Joint inversion of Time-Lapse Surface Gravity and Seismic Data

for Monitoring of 3D CO2 Plumes via Deep Learning

Adrian Celaya Mauricio Araya-Polo
Rice University and TotalEnergies EP Research and TotalEnergies EP Research and Technology USA
Technology USA Houston, TX, USA
Houston, TX, USA
arXiv:2310.04430v1 [physics.geo-ph] 24 Sep 2023

ABSTRACT Injecting CO2 is just part of the process; during and after injec-
We introduce a fully 3D, deep learning-based approach for the tion, ensuring the integrity of these geological sites necessitates
joint inversion of time-lapse surface gravity and seismic data for ongoing monitoring. Regulatory authorities mandate the demon-
reconstructing subsurface density and velocity models. The target stration of storage volume containment and the detection of any
application of this proposed inversion approach is the prediction potential CO2 leakage or unwanted migration. Given the time scales
of subsurface CO2 plumes as a complementary tool for monitor- involved in monitoring CO2 storage sites, traditional monitoring
ing CO2 sequestration deployments. Our joint inversion technique techniques based on borehole sensors or surface seismic monitoring
outperforms deep learning-based gravity-only and seismic-only may not be practical or economically viable. Remote sensing by
inversion models, achieving improved density and velocity recon- other modes, such as gravity, might be economically viable and
struction, accurate segmentation, and higher R-squared coefficients. technically feasible if combined with traditional seismic approaches.
These results indicate that deep learning-based joint inversion is an Once data from the field is recorded, geophysical inversion tech-
effective tool for CO2 storage monitoring. Future work will focus niques are deployed. These widely used techniques for interpreting
on validating our approach with larger datasets, simulations with data sets and recovering subsurface physical models can be cru-
other geological storage sites, and ultimately field data. cial in monitoring CO2 storage sites [3, 14]. The recovered models,
which encompass parameters such as velocity, density, resistivity,
CCS CONCEPTS or saturation, provide essential information about the structural
and compositional characteristics of the subsurface.
• Applied computing → Physics; Environmental sciences.
Integrating multiple datasets collected over the same area is
known as geophysical inversion [18, 38]. This approach enhances
KEYWORDS
the reconstruction of subsurface structures and facilitates the iden-
Deep learning, joint inversion, gravity, seismic, carbon capture tification of changes or anomalies associated with activities like
utilization and storage CO2 storage [41]. However, effectively integrating the information
ACM Reference Format: embedded in multiple geophysical datasets presents a practical chal-
Adrian Celaya and Mauricio Araya-Polo. 2018. Joint inversion of Time-Lapse lenge from a computational point of view (i.e., storage, memory,
Surface Gravity and Seismic Data for Monitoring of 3D CO2 Plumes via and compute time), which compounds the inherent difficulties of
Deep Learning. In Proceedings of Make sure to enter the correct conference solving nonlinear inversion problems [40]. These challenges espe-
title from your rights confirmation emai (SC23). ACM, New York, NY, USA, cially become apparent when working with realistic 3D synthetic
7 pages. https://doi.org/XXXXXXX.XXXXXXX
or field data.
In this work, we develop an effective supervised 3D deep learning
1 INTRODUCTION (DL)-based inversion method to recover high-resolution subsurface
Reducing CO2 concentration in the atmosphere is critical to control CO2 plumes from both surface gravity and seismic data. To the best
climate change. These efforts involve implementing various tech- of our knowledge, this is the first fully 3D approach for the joint
nologies, such as efficient fossil-based fuel consumption, expanding inversion of surface gravity and seismic data, which is tested on
absorption sources through afforestation/reforestation, and adopt- realistic, physics-simulated CO2 plumes.
ing CO2 capture, utilization, and storage (CCUS) techniques.
Among the CCUS technologies currently deployed worldwide,
CO2 geological storage has emerged as a promising approach. This 2 PREVIOUS WORK
technology involves capturing CO2 from fixed sources or directly Conventional inversion methods find a model with the minimum
from air, then injecting it into underground formations. possible structure and whose forward response fits the observed
data [11, 26, 31]. The minimum structure is achieved by minimizing
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed model roughness through a least squares regression, resulting in a
for profit or commercial advantage and that copies bear this notice and the full citation smooth model. While the least squares regression produces smooth
on the first page. Copyrights for components of this work owned by others than ACM subsurface models, the predicted models are often larger and exhibit
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a smaller density values than the actual model [6, 35].
fee. Request permissions from [email protected]. DL is an emerging alternative to traditional geophysical inver-
SC23, Nov. 12–17, 2023, Denver, CO sion [4, 20, 22, 42]. Over the last several years, deep convolutional
© 2018 Association for Computing Machinery.
ACM ISBN 978-1-4503-XXXX-X/18/06. . . $15.00 neural networks (CNNs) have achieved state-of-the-art results in
https://doi.org/XXXXXXX.XXXXXXX various computer vision applications such as image classification,
SC23, Nov. 12–17, 2023, Denver, CO Celaya and Araya-Polo

