aaaifss2022_14.txt

De-risking Carbon Capture and Sequestration with Explainable CO 2Leakage
Detection in Time-lapse Seismic Monitoring Images
Huseyin Tuna Erdinc,*1Abhinav Prakash Gahlot,*2Ziyi Yin,3
Mathias Louboutin,2Felix J. Herrmann,1,2,3
1School of Electrical and Computer Engineering, Georgia Institute of Technology
2School of Earth and Atmospheric Sciences, Georgia Institute of Technology
3School of Computational Science and Engineering, Georgia Institute of Technology
{herdinc3, agahlot8, ziyi.yin, mlouboutin3, felix.herrmann }@gatech.edu,
Abstract
With the growing global deployment of carbon capture and
sequestration technology to combat climate change, monitor-
ing and detection of potential CO 2leakage through existing
or storage induced faults are critical to the safe and long-term
viability of the technology. Recent work on time-lapse seis-
mic monitoring of CO 2storage has shown promising results
in its ability to monitor the growth of the CO 2plume from
surface recorded seismic data. However, due to the low sen-
sitivity of seismic imaging to CO 2concentration, additional
developments are required to efficiently interpret the seis-
mic images for leakage. In this work, we introduce a binary
classification of time-lapse seismic images to delineate CO 2
plumes (leakage) using state-of-the-art deep learning models.
Additionally, we localize the leakage region of CO 2plumes
by leveraging Class Activation Mapping methods.
Introduction
According to the International Energy Agency and the In-
ternational Panel on Climate Change report (IPCC 2018),
there is a need for a 50 percent reduction of greenhouse
gas emissions by 2050 to avoid an increase of 1.5 degrees
Celsius of Earth’s average temperature. This can only be
achieved by reduced dependence on fossil fuels, use of re-
newable sources of energy and large-scale global deploy-
ment of carbon reduction technologies such as carbon cap-
ture and sequestration (CCS). This technology consists of
collection, transportation, and injection of CO 2into an ap-
propriate geologic storage reservoir for extended time peri-
ods (tens of years). Especially, unlike other solutions, CCS
is considered a relatively low-cost, long-term and imminent
solution. However, potential CO 2leakage from the under-
ground reservoirs due to pre-existing or pressure-induced
faults poses risks (Ringrose 2020). Thus, it is necessary to
de-risk CCS projects by monitoring CO 2plumes in order to
accurately detect and predict potential leakages as early as
possible.
Time-lapse seismic monitoring has been introduced as
a reliable technology to monitor the CO 2dynamics in
the Earth’s subsurface during carbon sequestration (Lumley
*These authors contributed equally.
Copyright © 2022, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.2001) and is already in use at existing storage sites (Arts
et al. 2008; Chadwick et al. 2010; Ringrose et al. 2013; Furre
et al. 2017). In essence, sequential (i.e once every 6 month-
s/year/...) seismic datasets, called vintages, are collected in
the field over an area covering the storage reservoir. Then,
each seismic dataset is inverted to obtain high fidelity im-
ages of the subsurface over time (Arts et al. 2008; Ayeni
and Biondi 2010; Yin, Louboutin, and Herrmann 2021). The
evolution of the CO 2reservoir can finally be visualized by
subtracting the seismic images between different points in
time. However, due to the inherently weak and noisy ampli-
tudes of the CO 2reservoir’s response in those seismic im-
ages, detecting the presence of potential irregularities, such
as in CO 2plumes, corresponding to a leakage is a challeng-
ing problem. To tackle this difficulty, we propose a machine
learning based detection method based on standard binary
classification.
Recently, numerous methods leveraging machine learn-
ing have been introduced for the detection of CO 2leakage
based on a simple artificial neural network (ANN) (Li et al.
