Forward And Inversion Of Magnetotelluric Based On Deep Learning

Magnetotelluric(MT)is a geophysical electromagnetic exploration method,which can infer underground resistivity by measuring the natural geomagnetism and geoelectric field changes on the earth's surface.The detection depth ranges from tens of meters underground to tens of kilometers or more,and has the characteristics of long-period sounding.MT appeared in the 1940 s and gradually developed.Now it has become an international frontier subject.It is widely used in resource exploration(oil,natural gas,geothermal and solid minerals),groundwater monitoring,deep crust and mantle detection,and earthquake precursor prediction.In this paper,the development of magnetotelluric exploration is briefly reviewed,and the current popular forward and inverse methods and their advantages and disadvantages are introduced.In order to overcome the shortcomings of conventional linear inversion methods,this paper introduces a deep learning method based on neural networks to perform magnetotelluric inversion.In order to construct a large number of data for neural network training and learning,to extract the characteristics of different models and related characteristics of labels,this paper uses the finite element method for forward simulation.In order to simulate arbitrarily complex geoelectric structures,and uses unstructured meshing to simulate arbitrarily complex geoelectric structure,the data set used for deep learning is generated through numerical simulation.Secondly,explain the relevant knowledge of multilayer perceptron and convolutional neural networks,and the deep learning algorithm based on Tensor Flow framework is established.The activation function is used to enhance the expression ability of the entire neural network,and the loss function is used to measure the error between the true label and the predicted label;the function of the convolutional layer is to extract the characteristics of the training data set,and the method of upsampling and downsampling to change the characteristics space size;the optimization algorithm minimizes the loss function,and the use of the optimization algorithm can make the neural network learn accurately and efficiently.According to the choice of the deep learning model,the process of adjusting the neural network parameters and algorithms to reduce the risk of over-fitting and underfitting as much as possible.The deep learning inversion algorithm based on the Tensorflow framework realize the separation of forward and inversion.The learning of the model data set takes a long time,but once the learning is completed,the inversion can be completed quickly,which can be completed in a few seconds on a mainstream PC.At the same time,the trained model can be easily deployed on the Raspberry Pi and other single-board computers,which are convenient for the development of instruments.Finally,the deep learning interpretation method is applied to the magnetotelluric inversion.The inversion work is carried out in three steps: 1)establish a deep learning neural network;2)construct a variety of geological models,obtain the two-dimensional magnetotelluric response through forward modeling to construct data set;3)Learn magnetotelluric response characteristics through deep learning neural network and forward modeling data set,and apply it to the inversion prediction of unknown model.The inversion results of model examples show that the deep learning algorithm achieves satisfactory inversion results.Minimal errors are found in the spatial location,shape and size of the abnormal body,as well as the resistivity of the abnormal body and the background.At the same time,the inversion of deep learning has a certain anti-interference ability,and the inversion can still achieve a great fitting.