Elastic Impedance Inversion with GRU-CNN Hybrid Deep Learning: Visualizing the Black Box

Xiujuan Liu; Lifeng Liang; Zhiming Kang; Qiang Guo

doi:doi:10.11648/j.earth.20221104.15

| Peer-Reviewed

Elastic Impedance Inversion with GRU-CNN Hybrid Deep Learning: Visualizing the Black Box

Xiujuan Liu, Lifeng Liang, Zhiming Kang, Qiang Guo

Published in Earth Sciences (Volume 11, Issue 4)

Received: 22 June 2022 Accepted: 13 July 2022 Published: 22 July 2022

Views: Downloads:

Download PDF

Share This Article

Twitter
Linked In
Facebook

Abstract

The process of hybrid deep learning is highly integrated with the seismic inversion, leading to the black box phenomena, which hampers the understanding of deep-learning inversion process. In this paper, a numerical example is presented to visualize the process of deep learning from the perspective of elastic inversion. Synthetic seismic data is generated by forward modeling on a wedge model, after which the GRU-CNN hybrid deep learning algorithm is applied to obtain the inverted impedance method. In specific, the extraction of local seismic features by CNN, the extraction of low-frequency seismic features by GRU, activation layer, Adam and learning rate schedules, initialization model, loss function, and training process are detailed illustrated and visualized, all of which reveal the internal operating mechanism of the black box. The results show that: 1) after required epoch iterations, the features extracted by CNN becomes close to the real impedance, while the features extracted by GRU is close to the low-frequency information involved in conventional seismic inversion (which is consistent with the cognition from commercial software, e.g., Jason, Geoview, etc.), 2) The learning rate is a very critical parameter in the optimization process. Comparing with the constant learning rate, the cosine learning rate converges faster with better performance, and 3) the initial impedance model in hybrid deep learning is to initialize the weights of all neurons, which is essentially different from those of conventional seismic inversion scheme, e.g., constrained sparse pulse inversion.

Published in	Earth Sciences (Volume 11, Issue 4)
DOI	10.11648/j.earth.20221104.15
Page(s)	194-203
Creative Commons	This is an Open Access article, distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution and reproduction in any medium or format, provided the original work is properly cited.
Copyright	Copyright © The Author(s), 2022. Published by Science Publishing Group

