Inversion of 1D audio magnetotelluric data based on residual convolution neural network

Abstract

The 18-layer residual convolution neural network (ResNet18) were used for 1D audio magnetotelluric(AMT) data inversion. In order to avoid the dependence of the traditional iterative algorithm on the initial model and calculate sensitivity matrix, we have trained ResNet18 via providing model parameters instantaneously. The residual network was used to solve the problem of deep network gradient disappearing. Deep network and lots of sample data could improve the generalization of the network. The experimental results showed that it could obtain reliable inversion results for synthetic AMT data.