segmentation, and generation [5, 12, 21]. CNNs have recently been exceeding 1Gt of CO2 . The formation encompasses an aquifer with
used for inversion of seismic [1, 9, 25], electromagnetic imaging an approximate thickness of 100m, spans 100km in the north-south
[10, 32], electrical resistivity [27, 36], and time-lapse surface gravity direction and 60km from east to west, and is at a depth that ranges
data [8, 19, 43, 44]. However, these works focus solely on inverting from 2,200m to 3,100m below sea level. This setting provides op-
a single modality and do not explore joint inversion with multi- timal pressure and temperature conditions for injecting CO2 in a
physical data. supercritical state.
Using simulated CO2 plumes from the onshore Kimberlina site, We generate data based on the Johansen formation using the
Um et al. developed a 2D DL architecture to perform a joint in- process described in [8]. That is, we generate a number of distinct
version with seismic, electromagnetic, and gravity data [41]. Addi- geological realizations that vary in porosity and permeability. We
tionally, they use a modified version of their architecture to invert conduct a fluid flow simulation that assumes a 100-year injection pe-
their imaging data types individually. In each case, their DL-based riod followed by a 400-year migration period. This process produces
approach can recover CO2 plumes. However, their approach still subsurface changes in density. To convert these density models to
does not perform DL-based inversion in a fully 3D setting. Hu et al. 𝑉𝑝 models for forward seismic modeling, we use the following
present a physics-informed DL-based approach for inverting elec- conversion:
tromagnetic and seismic data for recovering subsurface anomalies √︄
[18]. While their approach successfully reconstructs the anomalies, 𝜅 + 34 𝜇
𝑉𝑝 (r) = , (3)
this approach also uses 2D data and does not consider 3D data or the 𝜌 (r)
computational cost of implementing physics-informed inversion where r is the spatial coordinate in our model, 𝜅 = 8.14GPa is
for such data; as opposed to our work, which implements fully 3D the bulk modulus, 𝜇 = 1.36GPa is the shear modulus, and 𝜌 is
joint inversion with seismic and surface gravity data. the density value [17, 33]. Here, we assume that the formation
comprises 80% shale and 20% sandstone to get the values for 𝜅 and
3 PROBLEM STATEMENT 𝜇 [13, 39]. Using this process, we generate 180 density/velocity pairs.
Classical inversion techniques aim to minimize a cost function that For preprocessing, we resample each density and velocity model
measures the difference between the forward response of a given from their original resolution of 440×530×145 to 256×256×128.
subsurface model and the observed data. Let 𝐹 be our forward
operator, m be a subsurface model, and dobs be the observed data.
Then the classical inversion problem can be written as
min ||𝐹 (m) − dobs || 22 . (1)
m
Note that this problem takes a single input (i.e., dobs ) and produces
a single predicted subsurface model.
In contrast, joint inversion is an extension of the classical formu-
lation given by (1) that maps multiple inputs (i.e., gravity, seismic,
and electromagnetic data) to multiple outputs (i.e., density, velocity,
and resistivity models). For a given number of inputs d1obs, . . . , d𝑛obs
and appropriate forward operators 𝐹 1, . . . , 𝐹𝑛 , our joint problem is
given by
𝑛
||𝐹𝑖 (m𝑖 ) − d𝑖obs || 22 + 𝛼
∑︁ ∑︁
min Φ(m𝑖 , m 𝑗 ), (2)
m1 ,...m𝑛 𝑖=1 𝑖≠𝑗
where Φ is a coupling function used to link different physical models
Figure 1: Illustration of a change in surface gravity for a
via known petrophysical relations or other metrics like SSIM, and 𝛼
given density perturbation in the subsurface [8].
is a parameter controlling the contribution of the coupling term [24,
30]. Joint inversion is much more time-consuming than independent
inversions because of the additional terms in the cost function and
the need to exchange information between different models via the 4.1 Modeling Gravity
coupling terms [18]. There also is a need to determine or adjust the Given a density perturbation Δ𝜌 observed between the current and
weighting parameter 𝛼, which adds another layer of complexity to original (i.e., base) acquisition, the gravity field recorded at a station
the standard joint inversion method [18]. located at r′ can be expressed as:
Our goal is to train a CNN that can accurately predict the changes
r − r′
∭
in subsurface density and velocity given corresponding variations g(r′ ) = 𝛾 Δ𝜌 (r)𝑑𝑉 (4)
3
in surface gravity and seismic data. 𝑉 ||r − r′ || 2

