Research On A Method Of 2D Joint Invertion Of Gravity And Magnetic Data Constrained By Deep Neural Network

Deep learning is constructed by multi-layer neural networks.Deep learning has the ability to extract the distribution features of input data,learning ability,and generalization ability through the powerful computing power of computers.In recent years,many excellent network structures and algorithms have emerged in deep learning,which have been widely applied in geophysical data inversion.Joint inversion of multiple geophysical data can improve resolution of exploration significantly and overcome the uncertainty problems of a single-property inversion method and the inconsistency problems of a single-parameter inversion model effectively.Joint inversion mainly includes two categories: the petrophysical coupling and structural similarity coupling,which have been widely applied.In order to combine the advantages of physical and structural coupling,this paper proposes a method of 2D joint inversion of gravity and magnetic data constrained by deep neural networks.In this method,a convolutional neural network is designed initially,and then trained through a large number of datasets reflecting the physical and structural coupling relationships,so that the network can simultaneously learn the physical and structural coupling relationships between gravity and magnetic models,Finally,the trained network is used to achieve a "soft" mapping between gravity and magnetic models,providing an accurate reference model for joint inversion and improving the effectiveness of joint inversion.The model trial results show that the method can balance the physical and structural coupling relationships between gravity and magnetic models.Compared with the joint inversion of cross gradient and the separate inversion of gravity and magnetic data,inversion results of the deep learning coupling method proposed in this paper are more accurate in their positions,more focused,and faster in the convergence speed.This provides a new approach for the application of deep learning in the joint inversion.