See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/376548515

Joint data and physics model driven full-waveform inversion using CMP gathers
and well-logging data

Conference Paper · December 2023

DOI: 10.1190/image2023-3911297.1

CITATIONS READS
0 18

2 authors:

Shuliang Wu Jianhua Geng

Tongji University Tongji University
3 PUBLICATIONS 2 CITATIONS 86 PUBLICATIONS 954 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Shuliang Wu on 23 December 2023.

The user has requested enhancement of the downloaded file.

Downloaded 12/23/23 to 111.187.54.13. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/page/policies/terms

Joint data- and physics-model-driven full-waveform inversion using CMP gathers and well-logging
data
Shuliang Wu and Jianhua Geng, Tongji University, State Key Laboratory of Marine Geology.

Summary learning platforms can avoid calculating the gradient

explicitly and reduce the complexity of the inversion, its
We propose a joint data- and physics-model-driven full- inversion results have low accuracy due to the structure
waveform inversion (FWI) method based on semi- limitations of RNNs and PINNs for the inability to use
supervised learning framework, which uses well-logging higher-order complex differential forms to solve the wave
data, pseudo labels produced from conventional FWI and equation. At the same time, some data-driven FWI methods
common mid-point (CMP) gathers to train neural network. based on neural networks have been proposed (Wang et al.,
Neural network builds mapping relationship between 2018; Yang and Ma, 2019; Sun et al., 2021; Wang and Ma,
several adjacent CMP gathers and the vertical profiles of 2020; Feng et al., 2022), which usually use large datasets to
the subsurface model. The conventional model-driven FWI train neural networks in a supervised learning way so that
is used to produce pseudo label datasets to train neural the neural networks learn the mapping relationships from
network, which can reduce the reliance on massive datasets. seismic data to a subsurface model. These data-driven
At the same time, FWI is also a physical model constraint methods can alleviate the problem that traditional model-
on the neural network to make the inversion result more driven methods tend to fall into local extremes to a certain
physical interpretable. Using our method, the neural extent, but such data-driven methods require a large
network is able to utilize information from both logging amount of training data to cover the complex geological
data and physical models, thus it can take advantage of conditions in the subsurface. Such training data are often
DOI:10.1190/image2023-3911297.1

well-logging data to make the inversion more accurate and difficult to be obtained in the actual data inversion. Besides,
stable. Synthetic tests on the Marmousi2 model and such methods are not supported by physical mechanisms
Overthrust model show that compared with the thus lack physical interpretability. Some researchers (Wu
conventional model-driven FWI, our method can get more and McMechan, 2019; He et al., 2021; Zhu et al., 2022;
accurate and stable inversion result in the situation of Dhara and Sen, 2022) proposed to employ neural networks
lacking low-frequency data and bad initial model. as a regularization method for FWI, which is effective but
computationally complex. None of the above methods uses
Introduction the well-logging data, which is the most truthful data in
actual seismic exploration.
Velocity model building is the key task in seismic To fully utilize the well-logging data to improve the
exploration (Doherty and Claerbout, 1976), and there are stability and accuracy, mitigate the reliance of the massive
various methods such as tomograph methods and migration dataset of the data-driven method, increase the physical
velocity analysis, which are often not accurate enough. Full interpretability, we proposed a joint data- and physics-
waveform inversion (FWI) uses wave equation to simulate model-driven method based on well-logging data and
wave phenomena, and get an accurate subsurface velocity common midpoint gathers. Our method trains neural
model through some optimization methods to minimize the networks in a semi-supervised way. A physical model is
error between actual observation waveform data (Tarantola, added into the inversion procedure, which reduces the
1986). However, FWI is a high nonlinear inverse problem reliance on massive datasets and increases the physical
(Virieux and Operto, 2009). Therefore, conventional model interpretability.
driven FWI usually requires a good initial velocity model
and low-frequency component in seismic data to get a good Method
result, otherwise it is easy to fall into local extrema. In
recent years, deep learning techniques get rapid The proposed method is based on semi-supervised learning
development. Some researchers start to use deep learning framework. Firstly, we use U-net (Ronneberger et al.,2015)
techniques to solve FWI problems. Richardson (2018), Sun to build mapping relationship between several adjacent
et al. (2020) and Zhang et al. (2020) proposed to use CMP gathers and the vertical profiles of the subsurface
recurrent neural networks (RNNs) to solve wave equations model, which can be expressed as
and invert the model parameters based on the automatic � = �(�)
differentiation mechanism of the deep learning platforms. where � represents the neural network, � represents the
Song et al. (2021) and Rasht-Behesht et al. (2022) proposed vertical profile of subsurface model and � represents
to use physics-informed neural networks (PINNs) for several adjacent CMP gathers. Since the structure of
wavefield simulation and full-waveform inversion. The subsurface model is complex and the single vertical profile
above methods are still model-driven methods. Although may be related to multiple adjacent CMP gathers,the
the automatic differentiation mechanism of the deep several adjacent CMP gathers are used to build mapping