where 𝑉 is the volume of the reservoir, and 𝛾 is Newton’s grav-

4 DATA PREPARATION itational constant. For more details on gravity modeling, see [8].
Located 60km off the western coast of Norway, the Johansen for- Assuming that gravity sensors are placed in a uniform grid every
mation is a promising CO2 storage site with a theoretical capacity 500m, our surface gravity maps are size 88×106. An illustration of
Joint inversion of Time-Lapse Surface Gravity and Seismic Data for Monitoring of 3D CO2 Plumes via Deep Learning SC23, Nov. 12–17, 2023, Denver, CO

a change in surface gravity for a given density permutation in the 5 METHODS

subsurface is shown in Figure 1.
5.1 Network Architecture
4.2 Forward Seismic Modeling We use a modified 3D U-Net architecture to map 2D surface gravity
maps and 3D seismic cubes to subsurface changes in density and
Forward seismic modeling approximates the behavior of seismic velocity. Each input is fed into a dedicated encoder branch. The first
waves propagating through a mechanical medium m and is given step for the surface gravity encoder branch is to resize the input
by the elastic wave equation: to match the reservoir geometry via linear interpolation. Then
𝜕 2𝑢 the 2D features are converted into a 3D volume using a pointwise
− V𝑝2 ∇(∇ · 𝑢) − V𝑠2 Δ𝑢 = f, (5) convolution, where the number of channels equals the subsurface
𝜕𝑡 2
model’s depth (i.e., the output). This resulting 3D volume is the
where u = u(𝑥, 𝑦, 𝑧, 𝑡), is the seismic wave displacement, V𝑝 is
input to the first 3D encoder branch of the network. The seismic
P-wave velocity (compression/rarefaction), V𝑠 is S-wave velocity
encoder branch starts with a 3D seismic cube as input. However, the
(shear stress), and f is the perturbation source (i.e., shot) function
seismic cubes are larger in the depth dimension than the network
[15]. While (5) more accurately describes seismic wave propagation,
output. To address this, we use two convolutional layers with strides
it is often preferred (as in this work) to approximate the solution u
1×1×2 to reduce the depth dimension of the seismic cube from 512
by the acoustic wave equation, which assumes only P-waves and
to 128. The resulting resized seismic features are the input to the
requires less computational resources and parameters, as compared
second encoder branch of our architecture. In each resolution level
to solving (5) [15].
of our encoder branch, we apply two convolutional layers and
3D seismic data consumes a large amount of memory and stor-
downsampling via max pooling four times. A convolutional layer
age, up to dozens of TB for the raw data in our case. We compute
consists of two convolutions, each followed by batch normalization
spatial decimation evenly, but temporal, which in seismic signals
and a ReLU non-linearity.
represents depth, decimation favors samples that convey informa-
At the bottleneck of our architecture, we concatenate the en-
tion in the area of interest (i.e., the reservoir). Further, we collapse
coded seismic and gravity features and apply two convolutional
our seismic data by adding the recorded data from each shot and
layers to merge the individual features. We also apply autoencoder
then boost the later time signals by multiplying by a monotoni-
regularization. However, because we have two inputs, we separately
cally increasing function. This method is fully described in [15].
decode the seismic and gravity features (before concatenation and
Finally, we resize each collapsed and boosted seismic cube from its
convolution) to reconstruct their respective inputs.
original resolution of 167×154×941 to 256×256×512 by first taking
Our decoder branch upsamples the combined features and con-
every other slice (i.e., time step) in the z-direction and then linearly
catenates them with the corresponding features from the encoder
interpolating to the final resolution.
branches in the skip connections. Like with the network proposed
Like forward gravity modeling, we look at the differences be-
by [8], we split the output of the decoder branch into three separate
tween the original or base seismic data and the current data. Figure
outputs: two regression branches to predict density and velocity
2 illustrates an example of the change in seismic data from the
and one for segmenting the plume. Figure 3 shows a detailed sketch
original and current acquisitions.
of our proposed architecture.

