BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CERN//INDICO//EN
BEGIN:VEVENT
SUMMARY:Accelerating Micro-Macro Models for Two-Mineral Reactive Systems w
ith Machine Learning
DTSTART;VALUE=DATE-TIME:20210531T084000Z
DTEND;VALUE=DATE-TIME:20210531T085500Z
DTSTAMP;VALUE=DATE-TIME:20221001T014141Z
UID:indico-contribution-238-3877@events.interpore.org
DESCRIPTION:Speakers: Stephan Gärttner (Friedrich-Alexander-Universität
Erlangen-Nürnberg)\nIn this talk\, we present an effective micro-macro mo
del for reactive flow and transport in evolving porous media exhibiting tw
o competing mineral phases. As such\, our approach comprises flow and tran
sport equations on the macroscopic scale including effective hydrodynamic
parameters calculated from representative unit cells. Conversely\, the mac
roscopic solutes’ concentrations alter the unit cells' geometrical struc
ture by triggering dissolution or precipitation processes on the distinct
mineral surfaces. Gradually\, such processes result in complex and hardly
predictable geometries. Therefore\, these do not allow for low dimensional
parameter representations. Accordingly\, associate effective parameters c
annot be covered by simple heuristic laws. Hence\, we derive hydrological
parameters directly from the full geometry represented by level-set method
s.\nThe numerical realization of such micro-macro models poses several cha
llenges\, especially in terms of computational complexity due to the incre
ased dimensionality of the problem. Costly computations of effective param
eters directly from the representative geometry often constitute a bottlen
eck in the simulation of highly heterogeneous media. In this talk\, the si
gnificant performance enhancements arising from machine learning technique
s are evaluated. To this end\, convolutional neural networks (CNNs) are tr
ained on geometries derived from geological microCT scans to predict hydro
dynamic parameters such as permeability and diffusivity. The pretrained ne
tworks are subsequently deployed in a micro-macro simulation. We investiga
ted the results in terms of computation time reduction and maintenance of
high predictive quality.\n\nhttps://events.interpore.org/event/25/contribu
tions/3877/
LOCATION:
URL:https://events.interpore.org/event/25/contributions/3877/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Deep learning enhancement of micro-CT images for large-scale flow
simulation
DTSTART;VALUE=DATE-TIME:20210531T082500Z
DTEND;VALUE=DATE-TIME:20210531T084000Z
DTSTAMP;VALUE=DATE-TIME:20221001T014141Z
UID:indico-contribution-238-3752@events.interpore.org
DESCRIPTION:Speakers: Samuel Jackson (CSIRO)\nThere are inherent resolutio
n and field-of-view trade-offs in X-Ray micro-computed tomography imaging\
, which limit the characterization\, analysis and model development of por
ous systems with multi-scale heterogeneities. In this work\, we overcome t
hese tradeoffs by utilising a deep convolution neural network to create en
hanced\, high-resolution data over large spatial scales from low-resolutio
n data.\n \nWe use paired high-resolution (2 micrometres) and low-resoluti
on (6 micrometres) images from two structurally-different Bentheimer rock
samples to train an Enhanced Deep Super Resolution (EDSR) convolutional ne
ural network. The generated high-resolution images are validated against t
he true high-resolution images through textual analysis\, segmentation beh
aviour and pore-network model (PNM) multiphase flow simulations. The final
trained EDSR network is then used to generate high-resolution digital roc
k cores of the whole samples with dimensions of 1.2cm × 1.2cm × 6-7cm. T
he 3D digital rock cores are populated with continuum properties predicted
from subvolume PNMs\, and used to simulate a range of experimental multip
hase flow experiments. We present a consistent workflow to analyse multi-s
cale heterogeneous systems that are otherwise intractable using convention
al methods.\n\nhttps://events.interpore.org/event/25/contributions/3752/
LOCATION:
URL:https://events.interpore.org/event/25/contributions/3752/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Integrating Machine Learning into a Methodology for Early Detectio
n of Wellbore Failure
DTSTART;VALUE=DATE-TIME:20210603T174500Z
DTEND;VALUE=DATE-TIME:20210603T180000Z
DTSTAMP;VALUE=DATE-TIME:20221001T014141Z
UID:indico-contribution-238-3908@events.interpore.org
DESCRIPTION:Speakers: Edward Matteo (Sandia National Laboratories)\nThere
are literally a few million boreholes in the continental US (both onshore
and offshore) that include abandoned wells\, production wells\, and wells
for underground hydrocarbon storage. Some are vulnerable to potentially ca
tastrophic loss of seal integrity\, largely owing to progressive damage of
the annular cement sheath. The Deepwater Horizon oil spill and the Aliso
Canyon natural gas leak have elevated wellbore integrity to national atten
tion\; new regulations have begun to address the effects of\, but not caus
es of\, well failure. It is conjectured that damage within the annular cem
ent between host rock and well casing\, engineered as the main seal betwee
n biosphere and subsurface\, is one of the main leakage pathways. One appr
oach to helping solve this problem is to utilize existing operational data
sets from monitored wellbores as a testbed for developing methodologies th
at can screen for early detection of damage and/or failure.\n\nThere are
few publicly available datasets of wellbore deformation\, damage\, and lea
kage due to geological forces. One existing group of datasets includes we
llbores for several underground hydrocarbon storage facilities consisting
of storage caverns built in salt domes. In these datasets\, histories of
wellbore casing damage have been determined from measurements taken by mul
ti-arm calipers over many years. The operators of these facilities have o
bserved some patterns to these deformation histories based on knowledge of
the geomechanics of these salt domes\, but a full explanation of these ev
ents is incomplete. This group of datasets has been selected for use in a
machine-learning study to evaluate\, interpret\, and predict patterns of
casing damage.\n\nIn our research\, we explore using data science and mach
ine learning (ML) methods to predict when a well might be approaching a s
tate of failed seal integrity over time. We use Subject Matter Expert (SME
) information as well as statistical techniques including correlation and
regression analysis to define the features for our ML models. Our time ser
ies prediction considers both next time-step modeling as well as longer te
rm time-series forecasting by utilizing random forests (RFs) and deep neur
al networks (DNNs) as well as recurrent neural networks (RNNs) for the pre
dictions. The RF models allow us to perform feature importance characteriz
ation\, while DNNs\, specifically convolutional DNNs\, facilitate utilizat
ion of spatial information including depth and volumetric data. We will ut
ilize these models to characterize and automate the identification of fact
ors that put wellbores at risk\, so as to be used as an early detection sy
stem for failure screening that outperforms existing analysis tools. \n\nS
NL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525.
