Deep Learning Enhanced Joint Inversion of

2021 IEEE International Symposium on Antennas and Propagation and USNC-URSI Radio Science Meeting (APS/URSI) | 978-1-7281-4670-6/21/$31.00 ©2021 IEEE | DOI: 10.1109/APS/URSI47566.2021.9703802

Multiphysics Data with Nonconforming

Discretization
Yanyan Hu, Jiefu Chen, Xuqing Wu Yueqin Huang
University of Houston Cyentech Consulting LLC
Houston, TX, United States Cypress, TX, United States

Abstract—Join inversion has been playing an important role the workflows for separate EM and acoustic inversions. After
in integrating multiphysics data to reduce inversion uncertainties the separately inverted models are obtained, we send them
and improve resolution. In this paper, we propose a deep learning into a well-trained network. This network extracts and fuses
enhanced (DLE) joint inversion framework which enforces struc-
tural similarity by a deep neural network (DNN) and considers information from the input models, and output two mod-
nonconforming discretizations of different physical models due els with improved structural consistency and better defined
to the resolution capability of different measuring data. The property values. The output models from neural network are
framework enables the DNN to fuse and extract the information then combined with inverted models from separate inversions
from separate inversion workflows in the model domain and to produce updated initial models for the next round of
improves the joint inversion result iteratively. Interpolation can
be performed due to the model domain fusion to deal with the inversions. The detailed implementation is shown in Algorithm
nonconforming discretizations. Numerical examples demonstrate 1, where P types of physical data are jointly inverted.
the inverted models by our DLE method have more accurate EM Acoustic
property values and better structural features compared to inversion inversion
traditional separate inversions.
EM data, Acoustic data,
I. I NTRODUCTION initial guess, initial guess,
a priori, etc Updated Updated a priori, etc
Multiple measuring methods with different sensitivities to EM model acoustic model
the physical properties have different resolution capabilities. Acoustic inverse
Joint inversion of multiphysics data is to leverage the under- Predicted Predicted solver
lying physical correlations or structural consistency to build EM model acoustic model

a unified inversion framework and reduce the uncertainty of

the recovered models [1]. Typically, traditional joint inversion
methods utilize the same discretization mesh without consid-
ering the resolution difference between different measuring
methods, which may cause biased model parameterization [2]. Fig. 1. The demonstration flowchart of the DLE joint inversion framework.
Deep learning techniques have also been used in joint inversion
Currently, we design a double-channel U-Net for jointly
but still do not contemplate the discretization mismatch [3]–
inverting two types of physical data. The loss function for
[5]. In this paper, we design a deep learning enhanced (DLE)
training the neural network is defined as
joint inversion framework to extract the complementary infor- X
mation in the model domain, which achieves great flexibility. L= λp Lmp
For each physics, we may want to choose a numerical scheme p
with its own discretization strategy specifically suitable for
X
λp MSE mpred − mtrue


= p p , p ∈ {EM, A} (1)
the corresponding forward modeling. Interpolation can be per- p
formed due to the model domain fusion so that the separately
where p represents different physics; Lmp is the mean square
inverted models can be concatenated as a multi-channel input
error (MSE) between the predicted model mpred p and the true
of the network.
model mtrue
p corresponding to physics p; λ p is the coefficient
II. J OINT I NVERSION M ETHODOLOGY to balance between Lmp .
A. Deep Learning Enhanced Joint Inversion Framework B. Discretization and Mapping
In Fig. 1, we use joint inversion of electromagnetic (EM) Acoustic data usually have higher resolution than EM data.
and acoustic data as the example to illustrate our DLE Hence we adopt fine mesh with the velocity model and coarse
framework. The corresponding inverted models are resistivity mesh with the resistivity model for the separate inversions.
and velocity. The leftmost and rightmost columns indicate Then we interpolate the inverted resistivity into the fine mesh

978-1-7281-4670-6/21/$31.00 ©2021 IEEE 1489 APS 2021

Authorized licensed use limited to: University of Houston. Downloaded on March 24,2025 at 01:34:35 UTC from IEEE Xplore. Restrictions apply.
Algorithm 1 DLE joint inversion Fig. 3 indicates that the DLE method achieves more accurate
obs
1: Input: dp , minit
p , inverse solverp , N , p = 1, ..., P inverted resistivity values as well as clearer boundaries to the
2: Output: mrec
p , p= 1, ..., P true model than conventional inversion. The jointly inverted
3: Initialize: f lagp ← 0, count ← 0, p = 1, ..., P velocity also has a cleaner background and better structural
4: while (not all f lagp = 2, p = 1, ..., P ) and count < N features.
do D E F
−600 120
5: if f lagp = 0, p = 1, ..., P then −650
← inverse solverp dobs