5.2 Loss Function

Our loss function consists of three components - segmentation,
regression, autoencoder losses.
The segmentation loss is the Generalized Dice Loss (GDL) pro-
posed by Sudre et al. [37]. Unlike the original Dice loss proposed by
[29], the GDL uses weighting terms to account for class imbalance.
This loss function is given by

𝑤𝑘 ||𝑇 𝑘 − 𝑃 𝑘 || 22
Í𝐶
L𝑔𝑑𝑙 = Í 𝑘=1  , (6)
𝐶 𝑤
𝑘=1 𝑘
||𝑇 𝑘 || 22 + ||𝑃 𝑘 || 22

where 𝐶 denotes the number of segmentation classes, 𝑃 𝑘 is the

𝑘-th class in the predicted mask, and 𝑇 𝑘 is the same for the ground
truth. The term 𝑤𝑘 is the !weighting term for the 𝑘 𝑡ℎ class and

is given by 𝑤𝑘 = Í𝐶 1 1
1
𝑁𝑘 . Here, 𝑁𝑘 is the total number of
𝑗 =1 𝑁 𝑗

pixels belonging to the class 𝑘 over the entire dataset. Note that
Figure 2: Example of seismic difference cube. the weights 𝑤𝑘 are pre-computed and remain constant throughout
training. In our case, the number of segmentation classes equals
two; background and foreground. The computed class weights are
SC23, Nov. 12–17, 2023, Denver, CO Celaya and Araya-Polo

Figure 3: Detailed sketch of our proposed architecture for jointly inverting surface gravity and seismic data. Our architecture
uses two separate encoder branches for each input and decodes them jointly. The output of the decoder branch splits into three
separate outputs: two regression branches to predict density and velocity and one for segmenting the plume. Additionally, this
architecture uses autoencoder regularization.

approximately 0.0015 and 0.9985 for the background and foreground use the architecture described in [8] for the portions of our dataset
classes, respectively. corresponding to either gravity or seismic inversion. Note that the
The regression and autoencoder losses use the mean squared models used for individual inversion only output the properties
error loss for their respective inputs. For a general input pair (𝑇 , 𝑃), corresponding to that particular inversion problem. For example,
this loss is given by 𝑁1 ||𝑇 − 𝑃 || 22 . For the regression tasks (i.e., the network used for purely seismic inversion only produces a
density and velocity reconstruction) the inputs to this loss function velocity model as a prediction.
are the true and predicted density and velocity models. For the To evaluate the validity of our predicted inversions, we utilize the
autoencoder branches, the inputs to this loss function are the true following metrics: mean squared error in kg/m3 between the true
and reconstructed gravity and seismic inputs. and predicted density models, mean squared error in m/s between
Our overall loss function is a weighted convex combination of the true and predicted velocity models, mean squared error in
the previously described components and is given by 𝜇Gals between the observed data and the gravity response of the
predicted density model, the R-squared coefficient between the
0.05  𝑔𝑟𝑎𝑣 
L = 0.375L𝜌 + 0.2L𝑔𝑑𝑙 + 0.375L𝑣 + L𝑎𝑒 + L𝑎𝑒 𝑠𝑒𝑖𝑠
, (7) true and predicted models (for density and velocity), and the Dice
2
coefficient between the non-zero masks of the true and predicted
where L𝜌 , L𝑔𝑑𝑙 , L𝑣 , L𝑎𝑒 , and L𝑎𝑒
𝑠𝑒𝑖𝑠 are the density model, seg- plume geometry.
𝑔𝑟𝑎𝑣

mentation, velocity model, gravity autoencoder, and seismic au- Our models are implemented in Python using PyTorch (v2.0.1)
toencoder losses, respectively. Note that we select these weights and trained on four NVIDIA A100 GPUs [34]. At test time, our
via a partial grid search. DL-based methods produce predictions of size 256×256×128. We
resample this output to the original grid resolution of 440×530×145
5.3 Training and Testing Protocols via linear interpolation to produce our final prediction. Given the
To train our neural networks, we use the Adam optimizer [23] with sample size, we develop a data parallelism approach by using Py-
an initial learning rate of 0.001 and a cosine decay schedule with Torch’s Distributed Data Parallel implementation. The in-node scal-
restarts [28]. We train our model to convergence (≈ 400 epochs) ability is nearly ideal, with epochs taking 190 seconds running on
with a batch size of 8. We use 80% of our dataset as a training 1 GPU and ending up in 45 seconds when running on 4 GPUs.
set, use 10% of the training data as a validation set, and use the
remaining 20% of the data as a test set. To compare the effect of
joint inversion vs. inversion with only gravity or seismic data, we
Joint inversion of Time-Lapse Surface Gravity and Seismic Data for Monitoring of 3D CO2 Plumes via Deep Learning SC23, Nov. 12–17, 2023, Denver, CO