Keywords

Elastic Impedance Inversion, GRU, CNN, Wedge Model, The Black Box

References

[1]	HINTONS G E, SALAKHUTDINOV R R., 2006, Reducing the dimensionality of data with neural networks: Science, 313, 504-507. doi: 10.1126/science.1127647.
[2]	KRIZHEVSKY A, SUTSKEVER I, HINTON G E., 2012, Imagenet classification with deep convolutional neural networks: Proceedings of the International Conference on Neural Information Processing Systems (NIPS), 2012: 1097-1105.
[3]	RUMELHART D, HINTON G, WILLIAMS R., 1986, Learning representations by back propagating errors: Nature, 323, 533-536. doi: 10.1038/323533a0.
[4]	ZHAO Gaishan., 2019, Road to intelligent petroleum geophysical exploration: From automatic to intelligent: Geophysical Prospecting for Petroleum, 58, 791-810. doi: 10.3969/j.issn.1000-1441.2019.06.002.
[5]	Wu X M, Shi Y Z, Fomel S, et al., 2019, FaultNet3D: Predicting Fault Probabilities, Strikes, and Dips With a Single Convolutional Neural Network: IEEE Transactions on Geoscience and Remote Sensing, 57, 9138-9155. doi: 10.1109/TGRS.2019.2925003.
[6]	HAN Wei-xue, ZHOU Ya-tong, CHI Yue., 2018, Deep learning convolutional neural networks for random noise attenuation in seismic data: Geophysical Prospecting for Petroleum., 57, 862-869. doi: 10.3969/j.issn.1000-1441.2018.06.008. (in Chinese).
[7]	LIU Li-hui, LU Rong, YANG Wen-kui., 2019, Seismic lithofacies inversion based on deep learning: Geophysical Prospecting for Petroleum, 58, 123-129. doi: 10.3969/j.issn.1000-1441.2019.01.014. (in Chinese).
[8]	AN Pen, CAO Dan-ping, ZHAO Bao-yin, et al., 2019, Reservoir physical parameters prediction based on LSTM recurrent neural network: Progress in Geophysics, 34, 1849-1858, doi: 10.6038/pg2019CC0366. (in Chinese).
[9]	ZHANG Dong-xiao, CHEN Yun-tian, MENG Jin., 2018, Synthetic well logs generation via Recurrent Neural Networks: Petroleum Exploration and Development, 45, 598-607. doi: 10.11698/PED.2018.04.06. (in Chinese).
[10]	WANG Wen-qiang, MENG Fan-shun, SUN Wenliang., 2019, Trace editing based on optimized convolutional neural network: Progress in Geophysics., 34, 0214-0220, doi: 10.6038/pg2019BB0387. (in Chinese).
[11]	Alfarraj M, AlRegib G., 2019, Semisupervised sequence modeling for elastic impedance inversion: Interpretation, 7, 237-249. doi: 10.1190/int-2018-0250.1.
[12]	Russell, B., 2019, Machine learning and geophysical inversion –A numerical study: The Leading Edge, 38, 498-576. doi: 10.1190/tle38070512.1.
[13]	Kim, Y., and N. Nakata, 2018, Geophysical inversion versus machine learning in inverse problems: The Leading Edge, 37, 894–901, doi: 10.1190/tle37120894.1.
[14]	Naeini, E. Z., and Prindle, K., 2018, Machine learning and learning from machines: The Leading Edge, 37, 886–893. doi: 10.1190/tle37120886.1.
[15]	Smith L N., 2017, Cyclical Learning Rates for Training Neural Networks: 2017 IEEE Winter Conference on Applications of Computer Vision (WACV), IEEE, 2017: 464-472. doi: 10.1109/WACV.2017.58.

Cite This Article

Plain Text BibTeX RIS

APA Style

Xiujuan Liu, Lifeng Liang, Zhiming Kang, Qiang Guo. (2022). Elastic Impedance Inversion with GRU-CNN Hybrid Deep Learning: Visualizing the Black Box. Earth Sciences, 11(4), 194-203. https://doi.org/10.11648/j.earth.20221104.15

Copy | Download

ACS Style

Xiujuan Liu; Lifeng Liang; Zhiming Kang; Qiang Guo. Elastic Impedance Inversion with GRU-CNN Hybrid Deep Learning: Visualizing the Black Box. Earth Sci. 2022, 11(4), 194-203. doi: 10.11648/j.earth.20221104.15

Copy | Download

AMA Style

Xiujuan Liu, Lifeng Liang, Zhiming Kang, Qiang Guo. Elastic Impedance Inversion with GRU-CNN Hybrid Deep Learning: Visualizing the Black Box. Earth Sci. 2022;11(4):194-203. doi: 10.11648/j.earth.20221104.15

Copy | Download

@article{10.11648/j.earth.20221104.15,
  author = {Xiujuan Liu and Lifeng Liang and Zhiming Kang and Qiang Guo},
  title = {Elastic Impedance Inversion with GRU-CNN Hybrid Deep Learning: Visualizing the Black Box},
  journal = {Earth Sciences},
  volume = {11},
  number = {4},
  pages = {194-203},
  doi = {10.11648/j.earth.20221104.15},
  url = {https://doi.org/10.11648/j.earth.20221104.15},
  eprint = {https://article.sciencepublishinggroup.com/pdf/10.11648.j.earth.20221104.15},
  abstract = {The process of hybrid deep learning is highly integrated with the seismic inversion, leading to the black box phenomena, which hampers the understanding of deep-learning inversion process. In this paper, a numerical example is presented to visualize the process of deep learning from the perspective of elastic inversion. Synthetic seismic data is generated by forward modeling on a wedge model, after which the GRU-CNN hybrid deep learning algorithm is applied to obtain the inverted impedance method. In specific, the extraction of local seismic features by CNN, the extraction of low-frequency seismic features by GRU, activation layer, Adam and learning rate schedules, initialization model, loss function, and training process are detailed illustrated and visualized, all of which reveal the internal operating mechanism of the black box. The results show that: 1) after required epoch iterations, the features extracted by CNN becomes close to the real impedance, while the features extracted by GRU is close to the low-frequency information involved in conventional seismic inversion (which is consistent with the cognition from commercial software, e.g., Jason, Geoview, etc.), 2) The learning rate is a very critical parameter in the optimization process. Comparing with the constant learning rate, the cosine learning rate converges faster with better performance, and 3) the initial impedance model in hybrid deep learning is to initialize the weights of all neurons, which is essentially different from those of conventional seismic inversion scheme, e.g., constrained sparse pulse inversion.},
 year = {2022}
}