2018a), and a combination of convolutional neural networks
(CNN) and Long Short-Term Memory (LSTM) networks
(Zhou et al. 2019). While leading to accurate predictions,
these methods usually rely solely on the field recorded data
rather than the subsurface seismic images. Besides, practical
considerations such as repeatability (the ability to record the
data in the exact same way every year) hinders their appli-
cability to real world cases. On the other hand, as we rely
on visualizing the CO 2plumes in the seismic image, we
can take advantage of advanced seismic imaging techniques
designed for non-repeated seismic acquisition such as the
joint recovery model (JRM) (Oghenekohwo and Herrmann
2017a; Wason, Oghenekohwo, and Herrmann 2017; Yin,
Louboutin, and Herrmann 2021). Additionally, this imag-
ing technique has demonstrated higher fidelity imaging than
sequential seismic imaging allowing for easier detection of
CO2leakage.
We will show in the following sections that we can effi-
ciently and accurately detect CO 2from realistic seismic im-
ages recovered by JRM on synthetic but representative mod-
els of the Earth subsurface. We demonstrate our method us-
ing different state-of-the-art deep learning models in a trans-
fer learning setting to classify CO 2plume seismic images

with regular (no-leakage) CO 2plume or with CO 2leakage.
As CO 2leakage detection needs trustworthiness, we further
unravel the decisions made by our models and utilize Class
Activation Mapping (CAM) methods (Zhou et al. 2015) to
identify and visualize seismic image areas crucial for model
classification results. We show that the CAM result accu-
rately focuses on the leakage portion of the CO 2plume and
reservoir, validating that our network detects leakage based
on state of the CO 2reservoir over time.
Our main contributions are organized as follows. First, we
introduce the classification models used for leakage detec-
tion and the CAM methods for visualizing the area of inter-
est in the seismic image. Second, we demonstrate the accu-
racy of our models and qualitatively examine the results of
our CAM methods on a realistic synthetic set of CO 2plume
images.
Methodology
In order to speed up the training process and to compensate
for the overfitting that may occur with modest sized datasets,
we rely on transfer learning (Yosinski et al. 2014) using pre-
trained state-of-the-art models as a starting point. In particu-
lar, we consider four modern architectures known to achieve
high accuracy on standard dataset such as ImageNet-1k
(Russakovsky et al. 2015). The models used are VGG (Si-
monyan and Zisserman 2014), ResNet (He et al. 2016), Vi-
sion Transformer (ViT) (Dosovitskiy et al. 2021), and Swin
Transformer (Swin) (Liu et al. 2021), all pre-trained on the
standardized ImageNet-1k dataset.
VGG: is a convolutional neural network (CNN) model
that achieved significant success in The ImageNet Large
Scale Visual Recognition Challenge (ILSVRC) competition
in 2014 (Simonyan and Zisserman 2014). VGG consists of
sequences of convolution and maxpool layers. In our numer-
ical experiments, the VGG16 variant with 16 trainable layers
is used.
ResNet: is a CNN architecture with residual connections
proposed to solve the vanishing gradient problem in very
deep networks (He et al. 2016). ResNet consists of resid-
ual blocks and each residual block has convolution layers
and shortcut connections performing identity mapping. In
our numerical experiments, the ResNet34 variant with 34
trainable layers is used.
ViT: is an architecture based on transformer which is used
in the field of Natural Language Processing (NLP) (Vaswani
et al. 2017). Internally, the transformer learns a relationship
between input token pairs, and uses 16x16 patches of im-
ages as input tokens (Dosovitskiy et al. 2021). In our numer-
ical experiments, the tiny ViT variant is used allowing lower
memory and computational imprint.
Swin: is a special type of ViT that represents image
patches hierarchically by starting from small-sized patches
and gradually increasing the size through merging to achieve
scale-invariance property (Liu et al. 2021). Compared to
ViT, Swin transformer has superior (linear) computational
efficiency by computing self-attention within certain patches
of a window. In our numerical experiments, tiny Swin vari-
ant is used allowing lower memory and computational im-
print.Hyperparameters VGG16 ResNet34 ViT Swin
Batch Size 8 8 8 8
Learning Rate 5x10−56x10−34x10−310−3
Exp Decay Rate( γ)0.95 0 .92 0 .98 0 .98
Table 1: Training hyperparameters for the four models. All models
were trained with the same number of epochs and optimizer.
