This paper proposes two-dimensional magnetotelluric inversion based on residual neural networks to improve the accuracy of conventional inversion with convolutional neural networks,which are affected by exces-sive layers.A large number of data sets are established through two-dimensional forward modeling of magneto-telluric data.The apparent resistivity and phase data in TE and TM modes are used as input to a four-channel network,and the corresponding geoelectric model is used as a label and output for supervised learning.The two-dimensional magnetotelluric inversion is achieved by utilizing residual neural networks.Based on the inversion results of geoelectric models with different noise levels,it is shown that residual networks can not only elimi-nate the problem of decreased accuracy caused by excessive layers but also exhibit strong noise resilience.Based on the inversion of the measured electromagnetic data in Jizhong depression,China,the resistivity distri-bution of deep carbonate rock is obtained,and the characteristics of the thermal storage structure in the working area are analyzed accordingly.Inversion results from theoretical models and measured data both show that the proposed method has excellent learning ability and noise resilience,and the inversion effect is stable and reli-able.
维普
万方数据