Magnetotelluric inversion based on convolutional neural network

In order to improve the accuracy of two-dimensional magnetotelluric inversion, a new method based on convolution neural network was proposed. The specific implementation steps were as follows. Firstly, the sample data set was obtained by two-dimensional forward modeling of different geoelectric models. Then, the convolutional neural network framework was constructed where the inputs were of apparent resistivity and phase and the outputs were corresponding geoelectric model parameters. And the network was supervised and adjusted to obtain the optimal inversion network arrangement and hyperparameter. Finally, the trained network was verified through the inversion of unknown geoelectric model. The effect of convolution neural network inversion of various geoelectric model bodies with TM mode, and the influence of input component and model body depth on the inversion effect were discussed. The results show that the inversion method proposed in this paper can realize accurate positioning and imaging of geoelectric model, and the "focusing" effect is better than that of the least square inversion. Meanwhile, the joint inversion results of apparent resistivity and phase are better than those of the single parameter inversion. the inversion effect of shallow model body is better than that of deep part, and the mean square error of joint inversion is 30%-50% of single inversion. The validity of the new method is verified by the measured results. © 2020, Central South University Press. All right reserved.