Computational Geosciences,
Год журнала:
2023,
Номер
27(2), С. 245 - 262
Опубликована: Янв. 31, 2023
Abstract
In
recent
years,
convolutional
neural
networks
(CNNs)
have
experienced
an
increasing
interest
in
their
ability
to
perform
a
fast
approximation
of
effective
hydrodynamic
parameters
porous
media
research
and
applications.
This
paper
presents
novel
methodology
for
permeability
prediction
from
micro-CT
scans
geological
rock
samples.
The
training
data
set
CNNs
dedicated
consists
labels
that
are
typically
generated
by
classical
lattice
Boltzmann
methods
(LBM)
simulate
the
flow
through
pore
space
segmented
image
data.
We
instead
direct
numerical
simulation
(DNS)
solving
stationary
Stokes
equation
efficient
distributed-parallel
manner.
As
such,
we
circumvent
convergence
issues
LBM
frequently
observed
on
complex
geometries,
therefore,
improve
generality
accuracy
our
set.
Using
DNS-computed
permeabilities,
physics-informed
CNN
(PhyCNN)
is
trained
additionally
providing
tailored
characteristic
quantity
space.
More
precisely,
exploiting
connection
problems
graph
representation
space,
additional
information
about
confined
structures
provided
network
terms
maximum
value,
which
key
innovative
component
workflow.
robustness
this
approach
reflected
very
high
accuracy,
variety
sandstone
samples
archetypal
formations.
Nature Communications,
Год журнала:
2023,
Номер
14(1)
Опубликована: Фев. 14, 2023
Abstract
Proton
exchange
membrane
fuel
cells,
consuming
hydrogen
and
oxygen
to
generate
clean
electricity
water,
suffer
acute
liquid
water
challenges.
Accurate
modelling
is
inherently
challenging
due
the
multi-phase,
multi-component,
reactive
dynamics
within
multi-scale,
multi-layered
porous
media.
In
addition,
currently
inadequate
imaging
capabilities
are
limiting
simulations
small
areas
(<1
mm
2
)
or
simplified
architectures.
Herein,
an
advancement
in
achieved
using
X-ray
micro-computed
tomography,
deep
learned
super-resolution,
multi-label
segmentation,
direct
multi-phase
simulation.
The
resulting
image
most
resolved
domain
(16
with
700
nm
voxel
resolution)
largest
flow
simulation
of
a
cell.
This
generalisable
approach
unveils
multi-scale
clustering
transport
mechanisms
over
large
dry
flooded
gas
diffusion
layer
fields,
paving
way
for
next
generation
proton
cells
optimised
structures
wettabilities.
Transport in Porous Media,
Год журнала:
2021,
Номер
140(1), С. 241 - 272
Опубликована: Май 29, 2021
Abstract
The
permeability
of
complex
porous
materials
is
interest
to
many
engineering
disciplines.
This
quantity
can
be
obtained
via
direct
flow
simulation,
which
provides
the
most
accurate
results,
but
very
computationally
expensive.
In
particular,
simulation
convergence
time
scales
poorly
as
domains
become
less
or
more
heterogeneous.
Semi-analytical
models
that
rely
on
averaged
structural
properties
(i.e.,
porosity
and
tortuosity)
have
been
proposed,
these
features
only
partly
summarize
domain,
resulting
in
limited
applicability.
On
other
hand,
data-driven
machine
learning
approaches
shown
great
promise
for
building
general
by
virtue
accounting
spatial
arrangement
domains’
solid
boundaries.
However,
prior
convolutional
neural
network
(ConvNet)
literature
concerning
2D
image
recognition
problems
do
not
scale
well
large
3D
required
obtain
a
representative
elementary
volume
(REV).
As
such,
work
focused
homogeneous
samples,
where
small
REV
entails
global
nature
fluid
could
mostly
neglected,
accordingly,
memory
bottleneck
addressing
with
ConvNets
was
side-stepped.
Therefore,
important
geometries
such
fractures
vuggy
modeled
properly.
this
work,
we
address
limitation
multiscale
deep
model
able
learn
from
media
data.
By
using
coupled
set
networks
view
domain
different
scales,
enable
evaluation
(
$$>512^3$$
>5123
)
images
approximately
one
second
single
graphics
processing
unit.
architecture
opens
up
possibility
modeling
sizes
would
feasible
traditional
tools
desktop
computer.
We
validate
our
method
laminar
case
samples
fractures.
result
viewing
entire
at
once,
perform
prediction
exhibiting
degree
heterogeneity.
expect
methodology
applicable
transport
play
central
role.