R-Squared
Method MSE (kg/m3 ) MSE (m/s) MSE (𝜇Gal) Dice
Density Velocity
Gravity 0.311 (0.215) - 0.601 (1.185) 0.589 (0.034) 0.769 (0.12) -
Seismic - 0.121 (0.100) - 0.579 (0.096) - 0.715 (0.175)
Joint 0.268 (0.174) 0.084 (0.057) 0.634 (1.241) 0.621 (0.038) 0.801 (0.095) 0.800 (0.095)
Table 1: Joint inversion vs. inversion with only gravity or seismic data. Here, we see that our proposed joint inversion generally
outperforms gravity and seismic-only inversion for all metrics except for the mean squared error between the observed data
and the gravity response of the predicted density model, where the results vs. the gravity-only model are comparable.

model in terms of density and velocity reconstruction, segmenta-

tion accuracy, and R-squared coefficients indicates the effectiveness
of the proposed approach in capturing the complexity of subsurface
CO2 plumes.
The benefits of DL-based joint inversion are limited from a com-
putational perspective because combining two different datasets
requires more memory and time during training. Our joint inversion
model takes roughly 45s per epoch on 4 A100 GPUs with a batch
size of 8. In contrast, DL-based gravity inversion takes roughly
20s per epoch, and DL-based seismic inversion takes roughly 30s
per epoch for the same number of GPUs and batch size. Addition-
Figure 4: Training and validation loss curves on a logarith- ally, joint inversion takes just over 400 epochs to converge to a
mic scale for our DL-based joint inversion. Here, both losses solution vs. 200 for both the gravity and seismic-only approaches.
converge, indicating that our proposed joint inversion ar- The greater number of epochs is possibly explained by the more
chitecture successfully learns a mapping from our surface complex relationship our joint inversion approach has to resolve
gravity and seismic data to 3D subsurface density and veloc- vs. the single mode methods. Regarding inference, our joint model
ity models. is comparable to the gravity and seismic-only models, producing
predictions in less than one second on a single A100 GPU.
We utilize the PocketNet approach proposed by [7] in our joint
6 RESULTS architecture. This approach takes advantage of the similarity be-
We train our joint inversion model using the methods described in tween the U-Net architecture and geometric multigrid methods to
Section 5. Figure 4 shows the training and validation loss curves drastically reduce the number of parameters [7, 16]. Additionally,
on a logarithmic scale. This figure shows that both losses converge, we replace the transposed convolution with trilinear upsampling.
indicating that our proposed joint inversion architecture success- With these modifications, we reduce the number of parameters from
fully learns a mapping from our surface gravity and seismic data roughly 33,000,000 to 349,000, yielding a roughly 40% decrease in
to 3D subsurface density and velocity models. the time per epoch. Additionally, these modifications allow us to
In Table 1, we see that our proposed joint inversion generally use a larger batch size (8 instead of 4).
outperforms gravity and seismic-only inversion for all metrics ex- While developing our joint inversion model, we observe that the
cept for the mean squared error between the observed data and optimization landscape is tricky, with some local minimums cor-
the gravity response of the predicted density model, where the responding to good values for density model misfit and R-squared
results vs. the gravity only model are comparable. Figure 5 shows a and vice versa for velocity models. Adding a coupling term like
2D cross-section slice from predictions from each model. Visually, in the classical joint inversion formulation given by 2 may help
the density models produced by the gravity-only and our joint ar- avoid getting trapped in local minimums during training. Future
chitectures are similar. However, visual differences exist between work will focus on formulating and testing a term based on the
the seismic-only reconstructed velocity model and the velocity Dice score. The intuition here is to enforce via our modified loss
model produced from our joint architecture (i.e., top right corner); function that the plumes occupy the same physical space.
those can also be observed in Figure 6 for a different sample (i.e., at Our data comes from simulations of a single geologic formation
different times of plume’s migration). (Johansen). Further work is needed to test the generalization ca-
pability of our joint inversion model to other datasets based on
7 DISCUSSION simulations of other CO2 storage sites (i.e., Snohvit and Kimberlina
Our results demonstrate the potential benefits of DL-based joint [2]) and on field data. However, time-lapse CCUS monitoring with
inversion. The joint inversion model consistently outperforms DL- gravity (and other non-seismic methods) is a data-poor field. Suf-
based gravity-only and seismic-only inversion models across vari- ficiently detailed reservoir models for CCUS, especially field data,
ous evaluation metrics. This performance suggests that the fusion are hard to come by, and access is limited [2, 8].
of surface gravity and seismic data can lead to more accurate sub-
surface models. The improved performance of the joint inversion
SC23, Nov. 12–17, 2023, Denver, CO Celaya and Araya-Polo

Figure 5: From left to right 2D cross-section slice from ground truth, gravity-only, seismic-only, and joint models. The top row
shows density models, and the bottom row shows velocity models. Visually, the density models produced by the gravity-only
and our joint architectures are similar. However, visual differences exist between the seismic-only reconstructed velocity model
and the velocity model produced from our joint architecture (i.e., top right corner).