CAM Methods
Deep learning models for classification are notoriously
treated as “black boxes” as they do not expose their inter-
nal knowledge or operations to its users and do not pro-
vide interpretable results. To solve this problem, CAM based
saliency maps (heatmaps) were introduced to highlight the
most class-discriminative regions of to-be-classified input
images (Zhou et al. 2015). Since CO 2leakage requires high
fidelity, transparent and interpretable models, we use CAM
to further make our model results explainable and highlight
the regions of the seismic image that are most relevant to
the classification results. In our study, we considered two
CAM methods. First, Grad-CAM (Selvaraju et al. 2019), a
gradient-based CAM method considered as the state-of-the-
art in terms of explainability of neural networks for classi-
fication. This CAM method extracts gradients from a spe-
cific layer of a model and computes the weighted aver-
age of that specific layer’s activations. Second, we consider
Score-CAM (Wang et al. 2020), a perturbation based CAM
method. Score-CAM also computes the weighted average of
activations of a user-specified layer but, unlike Grad-CAM,
Score-CAM relies on propagating (forward pass through the
network) a masked input image where the mask is obtained
via upsampling the activations of the user-defined layer.
This CAM method provides high accuracy and interpretable
heatmaps and alleviates potential noise and spread present
in the gradient used for the Grad-CAM heatmaps.
Numerical Case Study
To generate the training dataset of CO 2plume evolution, we
used five 2D vertical slices extracted from the 3D Compass
velocity model (E. Jones et al. 2012) shown in Fig. 1(a). This
model is a synthetic but realistic model representative of the
complex geology of the southeast of the North Sea. The di-
mension of each model (slice) used in our work is 2131 X
4062 m2. We used FwiFlow (Li et al. 2020), to simulate the
CO2flow dynamics and JUDI (Witte et al. 2019) to model
the seismic data and compute the seismic images of the sub-
surface.
Time-lapse reservoir and seismic simulation
We consider a realistic two well setting with a fixed injection
well injecting CO 2and a production well extracting brine
from subsurface storage reservoir. Injection of supercriti-
cal CO 2into saline aquifers is an example of multi-phase
flow in porous media. While we understand more compli-
cated geothermal, geochemical and geomechanical process
may eventually be considered to model the CO 2dynamics,
we follow the two-phase immiscible incompressible flow

Figure 1: Workflow for CO 2Leakage Monitoring
physics, which in its leading order describes the process
of supercritical CO 2displacing brine in the pore space of
the rock. The system is governed by conservation of mass
and Darcy’s law. We refer to the existing literature (Li et al.
2020; Wen, Tang, and Benson 2021) (Li et al. 2020) for more
details about this physical system.
Using empirical relation and the Kozeny-Carman equa-
tion(Costa 2006), the acoustic properties (velocity and den-
sity) from the Compass model were converted into perme-
ability and porosity (Fig. 1(b)) to simulate the multi-phase
flow (CO 2and brine in porous media) in the reservoir. We
used FwiFlow.jl (Li et al. 2020) to solve multi-phase flow
equations based on the finite volume method. We simulated
the CO 2flow for a duration varying between 7to12years
(Fig. 1(c)). The reservoir was initially filled with saline wa-
ter and we injected compressed CO 2at the rate of 1MT/-
day into the reservoir for all simulations. In order to mimic
CO2leakage, we then created a fracture at a random location
along the top seal of the reservoir when the pressure induced
by the CO 2injection reaches a threshold of 15MPa. We
then converted back these simulated CO 2saturation snap-
shots over time into wave properties with the patchy sat-
uration model (Avseth, Mukerji, and Mavko 2010) to ob-
tain time-lapse subsurface models (Fig. 1(d)). We used this
model because at higher pressure condition, local fluid flow
slows down resulting in an acoustic velocity trend which fol-
lows patchy saturation (Li et al. 2018b).