Third International Meeting for Applied Geoscience & Energy 10.1190/image2023-3911297.1

© 2023 Society of Exploration Geophysicists and the American Association of Petroleum Geologists Page 640
 Joint data- and physics-model-driven FWI using CMP gathers and well-logging data
Downloaded 12/23/23 to 111.187.54.13. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/page/policies/terms

relationship with the single vertical profile velocity, which

can be expressed by the equation:
�� = �(�� )，�� = ���� �−� , …，��
���
, …，����
�+�
���
where �� represents CMP gathers centered on the i-th
trace, �� represents the velocity profile on the i-th trace.
In our joint data- and physics-model-driven inversion
method, we use the well-logging velocity profiles ���� and
Figure 1. The inversion procedure diagram.
well side CMP gathers ����_� to build labeled datasets. In
non-well side positions, an initial model ������� (which
And then we perform one iteration of model-driven FWI
will be updated in the process of inversion) and CMP based on an initial model to get an updated velocity model
gathers ����_� are used to build pseudo label datasets. The as the new pseudo label ������� . New pseudo label datasets
inversion procedure diagram is shown in Figure 1. Firstly,
can be built using ������� and the corresponding CMP
convert the observation data into CMP gathers, and create
label and pseudo label datasets according to the location of gathers. These pseudo label datasets and label datasets are
the well. After inputting the label and pseudo label datasets used to update parameters of neural networks together.
into the neural network, the network outputs the vertical Therefore, we can obtain new pseudo label ������� in each
velocity profiles �����_� at the location of the well and iteration and input the new pseudo label dataset and the
�����_� at non-well locations. The formula can be expressed label dataset into the neural network � for iterative
optimization inversion. Finally, when the misfit of the
as follows:
seismic data in the full waveform inversion converges or
�����_� = �(����_� )
falls within a certain value, the whole process stops
DOI:10.1190/image2023-3911297.1

�����_� = �(����_�) iteratively and the neural network training is also finished.
We choose to use Huber norm to calculate losses, which After the training stage of the neural network is finished, all
can be expressed as CMP gathers are input to inversion networks to obtain the
0.5 ∗ (� − �)2, �� � − � < � final inversion results. In this semi-supervised learning
������ �, � =
� − � − 0.5, ��ℎ������ approach, a model-driven FWI is used to provide pseudo
So, the ����� and ����� can be calculated as follows: labels for network training, which both alleviates the
����� = ������ ������ , ���� reliance of massive datasets and serves as a physical
����� = ������ �����_� , ������� mechanism guide for neural network updates, enhancing
the physical interpretability and accuracy of the inversion.
Then the total loss is obtained by weighting these two
And the well logging data in this method is used to make
losses, which can be expressed as
the inversion more accurate and stable.
����� = �� ����� + �� �����
where �� and �� denote weighting factors. The parameters
of the neural network are updated by minimizing ����� .
Once the training stage is finished, we can get vertical
velocity profiles �����_� and �����_� by inputting ����_� and
����_� . So, a 2D velocity model can be obtained by
concatenating these 1D velocity profiles in order. Since the
inversion by CMP gathers is a single trace inversion, we
use a generative neural network to reparameterize the
concatenated 2D velocity profiles to improve the lateral
continuity of the inversed velocity model, and the equation
can be expressed as
� = �(�, �����_� , �����_� )
Figure 2． Structure of U-net in our method.
where, � represents the generative neural network, �
represents the random vector, � represents the
reparametrized velocity model.

Figure 3. Structure of generative neural network.

