One-dimensional magnetotelluric (MT) data inversion modeling using convolutional neural network

Abstract

The magnetotelluric method is a geophysical method commonly used to map subsurface resistivity. The subsurface’s true resistivity is generated by inversion of the magnetotelluric data. Inversions carried out using conventional methods such as linear and global approaches have several limitations including the need for an initial model, models trapped in local minima, a large number of iterations and long computation time. To overcome the drawbacks, this paper proposes to invert one-dimensional magnetotelluric data using one of the deep learning methods, the convolutional neural network, which is heavily inspired by the human nervous system. This method starts by training the network with large amounts of data. The trained network is then used for inversion by receiving input in the form of apparent resistivity data and generating true resistivity and thickness values instantly. This method has been tested on synthetic data with curves of type A, H, K, and Q. The inversion results show that the convolutional neural network could approach the true resistivity and thickness values with a fairly small error and extremely fast computation time without initial model guess and iteration.