Inversion of magnetic data using deep neural networks

As a novel tool, deep learning is used to solve complex problems in the real world and has been successfully applied to invert for seismic and electromagnetic data. In this study, we propose to use deep neural networks (DNNs) to recover the distribution of the physical properties of buried magnetic orebodies from the surface and airborne magnetic anomaly data. This approach is based on data training instead of prior-knowledge assumptions used in traditional inversion methods. Once our generalized network is established, the computing time to predict new magnetic data using this method can be significantly reduced. By implementing the forward modeling of different types of synthetic physical property models, we obtained enough datasets to train a DNN model so that the network can establish a nonlinear mapping directly from magnetic anomaly data to physical properties. The pre-trained network can be used to estimate the distribution of magnetization intensity from new input magnetic data. Two DNN structures were employed to test the feasibility and generalization of the proposed method by implementing Experiments in serval two-dimensional (2D) synthetic examples. Compared with the conventional method, the predicted distribution of magnetization intensity obtained by using our method is more concentrated and has better resolution to determine the boundary of the magnetic body. In a field example of Galinge iron ore deposits in China, the magnetization distribution of concealed iron orebodies inverted by the proposed approach was in good agreement with that detected by borehole data. DNNs have good nonlinear inversion capability and exhibit some excellent merits of strong learning ability, wide coverage and strong adaptability, which is a considerable application prospect in geophysical exploration.