Deep learning audio magnetotellurics inversion using residual-based deep convolution neural network

In this study, we developed a novel 18-layers residual full convolutional neural network (18RFCN) for audio magnetotellurics (AMT) data inversion. Different from traditional inversion methods, the 18RFCN-based inversion algorithm is based on big-data training rather than prior-information assumptions, and the prediction is almost no time-consuming after network training. The problem of gradient disappearance was solved by adding shortcut connections when training deep neural networks, and we proposed two novel methods to quickly generate millions of samples for network training to improve the network generalization. We first tested on four synthetic data sets with Gaussian noises, which are entirely different from the data of sample sets, and the inversion results showed that the deep learning algorithm could accurately recover the underground conductivity structure almost no time-consuming compared to the Gauss Newton (GN) algorithm. We further tested the inversion algorithm on a field AMT data set acquired in Qinghai province, China. The inversion results of the deep learning algorithm agree well with the GN algorithm and the drilling data.