SAND2021-1732 A\n\nhttps://events.interpore.org/event/25/contributions/390
8/
LOCATION:
URL:https://events.interpore.org/event/25/contributions/3908/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Physics Impact on Deep Neural Networks for Multiphase Flow in Poro
us Media
DTSTART;VALUE=DATE-TIME:20210603T173000Z
DTEND;VALUE=DATE-TIME:20210603T174500Z
DTSTAMP;VALUE=DATE-TIME:20221001T014141Z
UID:indico-contribution-238-3897@events.interpore.org
DESCRIPTION:Speakers: Bicheng Yan (Los Alamos National Laboratory)\, Rajes
h Pawar (Los Alamos National Laboratort)\nMany recent studies have demonst
rated the superior predictive interpretability and physics consistency by
anchoring deep feedforward neural networks (DNN) with physics laws. As thi
s type of network is fully connected\, it potentially suffers low training
efficiency when predicting complex problems such as multiphase flow in po
rous media.\n\nIn this study\, we propose a learning framework to predict
the state variables of pressure\, saturation and well flow rate in fluid f
low in porous media. Since fluid flow in porous media is constrained by p
hysics\, we allow the learning framework to flexibly intake either pure la
beled data\, or a combination of labeled data and physics data in terms of
governing equation or physics-based operators. Given that the loss evalu
ation of non-labeled data during backpropagation is expensive due to autom
atic differentiation\, a modified batch-mode training procedure with trans
fer learning is proposed to ensure that it balances the training efficienc
y and the contribution of all the samples on the training loss. As well f
low rates are time series data only associated with well locations\, it is
effectively predicted along with pressure and saturation by a simple spar
se operator in the same framework.\n\nNumerical experiments of multiphase
flow related to geologic carbon storage are used to gauge the performance
of learning framework. Our results show that DNN integrating labeled data
with physics or the physics-based operators doesn’t bring too much addit
ional CPU cost due to automatic differentiation\, but they effectively imp
roves the fidelity of pressure and saturation prediction compared to DNN w
ith labeled data only. Moreover\, the prediction of the well flow rate is
quite accurate with average error lower than 2.0%. Therefore\, the learn
ing framework helps us to investigate the impact of physics on DNN predict
ions\, and provides us the guidance to effectively train a DNN with a good
combination of physics and labeled data.\n\nhttps://events.interpore.org/
event/25/contributions/3897/
LOCATION:
URL:https://events.interpore.org/event/25/contributions/3897/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Machine learning prediction of Lennard-Jones fluid self-diffusion
in pores
DTSTART;VALUE=DATE-TIME:20210603T171500Z
DTEND;VALUE=DATE-TIME:20210603T173000Z
DTSTAMP;VALUE=DATE-TIME:20221001T014141Z
UID:indico-contribution-238-3894@events.interpore.org
DESCRIPTION:Speakers: Calen Leverant ()\nPorous materials are widely used
in industrial applications (e.g.\, catalysis and separations) and diffusio
n of liquids within these materials can often control performance. Self-di
ffusion coefficients are typically obtained from molecular dynamics (MD) s
imulations in which the forces and trajectories of particles are calculate
d via Newtonian physics for millions of time steps. While MD provides accu
rate diffusion coefficients and can be tailored to a variety of circumstan
ces (e.g.\, diffusion in pores)\, it is computationally expensive and requ
ires large systematic studies in order provide insight as to which molecul
ar properties affect diffusion. To provide a quicker\, more computationall
y efficient alternative\, we trained a variety of machine learning algorit
hms from simple linear models to complex convolutional neural networks to
predict the diffusion of Lennard-Jones particles in a variety of ideal por
e shapes. During the feature selection process\, extra consideration was g
iven to select features that are easy to obtain and understand by non-expe
rts. Not only can these machine learning algorithms accurately predict dif
fusion coefficients at a fraction of the computational cost of MD\, but th
ey provide the opportunity to study the important features that contribute
to the diffusion coefficient values. Insights obtained from studying the
feature importance of these models can provide further understanding to th
e diffusion in porous media and enhance the materials design process for f
uture porous materials used in industrial processes.\n\nThis work is suppo
rted entirely by the Laboratory Directed Research and Development Program
at Sandia National Laboratories. Sandia National Laboratories is a multimi
ssion laboratory managed and operated by National Technology and Engineeri
ng Solutions of Sandia\, LLC.\, a wholly owned subsidiary of Honeywell Int
ernational\, Inc.\, for the U.S. Department of Energy's National Nuclear S
ecurity Administration under contract DE-NA-0003525. The views expressed i
n this article do not necessarily represent the views of the U.S. Departme
nt of Energy or the United States Government.\nSAND2021-1589 A\n\nhttps://
events.interpore.org/event/25/contributions/3894/
LOCATION:
URL:https://events.interpore.org/event/25/contributions/3894/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Geostatistical Inversion in Geologic CO2 Sequestration Using a Var
iational Autoencoder
DTSTART;VALUE=DATE-TIME:20210603T170000Z
DTEND;VALUE=DATE-TIME:20210603T171500Z
DTSTAMP;VALUE=DATE-TIME:20221001T014141Z
UID:indico-contribution-238-3892@events.interpore.org
DESCRIPTION:Speakers: Bailian Chen (Los Alamos National Laboratory)\nGeost
atistical inversion problems in geologic CO2 sequestration (GCS) often inv
olve matching observational data using a physical model that takes a large
number of parameters. It is known that solving an inversion problem in a
high-dimensional space with complex structure is usually a very time consu
ming process. In this work\, a dimensionality reduction technique\, variat
ional autoencoder (VAE)\, was proposed to efficiently invert storage reser
voir parameter fields (e.g.\, permeability) with the aim of improving pred
ictions of important metrics such as pressure and CO2 saturation maps. A g
radient-based optimization algorithm\, L-BFGS\, is utilized to minimize th
e observational and predictive data mismatch function. A VAE is trained to
map to a low-dimensional set of latent variables with a simple structure