8 CONCLUSIONS
We developed an effective DL-based joint inversion method to re-
cover high-resolution, subsurface CO2 density and velocity models
from surface gravity and seismic data. We train our joint DL archi-
tecture on realistic, physics-simulated CO2 plumes, surface gravity,
and seismic data. This training approach mirrors real-world site
data collection. Our joint inversion outperforms gravity and seismic-
only inversion techniques for our selected metrics. While there is
room for improvement, the results presented here are promising
and represent, to the best of our knowledge, the first fully 3D DL-
based joint inversion of surface gravity and seismic data derived
from a physics simulation of a proposed CO2 storage site.

ACKNOWLEDGMENTS
This work was supported by TotalEnergies EP Research & Technol-
ogy USA, LLC.

REFERENCES
[1] Amir Adler, Mauricio Araya-Polo, and Tomaso Poggio. 2021. Deep Learning
for Seismic Inverse Problems: Toward the Acceleration of Geophysical Analysis
Workflows. IEEE Signal Processing Magazine 38, 2 (2021), 89–119. https://doi.
org/10.1109/MSP.2020.3037429
[2] David Alumbaugh, Erika Gasperikova, Dustin Crandall, Michael Commer, Shi-
hang Feng, William Harbert, Yaoguo Li, Youzuo Lin, and Savini Samarasinghe.
2023. The Kimberlina synthetic multiphysics dataset for CO2 monitoring investi-
Figure 6: 3D view of label (top) -prediction (bottom) pair of gations. Geoscience Data Journal (2023).
velocity variation given the CO2 plume location. [3] Taqi Alyousuf, Yaoguo Li, and Richard Krahenbuhl. 2022. Machine
learning inversion of time-lapse three-axis borehole gravity data for CO2
monitoring. 3099–3103. https://doi.org/10.1190/image2022-3745388.1
arXiv:https://library.seg.org/doi/pdf/10.1190/image2022-3745388.1
[4] Mauricio Araya-Polo, Joseph Jennings, Amir Adler, and Taylor Dahlke. 2018.
Deep-learning tomography. The Leading Edge 37, 1 (2018), 58–66. https://doi.
org/10.1190/tle37010058.1 arXiv:https://doi.org/10.1190/tle37010058.1
Joint inversion of Time-Lapse Surface Gravity and Seismic Data for Monitoring of 3D CO2 Plumes via Deep Learning SC23, Nov. 12–17, 2023, Denver, CO

