Research On Magnetotelluric Inversion Based On Deep Learning

Magnetotelluric exploration plays a very important role in geophysical exploration,which is a process of inferring the real underground resistivity model through surface observation data.At first,there were many traditional inversion methods,such as Gauss-Newton method,nonlinear conjugate method and marquardt method.These traditional inversion methods have the characteristics of large amount of calculation,long calculation time and slow convergence.In recent years,with the continuous development of artificial intelligence and deep learning,convolutional neural network inversion,deep confidence network inversion and residual network inversion have appeared.These intelligent inversion methods have the characteristics of fast inversion speed and high accuracy.Magnetotelluric inversion is a crucial step for magnetotelluric exploration,which is to learn the relationship between resistivity model parameters and observed data(apparent resistivity),and then solve the function in reverse according to prior knowledge,so the solution will not be unique.This paper mainly studies the problem of one-dimensional magnetotelluric inversion.The purpose is to obtain the resistance and thickness of the resistivity model through observation data(apparent resistivity).Through forward modeling,the data of the three-layer and four-layer model are mapped to the 40-layer homogeneous medium earth model with the thickness of 20 per layer.Among the obtained data,the apparent resistivity of the known model is used as the input of the network,and the model parameters of the resistivity are used as the output of the network.In order to make the inversion more effective,this paper chooses an 18-layer residual depth full convolution network,which solves the problem of gradient disappearance by adding quick connections during training,and consumes less time in the inversion process.According to the uniqueness of the inversion solution,this paper designs a series of experiments,which verify that the network has good training effect,strong generalization ability,good anti-S equivalence ability and anti-noise ability,and can solve the one-dimensional magnetotelluric inversion problem well,and is expected to solve the two-dimensional and three-dimensional magnetotelluric inversion problems.Magnetotelluric inversion based on deep neural network is a global search inversion method,which has the advantages of fast inversion speed,convenient operation and high inversion accuracy,far ahead of traditional inversion methods.The residual depth full convolution network selected in this paper solves the problem that the gradient disappears with the deepening of the network.