to the high-dimensional parameter space (i.e.\, original space) that has a
complex structure. The required optimization process to fit model to obse
rvational data will then be performed on a low dimensional latent space\,
making the gradient-based optimization (i.e.\, L-BFGS) more computationall
y efficient. The feasibility and efficiency of the proposed approach for G
CS inverse analysis were demonstrated with a 3D synthetic case. A prelimin
ary result is shown in the figure. (Figure caption: Predictions of CO2 sat
uration plume at the end of post-injection period based on the updated mod
els under different monitoring durations (1 yr\, 3 yrs\, etc.))\n\nhttps:/
/events.interpore.org/event/25/contributions/3892/
LOCATION:
URL:https://events.interpore.org/event/25/contributions/3892/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Non-intrusive reduced order modeling of natural convection in poro
us media
DTSTART;VALUE=DATE-TIME:20210603T164500Z
DTEND;VALUE=DATE-TIME:20210603T170000Z
DTSTAMP;VALUE=DATE-TIME:20221001T014141Z
UID:indico-contribution-238-3891@events.interpore.org
DESCRIPTION:Speakers: Teeratorn Kadeethum (Cornell University)\nA simulati
on tool capable of speeding up the calculation for natural convection in p
orous media is of sizeable practical interest for engineers\, in particula
r\, to effectively perform sensitivity analyses\, uncertainty quantificati
on\, and optimization of $\\mathrm{CO_2}$ sequestration and geothermal har
vesting. We present a non-intrusive reduced order model (ROM) using the ne
sted proper orthogonal decomposition (POD) and artificial neural networks
(ANN). In this study\, the nested POD refers to a compression strategy in
which time and uncertain parameter domains are compressed consecutively (i
n contrast to the classical POD method in which all domains are compressed
simultaneously). We utilize the two-field mixed finite element method and
interior penalty discontinuous Galerkin approximation for spatial discret
ization and the 4th-order backward differentiation formula for time-steppi
ng as our full order model (FOM). This combination is selected to avoid sp
urious oscillations resulting from the lack of local mass conservation and
accurately capture the gravity-driven flow in advection-dominated problem
s. The proposed framework is divided into an offline phase for the ROM con
struction\, which we will present through five consecutive steps and (sing
le-step) online stage for the ROM evaluation. The \\textbf{offline phase}
includes the following steps: (1) initialize a training set (uncertain par
ameters)\, which could correspond to material properties\, boundary condit
ions\, or geometric characteristics\, (2) query the FOM for each value in
the training set\, (3) compress the FOM results using the nested POD\, (4)
obtain the optimal representation of the FOM results employing an $L^2$ p
rojection\, and (5) train the ANN to map the set of uncertain parameters (
input) to the collection of coefficients calculated from an $L^2$ projecti
on over the reduced basis (output). During the \\textbf{online phase}\, fo
r given values of uncertain parameters\, we aim to recover the online appr
oximation of our primary variables by querying the ANN evaluation of the c
ollection of coefficients and reconstructing the resulting finite element
representation through the reduced basis functions. We conclude the presen
tation using a series of validations through the method of manufactured so
lutions used in and benchmark problems of the Horton–Rogers–Lapwood an
d Elder problems.\n\nSNL is managed and operated by NTESS under DOE NNSA c
ontract DE-NA0003525\n\nhttps://events.interpore.org/event/25/contribution
s/3891/
LOCATION:
URL:https://events.interpore.org/event/25/contributions/3891/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Heterogeneity Evaluation of Microstructures in a Sandstone Reservo
ir Using Micro-CT Imagery
DTSTART;VALUE=DATE-TIME:20210603T163000Z
DTEND;VALUE=DATE-TIME:20210603T164500Z
DTSTAMP;VALUE=DATE-TIME:20221001T014141Z
UID:indico-contribution-238-3889@events.interpore.org
DESCRIPTION:Speakers: Lingyun Kong (National Energy Technology Laboratory\
; Energy & Environmental Research Center)\nHeterogeneity of microstructure
s in clastic rocks is relevant to a wealth of subsurface properties (e.g.\
, porosity\, permeability\, fracture orientations) and\, yet\, is challeng
ing to effectively characterize because of the stochastic distribution of
grain deposition\, diagenesis\, and texture deformation. At the pore to co
re scale\, heterogeneity lies in the spatial variation of pore throat netw
orks and complicates the fluid flow mechanisms associated with those netwo
rks. Therefore\, it is essential to precisely comprehend the heterogeneity
at the fine scale\, which further facilitates understanding at a broader
scale (well to basin). In this study\, micro-CT (computed tomography) imag
es of samples from a conventional sandstone reservoir were collected. The
heterogeneity was evaluated in terms of the microstructure variation (pore
size distribution\, pore shape distribution\, and porosity variation) in
three dimensions spatially and the associated petrophysical property chang
es derived from images. Firstly\, the representative elementary volume (RE
V) was determined by extracting the subvolumes at increasing sizes inside
of CT image data sets. REV was set to 100 voxels (20.7 µm/voxel) to captu
re the representative area to assess heterogeneity. An ImageJ Macro algori
thm was coded to automatically resample the subvolume from a 100-voxel siz
e to a 600-voxel size for a cube-shaped region of interest. A machine lear
ning-assisted thresholding method was developed to segment the grain matri
x and pore structures. Further\, microstructure variation was calculated b
y processing subvolumes\, with the permeability derived by adopting the Ko
zeny–Carman equation. The statistics from the above parameters demonstra
te both the size effect and spatial effect of the region of interest exist
ing in the sandstone reservoir. Moreover\, the fractal dimension\, a mathe
matical parameter indicating the self-similarity and complexity of a subje
ct\, was utilized to quantify the heterogeneity of microstructures at incr
easing subvolume sizes\, where the same trend was revealed as fractal dime
nsion increasing from 2.6 to 2.95. The database obtained via quantitative
evaluation of the microstructure variation within the sandstone samples pr
ovides for machine learning-informed image analysis and is an essential st
