Deep Learning Inversion of Electrical Resistivity Data by One-Sided Mapping
Benchao Liu, Peng Jiang, Qian Guo, Chuanwu Wang
Abstract
Recently, deep learning-based electrical resistivity inversion (DL-ERI) became popular for its potential to achieve high-accuracy imaging of the subsurface's electrical properties. The most typical way is to directly learn the mapping from apparent resistivity data to the resistivity model with DNN. Therefore, they are doing cross-domain mapping, which usually causes learning difficulty. In this work, we propose to do DL-ERI by one-sided domain mapping that uses gradients domain as the input and target. Specifically, the target is represented as residuals between the resistivity model and a uniform initial model, while input is the gradients calculated from the first step of the traditional linear inversion method. Then, we could obtain the inversion results by adding the predicted target to the initial model. From experiments, our one-sided domain mapping method shows clear superiority over the cross-domain mapping baseline and demonstrates promising performance on physical experimental data.