Magnetic Inversion Based On Deep Learning

The earth,as the only home for human existence at present,is self-evident for human survival and reproduction.However,the earth is also a complex and changeable system.Throughout the ages,countless scholars have wanted to explore the mysteries behind the earth,which has derived related to the earth.Various research and applications.Geophysical inversion,as a practical research field,finds the deeper meaning behind it by observing the physical phenomena that appear on the earth itself.It has been widely concerned and developed vigorously.However,some problems such as ambiguity,initial model dependence,and long calculation time have always restricted the further development of this field.With the mature application of deep learning technology,as a novel and effective tool,it can well solve the existing problems that restrict the development of the field of geophysical inversion.On this basis,this article uses the inherent superior learning and mapping of deep learning Ability,proposed a Deep Neural Network(DNN)based magnetic anomaly data inversion method.This method first trains the neural network with a large amount of data obtained from the forward modeling,and then uses the trained network to directly predict the distribution of the underground medium.Although the training phase requires a lot of time,once the training is completed,it can significantly improve the subsequent inversion efficiency and quickly obtain Inversion results.In order to compare the effects of different deep learning network structures on the inversion results,this paper designs Fully Connect(FC)and Convolutional Neural Network(CNN)for magnetic inversion.The inversion results show that DNN has good nonlinear inversion capabilities,and its advantages such as strong learning ability,wide coverage,and strong adaptability make it irreplaceable in the field of geophysical exploration research.Therefore,this article conducts research according to the following content:First of all,this article explains the concepts and definitions of magnetic exploration and briefly describes its development history.It introduces traditional magnetic forward and inversion methods.At the same time,it explains and analyzes indepth learning in detail.The basic method of magnetic inversion based on deep learning.Secondly,enough training samples are needed to make the network reach a good state of convergence.In this paper,different types of physical models are synthesized,and then their physical properties are modeled forward.Sufficient data sets are obtained to train the DNN network so that the network can directly establish a nonlinear mapping from magnetic anomaly data to physical properties.The pre-trained network can be used to estimate the distribution of underground magnetization based on the new input magnetic anomaly data.Finally,by testing in a two-dimensional(2D)comprehensive example,two DNN structures(FC and CNN)are used to test the feasibility and applicability of the method.Compared with the conventional method,the predicted distribution of the magnetization obtained in a shorter time using this method is more concentrated,and it has better resolution to determine the boundary of the magnet.