Based on these models, we then simulated the baseline
seismic survey corresponding to the initial stage (before the
injection of CO 2) and the monitor seismic survey corre-
sponding to the final stage at the end of the reservoir sim-
ulation (Fig. 1(e)). As mentioned in the introduction, it is
very difficult to exactly replicate the baseline and moni-
tor surveys. In order to mimic the realistic scenario in the
field, the baseline and monitor datasets were simulated us-
ing different acquisition geometries (position of the mea-
surements). Finally, we recovered the time-lapse seismic im-
ages using JRM (Oghenekohwo and Herrmann 2017b; Wa-
son, Oghenekohwo, and Herrmann 2017; Yin, Louboutin,
and Herrmann 2021) to alleviate potential noise and inac-
curacies in the seismic images in the case of non-replicatedtime-lapse surveys. These recovered images along with the
label (leakage/no-leakage) serve as the input to the classi-
fication network. We generated a total of 1000 leakage and
870no-leakage scenarios, and computed the baseline, moni-
tor and difference images with the JRM method in each case.
Training
The seismic difference images (difference between baseline
and monitor recovery results) were converted to 224x224
gray-scale images with bi-linear interpolation and trans-
formed into three channel images where each channel is a
copy of the actual gray-scale image. For the classification,
the image dataset was randomly split into an 80% train-
ing set and 20% test set. The training set was then further
divided into two parts, one for model parameter updating
(training) and another for hyperparameter tuning (valida-
tion). The training hyperparameters from this second part
are summarized in Table 1. For training, we replaced the last
fully connected layers (classification layers) of each model
with a new fully connected layer. We then trained the net-
work (Fig. 1(g)) in two steps. First, we only trained the
last classification layer, by freezing all the other layers, for
100 epochs. Since most of the layers are fixed and do not
need gradient updates, this first stage is extremely cheap and
computationally efficient. Second, we further trained the full
model for an additional 30 epochs to allow fine-tuning of all
layers for our specific classification task. Following standard
practices in classification settings, we used a binary cross-
entropy loss function and the Adam optimizer (Kingma and
Ba 2015) for all models. Finally, after the training (Fig. 1
(h)), we implemented the CAM based methods (Fig. 1 (i)).
We used the last convolutional layer activations for the CNN
models, and the activations preceding the last attention layer
for the transformer-based models.
Analysis
We show on Table 2, different performance metrics on our
testing dataset, after training our four networks, with means
and confidence intervals after 15 different runs. In detail, we
show standard metrics such as accuracy, precision, and re-
call. Additionally, we also show F1 score (Chinchor 1992),

Model Accuracy Precision Recall F1 ROC-AUC
VGG16 0.920 ±0.089 0.941 ±0.133 0.921 ±0.081 0.927 ±0.075 0.920 ±0.076
ResNet34 0.948±0.020 0.982 ±0.028 0.928 ±0.044 0.948 ±0.040 0.967 ±0.019
ViT 0.857 ±0.018 0.910 ±0.102 0.820 ±0.098 0.859 ±0.036 0.923 ±0.023
Swin 0.836 ±0.036 0.881 ±0.108 0.818 ±0.078 0.841 ±0.076 0.909 ±0.007
Table 2: Comparison of performance (for precision and recall, positives represent leakage whereas negatives are no leakage) on the test
dataset for our four neural networks. The highest performance for each metric is highlighted in bold.