[5] Spyridon Bakas et al. 2018. Identifying the best machine learning algorithms 2016 fourth international conference on 3D vision (3DV). IEEE, 565–571.
for brain tumor segmentation, progression assessment, and overall survival [30] Max Moorkamp, Björn Heincke, Marion Jegen, Alan W Roberts, and Richard W
prediction in the BRATS challenge. arXiv preprint arXiv:1811.02629 (2018). Hobbs. 2011. A framework for 3-D joint inversion of MT, gravity and seismic
[6] O. Boulanger and M. Chouteau. 2001. Constraints in 3D gravity inversion. refraction data. Geophysical Journal International 184, 1 (2011), 477–493.
Geophysical Prospecting 49, 2 (2001), 265–280. https://doi.org/10.1046/j.1365- [31] MN Nabighian, VJS Grauch, RO Hansen, TR LaFehr, Y Li, JW Peirce, JD Phillips,
2478.2001.00254.x and ME Ruder. 2005. 75th Anniversary. The historical development of the magnetic
[7] Adrian Celaya, Jonas A. Actor, Rajarajesawari Muthusivarajan, Evan Gates, Car- method in exploration: Geophysics 70, 6 (2005).
oline Chung, Dawid Schellingerhout, Beatrice Riviere, and David Fuentes. 2023. [32] Seokmin Oh, Kyubo Noh, Soon Jee Seol, and Joongmoo Byun. 2020. Coopera-
PocketNet: A Smaller Neural Network for Medical Image Analysis. IEEE Trans- tive deep learning inversion of controlled-source electromagnetic data for salt
actions on Medical Imaging 42, 4 (2023), 1172–1184. https://doi.org/10.1109/TMI. delineation. GEOPHYSICS 85, 4 (2020), E121–E137.
2022.3224873 [33] William G Pariseau. 2017. Design analysis in rock mechanics. CRC Press.
[8] Adrian Celaya, Bertrand Denel, Yen Sun, Mauricio Araya-Polo, and Antony Price. [34] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory
2023. Inversion of Time-Lapse Surface Gravity Data for Detection of 3-D CO2 Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban
Plumes via Deep Learning. IEEE Transactions on Geoscience and Remote Sensing Desmaison, Andreas Köpf, Edward Z. Yang, Zach DeVito, Martin Raison, Alykhan
61 (2023), 1–11. https://doi.org/10.1109/TGRS.2023.3273149 Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith
[9] Jie Chen, Cara Schiek-Stewart, Ligang Lu, Susanne Witte, Karin Eres Guardia, Chintala. 2019. PyTorch: An Imperative Style, High-Performance Deep Learning
Francesco Menapace, Pandu Devarakota, and Mohamed Sidahmed. 2020. Machine Library. CoRR abs/1912.01703 (2019). arXiv:1912.01703 http://arxiv.org/abs/1912.
learning method to determine salt structures from gravity data. In SPE Annual 01703
Technical Conference and Exhibition. OnePetro. [35] Mohammad Rezaie, Ali Moradzadeh, and Ali Nejati Kalateh. 2017. Fast 3D
[10] Daniele Colombo, Weichang Li, Ernesto Sandoval-Curiel, and Gary W. McNeice. inversion of gravity data using solution space priorconditioned lanczos bidiago-
2020. Deep-learning electromagnetic monitoring coupled to fluid flow simulators. nalization. Journal of Applied Geophysics 136 (2017), 42–50. https://doi.org/10.
GEOPHYSICS 85, 4 (2020), WA1–WA12. https://doi.org/10.1190/geo2019-0428.1 1016/j.jappgeo.2016.10.019
arXiv:https://doi.org/10.1190/geo2019-0428.1 [36] Mostafa Shahriari, David Pardo, Artzai Picón, Adrian Galdran, Javier Del Ser, and
[11] Catherine De Groot-Hedlin and SC Constable. 1990. OCCAM’s inversion to gen- Carlos Torres-Verdín. 2020. A deep learning approach to the inversion of borehole
erate smooth, two-dimensional models from magnetotelluric data. GEOPHYSICS resistivity measurements. Computational Geosciences 24, 3 (2020), 971–994.
55 (12 1990), 1613–1624. https://doi.org/10.1190/1.1442813 [37] Carole H. Sudre, Wenqi Li, Tom Kamiel Magda Vercauteren, Sébastien Ourselin,
[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Im- and M. Jorge Cardoso. 2017. Generalised Dice Overlap as a Deep Learning Loss
ageNet: A large-scale hierarchical image database. In 2009 IEEE Conference on Function for Highly Unbalanced Segmentations. In Deep Learning in Medical
Computer Vision and Pattern Recognition. 248–255. https://doi.org/10.1109/CVPR. Image Analysis and Multimodal Learning for Clinical Decision Support. Springer
2009.5206848 International Publishing, 240–248.
[13] G. Eigestad, Helge Dahle, Bjarte Hellevang, Fridtjof Riis, Wenche Johansen, and [38] Yen Sun, Bertrand Denel, Norman Daril, Lory Evano, Paul Williamson, and
Erlend Øian. 2009. Geological modeling and simulation of CO2 injection in the Mauricio Araya-Polo. 2020. Deep learning joint inversion of seismic and elec-
Johansen formation. Computational Geosciences 13 (12 2009), 435–450. https: tromagnetic data for salt reconstruction. SEG Technical Program Expanded
//doi.org/10.1007/s10596-009-9153-y Abstracts 2020 (2020), 550–554. https://doi.org/10.1190/segam2020-3426925.1