6: minv
p p , mp
init
−700
100

Resistivity (Ωm)
7: end if −750
80

8: mpred , ..., mpred = N et(minv inv

1 , ..., mP , Θ)

z(m)
1 P −800 60
9: if f lagp = 0, p = 1, ..., P then −850

10: mupdate
p ← minvp + mpred
p −900
40

Set f lagp ← 1 if data misfit of mupdate p is within the −950

20

target range −1000

minit
p ← mupdate
p G H I
−600 4000
11: else if f lagp = 1, p = 1, ..., P then
obs −650


12: mrec
p ← inverse solverp dp , mp
init 3500

−700 3000
f lagp ← 2

Velocity (m/s)
−750
minit ← mrec 2500

z(m)
p p −800
2000
13: end if −850
1500
14: count ← count + 1 −900
1000
15: end while −950
500
−1000
0 50 100 150 200 0 50 100 150 200 0 50 100 150 200
x(m) x(m) x(m)
same with the velocity, concatenate the two models with fine
mesh, and input it into the network. The output of the network Fig. 3. The true and inverted resistivity and velocity models. (a), (d) True
models. (b), (e) Separately inverted models. (c), (f) DLE jointly inverted
is also double-channel and with the fine mesh. It will be models.
mapped back to the originally defined meshes and perform
IV. C ONCLUSION
as the initials for the separate inversions in the next loop. The
complete process is demonstrated in Fig. 2. In this work, we deployed a deep neural network to combine
conventional separate inversion workflows to achieve joint in-
version of multiphysics data. Information fusion implemented
in model domain enables our framework the ability to deal
A B C with different discretizations for different property models.
Numerical results verify our method can achieve an improved
value accuracy and better structural features compared to
traditional separate inversions.
Fig. 2. The discretization and mapping process. A: Interpolation. B: Con-
catenation. C: Mapping. ACKNOWLEDGMENT
III. E XPERIMENTS The authors acknowledge the support from the U.S. De-
Yanyan Hu (ECE, UH)
We experiment on joint inversion of cross-well DC resistiv- partment of Energy, Office of Science, and Office of Fossil
ity data and acoustic travel time. The distance between the two Energy, under Award Numbers DE-SC0020910.
boreholes is 200m. The depth ranges from 600m to 1000m R EFERENCES
underground. For the DC resistivity survey, we deployed a
[1] W. Hu, A. Abubakar, and T. M. Habashy, “Joint electromagnetic and
dipole-dipole AB-MN array. The distance between the current seismic inversion using structural constraints,” GEOPHYSICS, vol. 74,
electrodes is 40m and the spacing is 40m. The distance be- no. 6, pp. R99–R109, 2009.
tween the potential electrodes is 20m and the spacing is 40m. [2] X. Tu and M. S. Zhdanov, Joint Gramian inversion of geophysical data
with different resolution capabilities, 2020, pp. 1750–1754.
For the acoustic survey, the distances between the sources [3] R. Guo, H. M. Yao, M. Li, M. K. P. Ng, L. Jiang, and A. Abubakar,
and receivers are both 20m. The inversion domain is divided “Joint inversion of audio-magnetotelluric and seismic travel time data
into 10m × 10m for resistivity and 5m × 5m for velocity. with deep learning constraint,” IEEE Transactions on Geoscience and
Remote Sensing, pp. 1–14, 2020.
When constructing the training dataset, we do not assume any [4] D. Colombo, W. Li, D. Rovetta, E. Sandoval-Curiel, and E. Turkoglu,
specific relationship between resistivity and velocity values. Physics-driven deep learning joint inversion, 2020, pp. 1775–1779.
But we do assume structural similarity between the resistivity [5] Y. Hu, Y. Jin, X. Wu, J. Chen, J. Chen, Q. Shen, and Y. Huang, “Deep
learning enhanced joint geophysical inversion for crosswell monitoring,”
and velocity models. in 2021 United States National Committee of URSI National Radio
We provide one example in Fig. 3 to demonstrate the inver- Science Meeting (USNC-URSI NRSM), 2021, pp. 101–102.
sion results and compare to conventional separate inversions.

1490

Authorized licensed use limited to: University of Houston. Downloaded on March 24,2025 at 01:34:35 UTC from IEEE Xplore. Restrictions apply.