Third International Meeting for Applied Geoscience & Energy 10.1190/image2023-3911297.1

© 2023 Society of Exploration Geophysicists and the American Association of Petroleum Geologists Page 641
 Joint data- and physics-model-driven FWI using CMP gathers and well-logging data
Downloaded 12/23/23 to 111.187.54.13. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/page/policies/terms

Examples

To verify the validity of our method, we perform synthetic

experiments on the Marmousi2 model based on a linear
initial model and seismic data lack of low-frequency
component. The grid of this model is 160*80 with spacing
size of 20m. The Ricker wavelet with a dominate frequency
of 8 Hz is high-pass filtered at 4 Hz as the source to
simulate the seismic data. We select one velocity profile
located at 1200m (the red line in Figure 4a) as the well log
to train the neural network and it is also used as a reference
to build the initial model (Figure 4b) for inversion. The
number of shots is 16 with interval of 200m and that of
receiver per shot is 160 with interval of 20m. The
maximum recording time is 2048ms with 2ms interval.
DOI:10.1190/image2023-3911297.1

Figure 5. Seismic record from source position at 2400m. (a)

True field data, (b) the simulated data from Figure 4c, (c)
the simulated data from Figure 4d, (d) the simulated data
from Figure 4e, (e) residuals of (b), (f) residuals of (c), (g)
residuals of (d).

For comparison, we use conventional model-driven FWI

Figure 4. (a) Real model, (b) initial model, (c) the and our method to do inversion. As is shown in Figure 4c,
conventional model-driven FWI result, (d) the joint data- due to the bad initial model and lack of low-frequency data,
and physics-model-driven FWI result, (e) the conventional the conventional model-driven FWI result falls into local
model-driven FWI result using (d) as the initial model. extrema obviously, which is also shown in Figure 5. As for
the joint data- and physics-model-driven method, a good
inversion result can be obtained. However, due to the
spectrum bias of neural networks (Xu et al., 2019), the joint
data- and physics-model-driven method tends to get a low
frequency result. Therefore, conventional model-driven
FWI is used to get a higher frequency velocity model based
on the joint data- and physics-model-driven FWI result, as
shown in Figure 4e. From the comparisons of velocity
profiles shown in Figure 6, compared with model-driven
method, the joint data- and physics-model-driven FWI
method is more stable and accurate.

Third International Meeting for Applied Geoscience & Energy 10.1190/image2023-3911297.1

© 2023 Society of Exploration Geophysicists and the American Association of Petroleum Geologists Page 642
 Joint data- and physics-model-driven FWI using CMP gathers and well-logging data
Downloaded 12/23/23 to 111.187.54.13. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/page/policies/terms

Figure 6. Comparisons of velocity profile at (a) 1200m, (b)

1600m, (c) 2800m. Green line denotes real velocity; yellow
line denotes initial velocity; red line denotes the
conventional model-driven method result; blue line denotes
the joint driven method result; black line denotes the
conventional model-driven method result after the joint
driven method.

Conclusions

We propose a joint data- and physics-model-driven FWI

DOI:10.1190/image2023-3911297.1

method, which uses well-logging data and pseudo labels

produced from conventional FWI and CMP gathers to train
neural networks. Neural networks are used to build
mapping relationships between well-logging data and
subsurface velocity profiles. The conventional FWI is used
to produce pseudo label data to reduce the reliance of the
network on massive datasets. What’s more, using this
method, the neural network can take advantage of well-
logging data to make the inversion more accurate and
stable in the situation of lacking good initial model and
low-frequency component of seismic data.

Acknowledgments

The research leading to this paper was supported by

Sponsors of Tongji University and by State Key Laboratory
of Marine Geology of China.

Third International Meeting for Applied Geoscience & Energy 10.1190/image2023-3911297.1

© 2023 Society of Exploration Geophysicists and the American Association of Petroleum Geologists Page 643
 REFERENCES
Downloaded 12/23/23 to 111.187.54.13. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/page/policies/terms