ep toward heterogeneity characterization across various scales.\n\nhttps:/
/events.interpore.org/event/25/contributions/3889/
LOCATION:
URL:https://events.interpore.org/event/25/contributions/3889/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Semantic segmentation of microCT and FIB-SEM rock images using dee
p learning methods
DTSTART;VALUE=DATE-TIME:20210603T161500Z
DTEND;VALUE=DATE-TIME:20210603T163000Z
DTSTAMP;VALUE=DATE-TIME:20221001T014141Z
UID:indico-contribution-238-3888@events.interpore.org
DESCRIPTION:Speakers: Jack Ringer (Sandia National Laboratories)\nRecent a
dvances in multiscale imaging techniques for the analysis of complex pore
structures and compositions have revolutionized our ability to characteriz
e various porous media systems. Segmentation of images obtained from diffe
rent image techniques such as X-ray computed microtomography (μCT) and sc
anning electron microscopy (SEM) is the first step to quantitatively descr
ibe various features of geomaterials. However\, conventional methods such
as thresholding\, watershed segmentation\, and other predefined algorithms
are subject to user bias\, often require specific segmentation processes
for each image set\, and can fail to adapt to certain datasets. In this wo
rk we evaluate the capability of convolutional neural networks (CNNs)-base
d algorithms to segment both μCT and focused ion beam-SEM (FIB-SEM) image
s with varying degree of challenges for image segmentation. The performanc
e of three different 2D CNN architectures (VGG16\, ResNet\, and U-Net) as
well as a few 3D CNN architectures (U-Net and MultiResU-Net) is assessed o
n four independent datasets including sandstone\, carbonate chalks\, and s
hale. Each of these datasets is composed of three-dimensional image stacks
and corresponding ground truth labels that were constructed with various
image processing algorithms. Our results indicate that deep learning archi
tectures can successfully be applied to the task of semantic segmentation
for μCT and FIB-SEM images and perform better than manual segmentation to
recover natural morphology of original images. In addition\, our results
indicate that transfer learning can allow for models to converge more quic
kly during training and that generic image features (learned from a large
dataset such as ImageNet) can be applied to improve model performance in s
ome cases. Performance comparison among different CNN architectures highli
ghts the linkage of classification outcomes to underlying features of each
CNN architecture and hyperparameters.\n\nhttps://events.interpore.org/eve
nt/25/contributions/3888/
LOCATION:
URL:https://events.interpore.org/event/25/contributions/3888/
END:VEVENT
BEGIN:VEVENT
SUMMARY:CCSNet: a deep learning modeling suite for CO2 storage
DTSTART;VALUE=DATE-TIME:20210603T160000Z
DTEND;VALUE=DATE-TIME:20210603T161500Z
DTSTAMP;VALUE=DATE-TIME:20221001T014141Z
UID:indico-contribution-238-3887@events.interpore.org
DESCRIPTION:Speakers: Gege Wen (Stanford University)\nNumerical simulation
is an essential tool for understanding subsurface flow in porous media pr
oblems\, yet it often suffers from computational challenges due to these p
roblems' highly non-linear governing equations\, their multi-physics natu
re\, and the need for high spatial resolutions to capture multi-scale hete
rogeneity. The inherent parameter uncertainties in subsurface porous medi
a necessitate probabilistic assessments and history matching tasks\, which
often require prohibitively large numbers of simulation runs. To aid engi
neering decisions\, surrogate modeling methods are proposed by developing
lower-fidelity but computationally efficient models that can provide reaso
nably accurate results for specific tasks. Deep learning has recently show
n a growing potential for subsurface flow and transport problems. Specific
ally\, supervised learning approaches use data generated by numerical simu
lators to train deep learning models and have shown encouraging results fo
r uncertainty quantification or history matching tasks. Here\, we introduc
e the CCSNet\, a general-purpose deep-learning tool that can act as an alt
ernative to conventional numerical simulators for a class of subsurface fl
ow in porous media problems\, namely\, carbon capture and storage (CCS) p
roblems. Unlike most proxy or surrogate models\, which are developed on a
task basis\, we demonstrate that CCSNet provides solutions to an entire cl
ass of CCS problems where CO$_2$ is injected into saline aquifers in 2d-ra
dial systems. \n\n CCSNet is trained with a data set that represents almos
t all potential variables in the problem domain\, including an extensive r
ange of reservoir conditions\, fluid properties\, geological attributes\,
rock properties\, multiphase flow properties\, and injection designs. The
CCSNet consists of a sequence of deep learning models to collaboratively p
roduce salient outputs that a conventional numerical simulator can provide
\, including gas saturation distributions\, pressure buildup\, CO$_2$ diss
olution\, dry-out\, fluid densities in gas and liquid phases\, and mass ba
lance. The dynamic change of these outputs is captured by a tailored tempo
ral-3d convolutional neural network (CNN) architecture. The full set of ou
tputs also allow us to evaluate how well the results satisfy the governing
conservation equations without explicitly representing them in the loss f
unction. The results are highly resolved\, nearly as accurate as numerical
simulation outputs\, and have excellent computational efficiencies that a
re 10$^3$ to 10$^5$ times faster than conventional numerical simulators. F
or 2d-radial CO$_2$ injection problems\, our results show that CCSNet can
sufficiently act as an alternative to computationally intensive numerical
simulators.\n \nTo illustrate the high computational efficiency of CCSNet\
, we applied it to the development of rigorous estimation techniques for s
weep efficiency and solubility trapping based on a stochastic sampling of
the problem domains. Our results show that sweep efficiencies in homogeneo
us reservoirs can be predicted accurately using only three variables: the
Bond number\, injection rate\, and irreducible water saturation. Interest
ingly\, injection depth and injection interval had little influence on swe
ep efficiency. Predicting solubility trapping is more complex\, requiring
information about the formation permeability\, irreducible water saturatio
n\, pressure\, temperature\, Bond number\, and capillary pressure curves.