Copy | Download

TY  - JOUR
T1  - Elastic Impedance Inversion with GRU-CNN Hybrid Deep Learning: Visualizing the Black Box
AU  - Xiujuan Liu
AU  - Lifeng Liang
AU  - Zhiming Kang
AU  - Qiang Guo
Y1  - 2022/07/22
PY  - 2022
N1  - https://doi.org/10.11648/j.earth.20221104.15
DO  - 10.11648/j.earth.20221104.15
T2  - Earth Sciences
JF  - Earth Sciences
JO  - Earth Sciences
SP  - 194
EP  - 203
PB  - Science Publishing Group
SN  - 2328-5982
UR  - https://doi.org/10.11648/j.earth.20221104.15
AB  - The process of hybrid deep learning is highly integrated with the seismic inversion, leading to the black box phenomena, which hampers the understanding of deep-learning inversion process. In this paper, a numerical example is presented to visualize the process of deep learning from the perspective of elastic inversion. Synthetic seismic data is generated by forward modeling on a wedge model, after which the GRU-CNN hybrid deep learning algorithm is applied to obtain the inverted impedance method. In specific, the extraction of local seismic features by CNN, the extraction of low-frequency seismic features by GRU, activation layer, Adam and learning rate schedules, initialization model, loss function, and training process are detailed illustrated and visualized, all of which reveal the internal operating mechanism of the black box. The results show that: 1) after required epoch iterations, the features extracted by CNN becomes close to the real impedance, while the features extracted by GRU is close to the low-frequency information involved in conventional seismic inversion (which is consistent with the cognition from commercial software, e.g., Jason, Geoview, etc.), 2) The learning rate is a very critical parameter in the optimization process. Comparing with the constant learning rate, the cosine learning rate converges faster with better performance, and 3) the initial impedance model in hybrid deep learning is to initialize the weights of all neurons, which is essentially different from those of conventional seismic inversion scheme, e.g., constrained sparse pulse inversion.
VL  - 11
IS  - 4
ER  -

Copy | Download

Author Information

Xiujuan Liu

Department of Geography, Lingnan Normal University, Zhanjiang, China
Lifeng Liang

Department of Geography, Lingnan Normal University, Zhanjiang, China
Zhiming Kang

Department of Geography, Lingnan Normal University, Zhanjiang, China
Qiang Guo

College of Information Engineering, China Jiliang University, Hangzhou, China

Download PDF

Sections

Plain Text BibTeX RIS

APA Style

Xiujuan Liu, Lifeng Liang, Zhiming Kang, Qiang Guo. (2022). Elastic Impedance Inversion with GRU-CNN Hybrid Deep Learning: Visualizing the Black Box. Earth Sciences, 11(4), 194-203. https://doi.org/10.11648/j.earth.20221104.15

Copy | Download

ACS Style

Xiujuan Liu; Lifeng Liang; Zhiming Kang; Qiang Guo. Elastic Impedance Inversion with GRU-CNN Hybrid Deep Learning: Visualizing the Black Box. Earth Sci. 2022, 11(4), 194-203. doi: 10.11648/j.earth.20221104.15

Copy | Download

AMA Style

Xiujuan Liu, Lifeng Liang, Zhiming Kang, Qiang Guo. Elastic Impedance Inversion with GRU-CNN Hybrid Deep Learning: Visualizing the Black Box. Earth Sci. 2022;11(4):194-203. doi: 10.11648/j.earth.20221104.15