Figure 2: Grad-CAM and Score-CAM saliency maps overlayed on the corresponding input seismic image containing a CO 2plume from
leakage. The CO 2plume can be seen on the seismic image as the high amplitude event at 1.3km depth and 1.8km in X.
that combine recall and accuracy, and area under curve of re-
ceiver operating characteristic (ROC-AUC) (Bradley 1997)
to further evaluate the classification performance of mod-
els. We observe in Table 2 that the CNN models outperform
the transformer variants in all the metrics by a significant
margin and that ResNet34 achieves the best performance in
all the measures of evaluation. This result is consistent with
the literature, hinting that despite being very accurate on a
specific task, transformers do not generalize well with our
modest sized dataset (Dosovitskiy et al. 2021). Additionally,
we observe that all models lead to better precision compared
to recall (more false negatives than false positives). This dis-
crepancy can be attributed to the fact that certain leakage
images have very small CO 2leakage areas (up to a single
pixel) in the seismic images and are consequently very diffi-
cult to detect.
Second, we show in Fig. 2 the CAM results of each model
on a single seismic image from our test dataset. The high
amplitude area shows the regions of the seismic images
that are most important to the classifier. As expected, those
heatmaps provide an explainable representation of the clas-
sification as the high amplitudes align with the CO 2leakage
part of the seismic image. We observe that for the CNN, the
saliency maps are well centered on the CO 2leakage por-
tion despite being very coarse. Because of this coarseness,
both Grad-CAM and score-CAM provide similar results. On
the other hand, transformer-based networks lead to more fo-cused saliency maps that target the location of the CO 2leak-
age extremely well. We observe in that case, the Score-CAM
leads to reduction of aliases and noise compared to the Grad-
CAM results. This can be linked to the potential presence of
noise in the gradients of the transformers as the networks are
very deep (Wang et al. 2020).
Conclusion
We have introduced an interpretable deep-learning method
for CO 2leakage detection with very high accuracy on a
synthetic but realistic model of a CO 2sequestration reser-
voir. First, we showed through four state-of-the-art models
that we can detect potential CO 2leakage from the recov-
ered time-lapse seismic images. Second, we demonstrated
that CAM provides an interpretable and accurate visual-
ization of the CO 2plume in case of leakage. Addition-
ally, we showed that transformer-based models (ViT, Swin)
led to more focused CAM and that Score-CAM provided
cleaner and therefore more explainable heatmaps. On the
other hand, we found that standard CNNs led to better classi-
fication results and therefore better leakage detection. In par-
ticular, ResNet model performed best and achieved a very
high score above 90% in every evaluation metric. Future
work will focus on improving the classification network to
achieve higher accuracy in leakage detection and on refining
the heatmaps for better explainability.

Acknowledgments
This research was carried out with the support of Georgia
Research Alliance and partners of the ML4Seismic Center.
The authors thank Philipp A. Witte at Microsoft for the con-
structive discussion.
References
Arts, R. J.; Chadwick, A.; Eiken, O.; Thibeau, S.; and
Nooner, S. L. 2008. Ten years’ experience of monitoring
CO2 injection in the Utsira Sand at Sleipner, offshore Nor-
way. First Break , 26.
Avseth, P.; Mukerji, T.; and Mavko, G. 2010. Quantitative
seismic interpretation: Applying rock physics tools to reduce
interpretation risk . Cambridge university press.
Ayeni, G.; and Biondi, B. 2010. Target-oriented joint least-
squares migration/inversion of time-lapse seismic data sets.
Geophysics , 75.
Bradley, A. P. 1997. The use of the area under the ROC curve
in the evaluation of machine learning algorithms. Pattern
Recognition , 30(7): 1145–1159.
Chadwick, A.; Williams, G.; Delepine, N.; Clochard, V .; La-
bat, K.; Sturton, S.; Buddensiek, M.-L.; Dillen, M.; Nickel,
M.; Lima, A. L.; Arts, R.; Neele, F.; and Rossi, G. 2010.
Quantitative analysis of time-lapse seismic monitoring data
at the Sleipner CO2 storage operation. The Leading Edge ,
29(2): 170–177.
