[S-EM12] Electric, magnetic, and electromagnetic survey technologies and the scientific achievements

[SEM12-05] One-dimensional Deep Learning Inversion of Marine Controlled Source Electromagnetic Data

*Pan Li¹, Z Du¹, Yuguo Li^1,2 (1.Ocean University of China, 2.Key Lab of Submarine Geosciences and Prospecting Techniques of Ministry of Education)

Keywords:CSEM, Deep Learning, EM field, resistivity inversion

The EM inverse problem is highly ill-posed, and often requires significant pre-conditioning and regularization. In this paper we explore the use of a machine learning technique for inversion of the marine Controlled Source Electromagnetic (CSEM) data. We propose a novel approach based on the deep leaning (DL) using both the convolutional neural network (CNN) and the recurrent neural network (RNN) algorithms to reconstruct variations in earth’s resistivity. Through the learning of a large number of 1D forward synthetic data by CNN and RNN, the mapping relationship between CSEM data and resistivity model is constructed. We built a large number of geological models with different EM structural complexities. Both the resistivity models and the corresponding EM field data represent a good training source for a deep learning network. For the purpose of training and validation where EM models and corresponding CSEM data are used. The trained network is then used to predict the distribution of resistivity in the model by feeding it with CSEM responses. We also explore the sensitivity of DL inversion to different EM field with resistivity targets distributed at different depths, variable amount of noise, and compare the results between DL and a standard Occam’s inversion approach. The results of the inversion show that the DL method is highly efficient and can robustly reconstruct subsurface resistivity structures. This study suggests that a framework with adoption of various DL methods for 2D and 3D EM inversion is feasible.