Copy | Download

@article{10.11648/j.earth.20221104.15,
  author = {Xiujuan Liu and Lifeng Liang and Zhiming Kang and Qiang Guo},
  title = {Elastic Impedance Inversion with GRU-CNN Hybrid Deep Learning: Visualizing the Black Box},
  journal = {Earth Sciences},
  volume = {11},
  number = {4},
  pages = {194-203},
  doi = {10.11648/j.earth.20221104.15},
  url = {https://doi.org/10.11648/j.earth.20221104.15},
  eprint = {https://article.sciencepublishinggroup.com/pdf/10.11648.j.earth.20221104.15},
  abstract = {The process of hybrid deep learning is highly integrated with the seismic inversion, leading to the black box phenomena, which hampers the understanding of deep-learning inversion process. In this paper, a numerical example is presented to visualize the process of deep learning from the perspective of elastic inversion. Synthetic seismic data is generated by forward modeling on a wedge model, after which the GRU-CNN hybrid deep learning algorithm is applied to obtain the inverted impedance method. In specific, the extraction of local seismic features by CNN, the extraction of low-frequency seismic features by GRU, activation layer, Adam and learning rate schedules, initialization model, loss function, and training process are detailed illustrated and visualized, all of which reveal the internal operating mechanism of the black box. The results show that: 1) after required epoch iterations, the features extracted by CNN becomes close to the real impedance, while the features extracted by GRU is close to the low-frequency information involved in conventional seismic inversion (which is consistent with the cognition from commercial software, e.g., Jason, Geoview, etc.), 2) The learning rate is a very critical parameter in the optimization process. Comparing with the constant learning rate, the cosine learning rate converges faster with better performance, and 3) the initial impedance model in hybrid deep learning is to initialize the weights of all neurons, which is essentially different from those of conventional seismic inversion scheme, e.g., constrained sparse pulse inversion.},
 year = {2022}
}

Copy | Download

TY  - JOUR
T1  - Elastic Impedance Inversion with GRU-CNN Hybrid Deep Learning: Visualizing the Black Box
AU  - Xiujuan Liu
AU  - Lifeng Liang
AU  - Zhiming Kang
AU  - Qiang Guo
Y1  - 2022/07/22
PY  - 2022
N1  - https://doi.org/10.11648/j.earth.20221104.15
DO  - 10.11648/j.earth.20221104.15
T2  - Earth Sciences
JF  - Earth Sciences
JO  - Earth Sciences
SP  - 194
EP  - 203
PB  - Science Publishing Group
SN  - 2328-5982
UR  - https://doi.org/10.11648/j.earth.20221104.15
AB  - The process of hybrid deep learning is highly integrated with the seismic inversion, leading to the black box phenomena, which hampers the understanding of deep-learning inversion process. In this paper, a numerical example is presented to visualize the process of deep learning from the perspective of elastic inversion. Synthetic seismic data is generated by forward modeling on a wedge model, after which the GRU-CNN hybrid deep learning algorithm is applied to obtain the inverted impedance method. In specific, the extraction of local seismic features by CNN, the extraction of low-frequency seismic features by GRU, activation layer, Adam and learning rate schedules, initialization model, loss function, and training process are detailed illustrated and visualized, all of which reveal the internal operating mechanism of the black box. The results show that: 1) after required epoch iterations, the features extracted by CNN becomes close to the real impedance, while the features extracted by GRU is close to the low-frequency information involved in conventional seismic inversion (which is consistent with the cognition from commercial software, e.g., Jason, Geoview, etc.), 2) The learning rate is a very critical parameter in the optimization process. Comparing with the constant learning rate, the cosine learning rate converges faster with better performance, and 3) the initial impedance model in hybrid deep learning is to initialize the weights of all neurons, which is essentially different from those of conventional seismic inversion scheme, e.g., constrained sparse pulse inversion.
VL  - 11
IS  - 4
ER  -

Copy | Download