data. The data-driven pre-training is initially utilized by conducting end-to-end supervised training, learning the synthetic features from training dataset. With inverted density and magnetization distributions using pre-trained network, the data misfit and structural losses are calculated for physics-informed fine-tuning of the network. The embedment of physics-informed fine-tuning optimizes data-driven pre-trained network while retaining swift model reconstruction ability, generating models with improved data fitting and model reconstruction of the consistent source regions. The proposed framework is tested on two sets of synthetic examples with different structural homologies and applied to the field data of the Jining iron deposit (northern China). The joint inversion generates density and magnetization distributions for hematite and magnetite and indicates the possibile presence of a regional magnetic basement caused by the high-susceptibility amphibole magnetite quartzite in the Taishan Group. The proposed physics-informed deep learning framework for joint inversion demonstrates the potential of integrating multiple geophysical data and enhances the geophysical consistency in geological modeling.