2D Magnetotelluric Inversion using Supervised Descent Method

In this work, we study the 2D magnetotelluric (MT) inversion by a machine learning scheme. Specifically, supervised descent method (SDM), which learns the descent directions of an optimization problem, is applied to MT data inversion. The inversion is composed of offline training and online prediction. In the offline stage, the descent directions can be learned from a set of training data; and in the online stage, data inversion can be achieved by updating the models using the learned directions. SDM offers a more convinient way to incorporate prior information into MT inversion. Furthermore, the learning scheme can achieve fast computation. Numerical tests show that we may be able to reconstruct thin layers from MT data with enough information.