Supervised Descent Method for 2D Magnetotelluric Inversion using Adam Optimization

In this work, we apply the Adam optimization to the training process of the supervised descent method (SDM) for 2D magnetotelluric (MT) data inversion. Instead of solving the linear regression by direct methods such as the singular value decomposition (SVD), we use the Adam optimization to minimize the objective function during the SDM training process. This method has lower time and memory cost compared with the direct method when the training dataset is massive and high-dimensional. Also, it is capable of drawing support from the deep-learning framework and can be further accelerated using graphical processing units (GPU). Numerical tests on reconstructing conductivity distribution from MT data has validated the feasibility of the Adam based SDM inversion.