Chinchor, N. 1992. MUC-4 Evaluation Metrics. In Pro-
ceedings of the 4th Conference on Message Understanding ,
MUC4 ’92, 22–29. USA: Association for Computational
Linguistics. ISBN 1558602739.
Costa, A. 2006. Permeability-porosity relationship: A reex-
amination of the Kozeny-Carman equation based on a frac-
tal pore-space geometry assumption. Geophysical Research
Letters , 33(2).
Dosovitskiy, A.; Beyer, L.; Kolesnikov, A.; Weissenborn,
D.; Zhai, X.; Unterthiner, T.; Dehghani, M.; Minderer, M.;
Heigold, G.; Gelly, S.; Uszkoreit, J.; and Houlsby, N. 2021.
An Image is Worth 16x16 Words: Transformers for Image
Recognition at Scale. ICLR .
E. Jones, C.; A. Edgar, J.; I. Selvage, J.; and Crook, H. 2012.
Building Complex Synthetic Models to Evaluate Acquisi-
tion Geometries and Velocity Inversion Technologies. Eu-
ropean Association of Geoscientists & Engineers , cp-293-
00580.
Furre, A.-K.; Eiken, O.; Alnes, H.; Vevatne, J. N.; and
Kiær, A. F. 2017. 20 Years of Monitoring CO2-injection
at Sleipner. Energy Procedia , 114: 3916–3926. 13th Inter-
national Conference on Greenhouse Gas Control Technolo-
gies, GHGT-13, 14-18 November 2016, Lausanne, Switzer-
land.
He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep Resid-
ual Learning for Image Recognition. In Proceedings of the
IEEE conference on computer vision and pattern recogni-
tion,, 770–778.
IPCC. 2018. Global warming of 1.5°C. An IPCC Special
Report on the impacts of global warming of 1.5°C abovepre-industrial levels and related global greenhouse gas emis-
sion pathways, in the context of strengthening the global re-
sponse to the threat of climate change, sustainable develop-
ment, and efforts to eradicate poverty. In Press .
Kingma, D. P.; and Ba, J. 2015. Adam: A Method for
Stochastic Optimization. arXiv preprint arXiv:1412.6980.
Li, B.; Zhou, F.; Li, H.; Duguid, A.; Que, L.; Xue, Y .; and
Tan, Y . 2018a. Prediction of CO2 leakage risk for wells in
carbon sequestration fields with an optimal artificial neural
network. International Journal of Greenhouse Gas Control ,
68: 276–286.
Li, D.; Wei, J.; Di, B.; Ding, P.; Huang, S.; and Shuai, D.
2018b. Experimental study and theoretical interpretation of
saturation effect on ultrasonic velocity in tight sandstones
under different pressure conditions. Geophysical Journal
International , 212: 2226–2237.
Li, D.; Xu, K.; Harris, J. M.; and Darve, E. 2020. Cou-
pled Time-Lapse Full-Waveform Inversion for Subsurface
Flow Problems Using Intrusive Automatic Differentia-
tion. Water Resources Research , 56(8): e2019WR027032.
E2019WR027032 10.1029/2019WR027032.
Liu, Z.; Lin, Y .; Cao, Y .; Hu, H.; Wei, Y .; Zhang, Z.; Lin, S.;
and Guo, B. 2021. Swin Transformer: Hierarchical Vision
Transformer using Shifted Windows. In ICCV .
Lumley, D. E. 2001. Time-lapse seismic reservoir monitor-
ing.GEOPHYSICS , 66(1): 50–53.
Oghenekohwo, F.; and Herrmann, F. J. 2017a. Highly re-
peatable time-lapse seismic with distributed Compressive
Sensing–-mitigating effects of calibration errors. The Lead-
ing Edge , 36(8): 688–694. (The Leading Edge).