Simple equations are now available to estimate both of these parameters as
part of site screening process.\n\nhttps://events.interpore.org/event/25/
contributions/3887/
LOCATION:
URL:https://events.interpore.org/event/25/contributions/3887/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Robust porous media flow control using Deep Reinforcement Learning
DTSTART;VALUE=DATE-TIME:20210604T091000Z
DTEND;VALUE=DATE-TIME:20210604T092500Z
DTSTAMP;VALUE=DATE-TIME:20221001T014141Z
UID:indico-contribution-238-3902@events.interpore.org
DESCRIPTION:Speakers: Atish Dixit (PhD student)\nWith the recent progress
in reinforcement learning (RL) research\, we investigate whether it would
be suitable to use RL in solving optimal well control problem with uncerta
in reservoir models. In principle\, RL algorithms are capable of learning
optimal action policies — a map from states to actions — to maximize a
numerical reward signal. In the RL formulation of porous media flow contr
ol problems\, we represent the state with snapshots of subsurface flow sim
ulation\; the action with valve openings controlling flow through sources/
sinks (i.e.\, injection/production) wells while the numerical reward refer
s to the total sweep efficiency. Optimal control policies are learned by n
umerous episodes of simulation trials (referred to as agent-environment in
teractions in the RL literature). \n\nThe major challenge in learning an
optimal flow control policy for well control is that the reservoir simulat
ion often comprises of uncertain parameters (e.g.\, permeability fields).
To the best of our knowledge\, so far\, such policies are learned by simpl
y incurring samples of parameter uncertainty distribution in each episode
of agent-environment interactions. Such a policy learning process is often
very unstable. furthermore\, it requires a very high number of episodes\,
such that the variety of parameter uncertainty domain is thoroughly explo
red. This is computationally quite intensive for porous media flow problem
s for subsurface reservoir. Therefore\, we investigate if we can learn the
robust optimal policy with just few samples of uncertainty distribution i
n order to cope with these limitations. \n\nWe present two test cases rep
resenting two distinct permeability uncertainty distributions as a proof o
f concept for our study. Policy based model-free RL algorithms like PPO (p
roximal policy optimization) and A2C (advantage actor-critic) are employed
to solve the robust optimal control problem for both test cases. The resu
lts are benchmarked with the optimization results obtained using different
ial evolution algorithm.\n\nhttps://events.interpore.org/event/25/contribu
tions/3902/
LOCATION:
URL:https://events.interpore.org/event/25/contributions/3902/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Integrating process-based reactive transport modeling and machine
learning for surrogate model development: an application to electrokinetic
remediation of contaminated groundwater
DTSTART;VALUE=DATE-TIME:20210604T085500Z
DTEND;VALUE=DATE-TIME:20210604T091000Z
DTSTAMP;VALUE=DATE-TIME:20221001T014141Z
UID:indico-contribution-238-3884@events.interpore.org
DESCRIPTION:Speakers: Riccardo Sprocati (Technical University of Denmark)\
nMultiphysics reactive transport models are nowadays widely employed in su
bsurface porous media and are able to account for several fully coupled ph
ysical and biogeochemical processes. However\, the increasing level of det
ail of such models comes at the price of an increased complexity which oft
en leads to long runtimes. As a result\, explorative and probabilistic ana
lysis that require numerous model evaluations are hampered by the excessiv
e simulation time required. \nTo overcome this limitation\, we show that i
t is possible to develop machine learning surrogate models which are able
to predict the evolution of complex subsurface remediation systems using a
s training data a limited number of process-based reactive transport simul
ations [1]. \nWe focus on an application of electrokinetic enhanced biorem
ediation (EK-Bio)\, which aims at the removal of chlorinated solvents in l
ow-permeability porous media. The modeling of this in situ remediation tec
hnology is challenging as requires the solution of coupled physical\, elec
trostatic\, chemical and biological processes. We developed a process-base
d\, multicomponent reactive transport model of EK-Bio using the code NP-Ph
reeqc-EK [2]\, which can simulate the key mechanisms of electrokinetic flo
w and transport in multidimensional domains. The model accounts for electr
omigration and electroosmosis\, Coulombic interactions\, interphase mass-t
ransfer\, equilibrium and kinetically controlled reactions\, including con
taminant degradation and biomass dynamics [3]. \nThe machine learning surr
ogate model was developed with a response surface approach using an approx
imation function based on an artificial neural network with a stack of mul
ti-layer perceptrons. The surrogate model uses as inputs for the training
the outputs of multiple runs of the process-based model\, defined accordin
g to a design of experiments procedure. Subsequently\, the surrogate model
was trained with randomized cross-validation of hyperparameters. Comparin
g the surrogate model performances on a test set\, the neural network demo
nstrated excellent prediction and generalization performances on all outpu
t variables. The developed surrogate allowed us to perform computationally
efficient model exploration\, global sensitivity analysis\, and probabili
stic uncertainty quantification.\n\nhttps://events.interpore.org/event/25/
contributions/3884/
LOCATION:
URL:https://events.interpore.org/event/25/contributions/3884/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Prediction of Flow and Reactive Transport using Physics-Informed N
eural Networks
DTSTART;VALUE=DATE-TIME:20210604T084000Z
DTEND;VALUE=DATE-TIME:20210604T085500Z
DTSTAMP;VALUE=DATE-TIME:20221001T014141Z
UID:indico-contribution-238-3880@events.interpore.org
DESCRIPTION:Speakers: Vincent Liu (Sandia National Laboratories)\nFlow and
reactive transport in fractured and porous media are fundamental to under
standing coupled multiphysics processes critical to various geoscience and
environmental applications such as geologic carbon storage\, subsurface e
nergy recovery\, and environmental biogeochemical processes. Although flui
d dynamics simulations provide fundamental solutions to flow and reactive
transport processes\, these computational simulations are often computatio
nally intensive and would not be scalable to high dimensional applications
. Deep learning can offer computationally efficient solutions to such prob
lems while reliable neural network models require a large number of traini
ng samples. Physics-informed neural network approaches can provide machine
learning solutions to physical systems respecting the laws of physics giv
en by general nonlinear differential equations with a small number of trai
ning data\, but training such networks require domain-specific expertise f