Dhara, A., and M. K. Sen, 2022, Physics-guided deep autoencoder to overcome the need for a starting model in full-waveform inversion: The Leading
Edge, 41, 375–381, doi: https://doi.org/10.1190/tle41060375.1.
Doherty, S. M., and J. F. Claerbout, 1976, Structure independent velocity estimation: Geophysics, 41, 850–881, doi: https://doi.org/10.1190/1
.1440669.
Feng, S., Y. Lin, and B. Wohlberg, 2021, Multiscale data-driven seismic full-waveform inversion with field data study: IEEE Transactions on Geo-
science and Remote Sensing, 60, 1–14, doi: https://doi.org/10.1109/TGRS.2021.3114101.
Rasht‐Behesht, M., C. Huber, K. Shukla, and G. E. Karniadakis, 2022, Physics‐informed neural networks (PINNs) for wave propagation and full
waveform inversions: Journal of Geophysical Research: Solid Earth, 127, e2021JB023120, doi: https://doi.org/10.1029/2021JB023120.
Richardson, A., 2018, Seismic full-waveform inversion using deep learning tools and techniques: arXiv preprint, doi: https://doi.org/10.48550/arXiv
.1801.07232.
Ronneberger, O., P. Fischer, and T. Brox, 2015, U-Net: Convolutional networks for biomedical image segmentation: International Conference on
Medical Image Computing and Computer-Assisted Intervention, Springer, 234–241.
Song, C., and T.A. Alkhalifah, 2021, Wavefield reconstruction inversion via physics-informed neural networks: IEEE Transactions on Geoscience and
Remote Sensing, 60, 1–12, doi: https://doi.org/10.1109/TGRS.2021.3123122.
Sun, J., K. A. Innanen, and C. Huang, 2021, Physics-guided deep learning for seismic inversion with hybrid training and uncertainty analysis:
Geophysics, 86, no. 3, R303–R317, doi: https://doi.org/10.1190/geo2020-0312.1.
Sun, J., Z. Niu, K. A. Innanen, J. Li, and D. O. Trad, 2020, A theory-guided deep-learning formulation and optimization of seismic waveform
inversion: Geophysics, 85, no. 2, R87–R99, doi: https://doi.org/10.1190/geo2019-0138.1.
Tarantola, A., 1986, A strategy for nonlinear elastic inversion of seismic reflection data: Geophysics, 51, 1893–1903, doi: https://doi.org/10.1190/1
.1442046.
Virieux, J., and S. Operto, 2009, An overview of full-waveform inversion in exploration geophysics: Geophysics, 74, no. 6, WCC1–WCC26, doi:
https://doi.org/10.1190/1.3238367.
Wang, W., and J. Ma, 2020, Velocity model building in a crosswell acquisition geometry with image-trained artificial neural networks: Geophysics,
85, no. 2, U31–U46, doi: https://doi.org/10.1190/geo2018-0591.1.
Wang, W., F. Yang, and J. Ma, 2018, Velociy model building with a modified fully convolutional network: 88th Annual International Meeting, SEG,
Expanded Abstracts, 2086–2089, doi: https://doi.org/10.1190/segam2018-2997566.1.
Wu, Y., and G. A. McMechan, 2019, Parametric convolutional neural network-domain full-waveform inversion: Geophysics, 84, no. 6, R881–R896,
doi: https://doi.org/10.1190/geo2018-0224.1.
Xu, Z. Q. J., Y. Zhang, and Y. Xiao, 2019, Training behavior of deep neural network in frequency domain: 26th International Conference on Neural
DOI:10.1190/image2023-3911297.1

Information Processing, 264–274.

Yang, F., and J. Ma, 2019, Deep-learning inversion: A next-generation seismic velocity model building method: Geophysics, 84, no. 4, R583–R599,
doi: https://doi.org/10.1190/geo2018-0249.1.
Zhang, T., K. Innanen, J. Sun, and D. Trad, 2020, Numerical analysis of a deep learning formulation of elastic full waveform inversion: 90th Annual
International Meeting, SEG, Expanded Abstracts, 1531–1535, doi: https://doi.org/10.1190/segam2020-3426826.1.
Zhu, W., K. Xu, E. Darve, B. Biondi, and G Beroza, 2022, Integrating deep neural networks with full-waveform inversion: Reparameterization,
regularization, and uncertainty quantification: Geophysics, 87, no. 1, R93–R109, doi: https://doi.org/10.1190/geo2020-0933.1.

Third International Meeting for Applied Geoscience & Energy 10.1190/image2023-3911297.1

© 2023 Society of Exploration Geophysicists and the American Association of Petroleum Geologists Page 644

View publication stats