[14] Manzar Fawad and Nazmul Haque Mondol. 2021. Monitoring geological storage arXiv:https://library.seg.org/doi/pdf/10.1190/segam2020-3426925.1
of CO2: A new approach. Scientific Reports 11, 1 (2021), 5942. [39] Anja Sundal, Johan Petter Nystuen, Kari-Lise Rørvik, Henning Dypvik, and Per
[15] Maayan Gelboim, Amir Adler, Yen Sun, and Mauricio Araya-Polo. 2022. Encoder– Aagaard. 2016. The Lower Jurassic Johansen Formation, northern North Sea –
Decoder Architecture for 3D Seismic Inversion. Sensors 23, 1 (2022), 61. Depositional model and reservoir characterization for CO2 storage. Marine and
[16] Juncai He and Jinchao Xu. 2019. MgNet: A unified framework of multigrid and Petroleum Geology 77 (2016), 1376–1401. https://doi.org/10.1016/j.marpetgeo.
convolutional neural network. Science china mathematics 62 (2019), 1331–1354. 2016.01.021
[17] Evert Hoek and Jonathan D Bray. 1981. Rock slope engineering. CRC press. [40] Albert Tarantola and Bernard Valette. 1981. Inverse problems = Quest for infor-
[18] Yanyan Hu, Xiaolong Wei, Xuqing Wu, Jiajia Sun, Jiuping Chen, Yueqin Huang, mation. Journal of Geophysics 50, 1 (October 1981), 159–170.
and Jiefu Chen. 2023. A deep learning-enhanced framework for multiphysics joint [41] Evan Schankee Um, David Alumbaugh, Michael Commer, Shihang Feng,
inversion. GEOPHYSICS 88, 1 (2023), K13–K26. https://doi.org/10.1190/geo2021- Erika Gasperikova, Yaoguo Li, Youzuo Lin, and Savini Samarasinghe.
0589.1 arXiv:https://doi.org/10.1190/geo2021-0589.1 2022. Deep-learning multiphysics network for imaging CO2 satura-
[19] Rui Huang, Shuang Liu, Rui Qi, and Yujie Zhang. 2021. Deep Learning 3D Sparse tion and estimating uncertainty in geological carbon storage. Geo-
Inversion of Gravity Data. Journal of Geophysical Research: Solid Earth 126, 11 physical Prospecting (2022). https://doi.org/10.1111/1365-2478.13257
(2021), e2021JB022476. arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1111/1365-2478.13257
[20] Kyong Hwan Jin, Michael T. McCann, Emmanuel Froustey, and Michael Unser. [42] Fangshu Yang and Jianwei Ma. 2019. Deep-learning inversion: A next-generation
2017. Deep Convolutional Neural Network for Inverse Problems in Imaging. seismic velocity model building method. GEOPHYSICS 84, 4 (2019), R583–R599.
IEEE Transactions on Image Processing 26, 9 (2017), 4509–4522. https://doi.org/10. https://doi.org/10.1190/geo2018-0249.1 arXiv:https://doi.org/10.1190/geo2018-
1109/TIP.2017.2713099 0249.1
[21] Tero Karras, Samuli Laine, and Timo Aila. 2019. A style-based generator ar- [43] Qianguo Yang, Xiangyun Hu, Shuang Liu, Qu Jie, Huaijiang Wang, and Qiuhua
chitecture for generative adversarial networks. In Proceedings of the IEEE/CVF Chen. 2022. 3-D Gravity Inversion Based on Deep Convolution Neural Networks.
conference on computer vision and pattern recognition. 4401–4410. IEEE geoscience and remote sensing letters 19 (2022), 1–5.
[22] Yuji Kim and Nori Nakata. 2018. Geophysical inversion versus machine learning [44] Wang Yu-Feng, Zhang Yu-Jie, Fu Li-Hua, and Li Hong-Wei. 2021. Three-
in inverse problems. The Leading Edge 37, 12 (2018), 894–901. https://doi.org/10. dimensional gravity inversion based on 3D U-Net++. Applied Geophysics 18,
1190/tle37120894.1 arXiv:https://doi.org/10.1190/tle37120894.1 4 (2021), 451–460.
[23] Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti-
mization. arXiv preprint arXiv:1412.6980 (2014).
[24] Peter G Lelièvre, Colin G Farquharson, and Charles A Hurich. 2012. Joint inver-
sion of seismic traveltimes and gravity data on unstructured grids with application
to mineral exploration. Geophysics 77, 1 (2012), K1–K15.
[25] Shucai Li, Bin Liu, Yuxiao Ren, Yangkang Chen, Senlin Yang, Yunhai Wang, and
Peng Jiang. 2020. Deep-Learning Inversion of Seismic Data. IEEE Transactions on
Geoscience and Remote Sensing 58, 3 (2020), 2135–2149. https://doi.org/10.1109/
TGRS.2019.2953473
[26] Yaoguo Li and Douglas W. Oldenburg. 1998. 3-D inversion of gravity
data. GEOPHYSICS 63, 1 (1998), 109–119. https://doi.org/10.1190/1.1444302
arXiv:https://doi.org/10.1190/1.1444302
[27] Bin Liu et al. 2020. Deep Learning Inversion of Electrical Resistivity Data. IEEE
Transactions on Geoscience and Remote Sensing 58, 8 (2020), 5715–5728. https:
//doi.org/10.1109/TGRS.2020.2969040
[28] Ilya Loshchilov and Frank Hutter. 2017. SGDR: Stochastic Gradient Descent
with Warm Restarts. In International Conference on Learning Representations.
https://openreview.net/forum?id=Skq89Scxx
[29] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. 2016. V-net: Fully
convolutional neural networks for volumetric medical image segmentation. In