or better convergence. In this work\, we apply hybrid physics informed neu
ral networks and data augmentation to predict fluid flow in a constrained
geometry. We test our models to evaluate various flow and reactive transpo
rt problems in 2D domains using the advection-diffusion(or dispersion)-rea
ction and Navier Stokes/Darcy equations. Additionally\, we test flow and t
ransport problems in the presence of an obstructing cylinder to analyze fl
uid velocity and concentration distribution from advection-diffusion-react
ion. Comparison of results between the physics-informed deep learning appr
oaches and computational simulations will be presented to highlight the ac
curacy of physics-informed neural networks and advance computational effic
iency.\nSNL is managed and operated by NTESS under DOE NNSA contract DE-NA
0003525.\n\nhttps://events.interpore.org/event/25/contributions/3880/
LOCATION:
URL:https://events.interpore.org/event/25/contributions/3880/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Inter-well Connectivity Analysis and Productivity Prediction Based
on Intelligent Connectivity Model
DTSTART;VALUE=DATE-TIME:20210531T094000Z
DTEND;VALUE=DATE-TIME:20210531T095500Z
DTSTAMP;VALUE=DATE-TIME:20221001T014141Z
UID:indico-contribution-238-3899@events.interpore.org
DESCRIPTION:Speakers: Yunqi Jiang (China University of Petroleum (East Chi
na))\, Kai Zhang (China University of Petroleum (East China))\nArtificial
neural networks (ANNs) are well known for its strong learning ability and
have been widely used in the petroleum industry\, such as history matching
\, production optimization and productivity forecast. However\, ANNs are a
lso a typical kind of “black box” models for their weakness in the mod
el interpretability\, causing their results less reliable than those from
other physics based models. This paper proposes an integrated model named
intelligent connectivity model (ICM)\, which incorporates ANNs with the ma
terial balance equation within a machine learning (ML) framework. ICM is a
modular model\, and each module keeps correspondence with each item in th
e material balance equation\, improving the model transparency and general
ization capability significantly. The results of simulation experiments sh
ow that ICM enables to generate comparable prediction results and provide
more reasonable characterizations on inter-well connectivity than the clas
sical physical model\, and meanwhile ICM is more computationally efficient
.\n\nhttps://events.interpore.org/event/25/contributions/3899/
LOCATION:
URL:https://events.interpore.org/event/25/contributions/3899/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Particle transport and filtration in 2D and 3D porous media: coupl
ing CFD and Deep Learning
DTSTART;VALUE=DATE-TIME:20210531T092500Z
DTEND;VALUE=DATE-TIME:20210531T094000Z
DTSTAMP;VALUE=DATE-TIME:20221001T014141Z
UID:indico-contribution-238-3896@events.interpore.org
DESCRIPTION:Speakers: Agnese Marcato (Politecnico di Torino)\nThe study of
particle transport in porous media is a research field of great interest
as it is involved in a wide variety of applications [1]. The random natur
e of porous media systems makes it difficult to analytically correlate the
impact and the synergy of the their geometrical parameters. Since these f
eatures make these systems a suitable candidate for machine learning (ML)
approaches\, in our work we employed neural networks for the realization o
f data-driven models. These techniques are able to grasp non-linear corre
lations between data and to account for a large number of input parameters
. Moreover in the case of convolutional neural networks the entire system
geometry can be used as input for the model\, in this way it is possible
to avoid the selection of the geometrical features [2-3]. \nIn this work
we coupled computational fluid dynamics (CFD) simulations with machine lea
rning models. The results of a CFD simulations campaign are employed as a
training set for the neural networks in order to obtain a computationally
inexpensive data-driven surrogate model which is able to replace the CFD s
imulation\, while keeping a good accuracy. The aim of the CFD investigatio
n is the flow\, transport\, and filtration at the pore-scale\, in this fra
mework the first step is the creation of the geometries. We designed bi-di
mensional [4] and three-dimensional [5] periodic packings of spheres via t
he open-source framework YADE DEM. \nFor each kind of geometry\, hundreds
of simulations are solved\, each differing randomly in their geometrical p
arameters and input operating conditions. The CFD simulations are performe
d on the open-source code OpenFOAM. At first\, the fluid flow is evaluated
in the limit of small Reynolds numbers (<0.1)\, thus obtaining the medium
permeability. Then the transport of dilute colloid particles is studied b
y solving the advection-diffusion equation\, and the filtration rate is ca
lculated [6]. \nTwo kinds of models have been built: both for the predicti
on of the permeability of the porous media\, and the filtration rate of th
e colloid through the grains. The first one is a simple fully-connected ne
ural network whose input features are the geometric parameters and operat
ing conditions. The second one is a convolutional neural network whose inp
ut is a porous medium geometry\, in the form of a binary matrix. After th
e neural network training process\, the end result is a surrogate black-bo
x model capable of predicting the output values when given a new set of in
put features\, or a new geometry\; notably\, the accuracy of this data-dri
ven model is on-par or better than other analytical or empirical correlati
ons.\nThis simple data-driven models can then be reliably used in place of
expensive CFD simulations (or in general\, all “first principles” met
hods)\, as one single call of the neural network has a computational cost
which is orders of magnitude lower than the full CFD simulation: in our te
st problems\, under a second versus several hours – with a total neural
network training time of around four minutes\, for the fully connected one
\, and of several hours\, for the convolutional one.\n\nhttps://events.int
erpore.org/event/25/contributions/3896/
LOCATION:
URL:https://events.interpore.org/event/25/contributions/3896/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Deep-learning-based surrogate model for brine extraction well plac
ement for geological carbon storage
DTSTART;VALUE=DATE-TIME:20210531T091000Z
DTEND;VALUE=DATE-TIME:20210531T092500Z
DTSTAMP;VALUE=DATE-TIME:20221001T014141Z
UID:indico-contribution-238-3886@events.interpore.org
DESCRIPTION:Speakers: Hyunjee Yoon ()\nIn a geological carbon storage proj
ect\, management of reservoir pressure buildup is essential for long-term
safe carbon storage. A reservoir pressure buildup caused by CO2 injection
may lead to serious safety issues such as induced seismicity\, caprock dam
age\, and leakage of brine and CO2. Brine extraction is a practical soluti
on to mitigate the reservoir pressure buildup. In heterogeneous reservoirs
