Application of Gradient Learning Scheme to Pixel-Based Inversion for Transient EM Data

Abstract

Traditional gradient descent inversion for transient electromagnetic (TEM) data is time and memory comsuming because the derivative matrices need to be computed repeatly. In this paper, we apply the Supervised Descent Method (SDM) into pixel-based inversion for TEM data. This method is based on the concept of gradient learning. In an offline stage, the average descent direction can be learned from a set of training data; and in an online stage, data inversion can be achieved by the learned descent directions without computing the derivative matrices. Numerical tests verify that this algorithm converges faster and is also more efficient. Moreover, SDM offers a more convinient way to incorporate prior information into inversion that could improve the efficiency of data interpretation.