Geophysical Inversion Enhanced by Deep Learning



Journal Title

Journal ISSN

Volume Title



Inversion is an essential technology for understanding geophysical data. In industrial practice, geophysical inversion is mainly used for high-resolution subsurface imaging. The reconstructed geological parameters include velocity, permittivity, permeability, and resistivity. The inversion results reveal interpretable features of subsurface structures and contribute to reservoir monitoring, logging while drilling, and other applications. Conventionally, solving the geophysical inversion relies on calculating the rigorous forward modeling function. Strictly following physics rules, the rigorous inversion can provide accurate model parameters. However, since the observations obtained by geophysical surveys are usually limited, geophysical inversions are usually highly non-linear and underdetermined. The complicated physical constraints may lead to local minima and slow convergence. And the rigorous numerical methods may demand a lot of computational resources. Therefore, using deep learning to enhance geophysical inversion has recently been an important topic.

This study mainly focuses on deep-learning-enhanced geophysical inversion with limited training data. Deep learning can fill the blank of the domain experience during the inversion. The deep learning approach can learn the specific data pattern from the historical data and provide different enhancements like data augmentation, data completion, and computing acceleration. However, many deep-learning approaches depend highly on a large amount of training data, which may be unavailable in field surveys. This work explores several deep-learning enhancement approaches for geophysical inversion with limited training data. The proposed methods are designed for solving different problems, including (1) the observation enhancement, (2) the end-to-end inversion surrogate, and (3) the end-to-end forward modeling surrogate. The physics rules are incorporated with the deep learning approaches by algorithms, regularizations, or constraints. For some problems, sensitivity analysis and uncertainty quantification are also studied. The experiment results show the promising performance of the proposed methods.



Deep learning, Geophysical inversion, Full-waveform inversion, Geosteering inversion, Surrogate, Signal processing


Portions of this document appear in: Yuchen Jin, Wenyi Hu, Shirui Wang, Yuan Zi, Xuqing Wu, and Jiefu Chen, “Efficient progressive transfer learning for full waveform inversion with extrapolated low-frequency reflection seismic data”, IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1-10, 2022; and in: Yuchen Jin, Wenyi Hu, Shirui Wang, Yuan Zi, Xuqing Wu, and Jiefu Chen, “A robust learning method for low-frequency extrapolation in GPR full waveform inversion”, IEEE Geoscience and Remote Sensing Letters, under review, vol. 19, pp. 1-5, 2022; and in: Yuchen Jin, Qiuyang Shen, Xuqing Wu, and Jiefu Chen, “A Physics-Driven Deep-Learning Network for Solving Nonlinear Inverse Problems”, Petrophysics, vol. 61, no. 1, pp. 86-98, Feb. 2020; and in: Yuchen Jin, and Weichang Li, “Deep network based geosteering inversion with crossgradients”, First International Meeting for Applied Geoscience & Energy, Society of Exploration Geophysicists, 2021, pp. 562-566; and in: Yuchen Jin, Chaoxian Qi, Li Yan, Yueqin Huang, XuqingWu, and Jiefu Chen, “A deep learning based surrogate and uncertainty quantification for fast electromagnetic modeling in complex formations”, 2022 International Meeting for Applied Geoscience & Energy, Society of Exploration Geophysicists, 2022.