\, the performance of brine extraction is significantly affected by where
to place a brine extraction well because the mitigation of pressure buildu
p and the arrival time of injected CO2 to the brine extraction well are de
termined by the hydraulic connectivity map. The optimization of a brine ex
traction well location is computationally expensive because many reservoir
simulation runs are required to seek optimal locations in potential well
locations. We propose an efficient surrogate model that computes the optim
ality of a brine extraction well quickly using the fast marching method an
d a convolutional neural network. The arrival time map of a pressure pulse
that the fast marching method provides rapidly can be used as a good repr
esentation of the hydraulic connectivity map for a brine extraction well l
ocation. The performance of our surrogate model is demonstrated in a CO2 i
njection site in the Pohang basin. The computational cost of optimization
of a brine extraction well is significantly saved using our accurate surro
gate model compared to a normal optimization process.\n\nhttps://events.in
terpore.org/event/25/contributions/3886/
LOCATION:
URL:https://events.interpore.org/event/25/contributions/3886/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Flux Regression Performances of Deep Learning in Discrete Fracture
Networks
DTSTART;VALUE=DATE-TIME:20210531T085500Z
DTEND;VALUE=DATE-TIME:20210531T091000Z
DTSTAMP;VALUE=DATE-TIME:20221001T014141Z
UID:indico-contribution-238-3885@events.interpore.org
DESCRIPTION:Speakers: Francesco Della Santa (Politecnico di Torino)\nThe n
eed of flow and transport characterization in underground fractured media
is critical in many engineering applications\, like fossil fuel extraction
and water resources analysis. However\, there is a lack of full knowledge
(geometrical and hydrogeological) of these fracture systems and\, therefo
re\, statistical representations of the fractured media are given. In this
context\, we perform flow simulations in underground fractures with Discr
ete Fracture Network (DFN) models.\n\nThe stochastic representation of the
fracture systems requires thousands of DFN generations and simulations to
characterize the flow in a real fractured medium. For this reason\, it is
desirable to consider the application of Deep Learning models and use the
m as alternative model reduction methods to speed up the flow characteriza
tion process.\n\nIn this work we show the application of a set of Deep Lea
rning models for flux regression in Discrete Fracture Networks\, analyzing
the regression quality and revealing suitable enhancements of the already
existing encouraging results [1].\n\nhttps://events.interpore.org/event/2
5/contributions/3885/
LOCATION:
URL:https://events.interpore.org/event/25/contributions/3885/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Applying Machine Learning Methods to Speed Up Two-Phase Relative P
ermeability Upscaling
DTSTART;VALUE=DATE-TIME:20210531T081000Z
DTEND;VALUE=DATE-TIME:20210531T082500Z
DTSTAMP;VALUE=DATE-TIME:20221001T014141Z
UID:indico-contribution-238-3870@events.interpore.org
DESCRIPTION:Speakers: Yanji Wang (School of Petroleum Engineering\, China
University of Petroleum (East China))\nTraditional flow-based two-phase up
scaling entails the computation of upscaled relative permeability function
s for each coarse block or interface. It can be very time-consuming especi
ally for large models with a large quantity of coarse gird blocks or for c
ases that requires simulation runs over multiple geological realizations (
as commonly used in uncertainty quantification or optimization). In this w
ork\, we introduce machine learning (ML) methods into the two-phase upscal
ing procedure to significantly speed up the upscaling computations. In the
new procedure\, the flow-based relative permeability upscaling is perform
ed only for representative coarse blocks/interfaces\, while the upscaled f
unctions for the majority of the coarse blocks are provided by the ML meth
ods.\n\nThe new upscaling procedure entails a few steps. First\, a ML meth
od is applied to select the representative coarse blocks/interfaces based
on the static permeability distribution associated with the target regions
. Flow-based two-phase upscaling is then performed for the selected blocks
/interfaces to build a database. A different ML model can then be construc
ted to reveal the relationship between the upscaled relative-permeability
functions and the corresponding static permeability distribution. This ML
model is finally used to give the upscaled relative permeability functions
for the rest of the coarse blocks/interfaces. In this work\, both the loc
al and extended local two-phase upscaling approaches with generic pressure
and saturation boundary conditions and effective flux boundary conditions
are incorporated with the ML-based upscaling procedure. \n\nWe test the n
ew upscaling procedure for generic (left to right) flow problems using 2D
models for oil-water two-phase systems. Both Gaussian and channelized perm
eability fields are considered. Extensive numerical results have shown tha
t the coarse-scale simulation results using the ML-based upscaling procedu
re are of similar accuracy compared to those using full flow-based upscali
ng. The relative errors of the total production rate and water cut are wit
hin 5%. Besides\, at least one order of magnitude speedups achieved\, whic
h are quite significant. Higher speedup is observed for models with larger
dimensions. \n\nThe ongoing work includes extending the procedure into 3D
models\, and testing it for actual field problems with more complex model
geometry.\n\nhttps://events.interpore.org/event/25/contributions/3870/
LOCATION:
URL:https://events.interpore.org/event/25/contributions/3870/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Research on pore-scale hydrate permeability prediction based on ma
chine learning
DTSTART;VALUE=DATE-TIME:20210531T075500Z
DTEND;VALUE=DATE-TIME:20210531T081000Z
DTSTAMP;VALUE=DATE-TIME:20221001T014141Z
UID:indico-contribution-238-3869@events.interpore.org
DESCRIPTION:Speakers: Ziwei Bu (China University of Petroleum (East China)
)\nNatural gas hydrate has huge reserves and is one of the most potential
carbon energy resources. In the process of natural gas hydrate production\
, the phase state changes in the formation. Until now\, the gas-liquid two
-phase flow mechanism is not well understood for gas hydrate formation. Th
e permeability of gas and water determines the flow capacity of fluids in
hydrate formation and directly affects the efficiency of natural gas produ
ction. Since gas-water two-phase flow can cause changes of hydrate saturat
ion and pore structure\, the studies on the relative permeability is not i
nadequate. This study uses a combination of numerical simulation and machi
ne learning to learn the relationship among pore statistical characteristi
c\, the pore habits of hydrates\, hydrates saturation and permeability. Th
e goal is to reveal the seepage characteristics of hydrates at pore scale.