Oghenekohwo, F.; and Herrmann, F. J. 2017b. Improved
time-lapse data repeatability with randomized sampling and
distributed compressive sensing. In EAGE Annual Confer-
ence Proceedings . (EAGE, Paris).
Ringrose, P. 2020. How to store CO2 underground: Insights
from early-mover CCS Projects , volume 129. Springer.
Ringrose, P.; Mathieson, A.; Wright, I.; Selama, F.; Hansen,
O.; Bissell, R.; Saoula, N.; and Midgley, J. 2013. The In
Salah CO2 Storage Project: Lessons Learned and Knowl-
edge Transfer. Energy Procedia , 37: 6226–6236. GHGT-11
Proceedings of the 11th International Conference on Green-
house Gas Control Technologies, 18-22 November 2012,
Kyoto, Japan.
Russakovsky, O.; Deng, J.; Su, H.; Krause, J.; Satheesh, S.;
Ma, S.; Huang, Z.; Karpathy, A.; Khosla, A.; Bernstein, M.;
et al. 2015. Imagenet large scale visual recognition chal-
lenge. International journal of computer vision , 115(3):
211–252.
Selvaraju, R. R.; Cogswell, M.; Das, A.; Vedantam, R.;
Parikh, D.; and Batra, D. 2019. Grad-CAM: Visual Explana-
tions from Deep Networks via Gradient-Based Localization.
International Journal of Computer Vision , 128(2): 336–359.
Simonyan, K.; and Zisserman, A. 2014. Very Deep Convolu-
tional Networks for Large-Scale Image Recognition. arXiv
preprint arXiv:1409.1556 .

Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones,
L.; Gomez, A. N.; Kaiser, L.; and Polosukhin, I. 2017. At-
tention Is All You Need. In Advances in Neural Information
Processing Systems , 6000–6010.
Wang, H.; Wang, Z.; Du, M.; Yang, F.; Zhang, Z.; Ding, S.;
Mardziel, P.; and Hu, X. 2020. Score-CAM: Score-Weighted
Visual Explanations for Convolutional Neural Networks. In
CVPR .
Wason, H.; Oghenekohwo, F.; and Herrmann, F. J. 2017.
Low-cost time-lapse seismic with distributed compressive
sensing–-Part 2: impact on repeatability. Geophysics , 82(3):
P15–P30. (Geophysics).
Wen, G.; Tang, M.; and Benson, S. M. 2021. Towards a
predictor for CO2 plume migration using deep neural net-
works. International Journal of Greenhouse Gas Control ,
105: 103223.
Witte, P. A.; Louboutin, M.; Kukreja, N.; Luporini, F.;
Lange, M.; Gorman, G. J.; and Herrmann, F. J. 2019.
A large-scale framework for symbolic implementations of
seismic inversion algorithms in Julia. Geophysics , 84(3):
F57–F71. (Geophysics).
Yin, Z.; Louboutin, M.; and Herrmann, F. J. 2021. Compres-
sive time-lapse seismic monitoring of carbon storage and se-
questration with the joint recovery model. In SEG Technical
Program Expanded Abstracts , 3434–3438. (IMAGE, Den-
ver).
Yosinski, J.; Clune, J.; Bengio, Y .; and Lipson, H. 2014.
How transferable are features in deep neural networks?.
In Advances in neural information processing systems ,
3320–3328.
Zhou, B.; Khosla, A.; Lapedriza, A.; Oliva, A.; and Torralba,
A. 2015. Learning Deep Features for Discriminative Local-
ization. In IEEE CVPR, , 2921–2929.
Zhou, Z.; Lin, Y .; Zhang, Z.; Wu, Y .; Wang, Z.; Dilmore,
R.; and Guthrie, G. 2019. A data-driven CO2 leakage de-
tection using seismic data and spatial-temporal densely con-
nected convolutional neural networks. International Journal
of Greenhouse Gas Control , 90: 102790.