\nUsing COMSOL Multiphysics software\, pore-scale hydrate models are estab
lished\, the N-S equation is used to describe the gas-water flow. Gas-wate
r two-phase flow are simulated. A large number of data samples are generat
ed and the pore-scale permeability prediction database is conducted. Based
on the data samples generated by COMSOL Multiphysics\, machine learning a
lgorithms are used for permeability analysis. The hydrate permeability cal
culation model considering different hydrate pore habits (pore filling\, p
article coating\, et al.) \, pore statistical characteristic\, and saturat
ion is established. Then\, the model is verified by comparing it with the
classical capillary model and Kozeny particle model. The new model provide
s theoretical support for flow prediction of hydrate porous media.\n\nhttp
s://events.interpore.org/event/25/contributions/3869/
LOCATION:
URL:https://events.interpore.org/event/25/contributions/3869/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Promises\, Challenges and Prospects of Deep Learning for Providing
Insight into Multi-phase Flow Through Porous Media
DTSTART;VALUE=DATE-TIME:20210531T074000Z
DTEND;VALUE=DATE-TIME:20210531T075500Z
DTSTAMP;VALUE=DATE-TIME:20221001T014141Z
UID:indico-contribution-238-3867@events.interpore.org
DESCRIPTION:Speakers: Seyed Reza Asadolahpour (Institute of GeoEnergy Engi
neering\, Heriot-Watt University)\nThe advent of deep learning marked a mi
lestone in the real-life applicability of machine learning tools\, as now
very complex problems can be solved with unprecedented accuracy. Deep neur
al networks generally require little explicit prior knowledge and are dist
inctively efficient in extracting complicated patterns. These capabilities
turn them into feasible candidates for replacing and/or assisting convent
ional time-consuming and computationally-expensive methods involved in por
e-scale modelling\, such as reconstruction\, segmentation and single-/mult
i-phase simulations.\n\nThis work aims to show how the power of deep learn
ing can be harnessed to both estimate porous-media properties and develop
new insights. Our main objectives are: (1) provide a general overview of h
ow deep neural networks have already been used in terms of single/multi-ph
ase flow characterization\; (2) demonstrate the potentials of deep learnin
g in digital rock physics through case studies\; (3) discuss deep-learning
-based approaches to explore the physics of the porous media.\n\nFirst\, t
he relevant body of research is considered so that advancements\, gaps and
potentials can be identified. Then\, an implementation map is laid out\,
encompassing the simplest to most comprehensive applications. Inputs can r
ange from grey-level images to customized feature maps\, while targets can
span from static properties to complex\, dynamic multi-phase properties (
e.g.\, resistivity index and fluid distribution). Secondly\, case studies
are presented where porosity\, permeability and relative permeability are
predicted from micro-CT (e.g.\, synchrotron beamline) images and rock-flui
d characteristics. A great challenge is to achieve the simulations at repr
esentative sample image sizes\, which makes hyperparameter sweeping extrem
ely taxing for the researcher and demanding on the hardware.\n\nThirdly\,
future research is discussed. It is proposed that to develop reliable mult
i-phase predictors\, large databases must be synthesized by collecting\, r
esampling\, augmenting\, and grouping images and the corresponding propert
ies. Consequently\, deep neural networks can be trained for various rock t
ypes (e.g.\, carbonate) and processes (e.g.\, two-phase unsteady-state dra
inage). Singular or ensembles of networks may either be used to make predi
ctions or to serve as the base to be customized for other applications\, i
.e.\, transfer learning. Final models can be put to ultimate real-life tes
ting by comparing against experimental data\, e.g.\, phase distributions f
rom synchrotron imaging.\n\nRather than trying to create mere black-box es
timators\, one must strive to understand how the networks extract informat
ion\, by looking at layer architectures\, weights and other elements. The
goal should be to gain insights into various flow functions (e.g.\, uncove
r the link between macroscopic properties and pore morphology and/or wetta
bility) and the physics of certain flow behaviours (e.g.\, snap-off). This
has already been done in such fields as object recognition\, for instance
\, to figure out the level of feature abstraction at different layers. Fur
thermore\, since trained models are very fast to run\, they make perfect a
ssets for such tasks as sensitivity/uncertainty analysis and back-calculat
ion of input features\, for instance\, to see what wettability distributio
n can result in a specific flow parameter.\n\nhttps://events.interpore.org
/event/25/contributions/3867/
LOCATION:
URL:https://events.interpore.org/event/25/contributions/3867/
END:VEVENT
END:VCALENDAR