Physics of Fluids,
Journal Year:
2025,
Volume and Issue:
37(3)
Published: March 1, 2025
This
work
presents
a
large
language
model
(LLM)-based
agent
OpenFOAMGPT
tailored
for
OpenFOAM-centric
computational
fluid
dynamics
(CFD)
simulations,
leveraging
two
foundation
models
from
OpenAI:
the
GPT-4o
(GPT
means
Generative
Pre-trained
Transformer)
and
chain-of-thought–enabled
o1
preview
model.
Both
agents
demonstrate
success
across
multiple
tasks.
While
price
of
token
with
is
six
times
as
that
GPT-4o,
it
consistently
exhibits
superior
performance
in
handling
complex
tasks,
zero-shot/few-shot
case
setup
to
boundary
condition
modifications,
zero-shot
turbulence
adjustments,
code
translation.
Through
an
iterative
correction
loop,
efficiently
addressed
single-phase
multiphase
flow,
heat
transfer,
Reynolds-averaged
Navier–Stokes
modeling,
eddy
simulation,
other
engineering
scenarios,
often
converging
limited
number
iterations
at
low
costs.
To
embed
domain-specific
knowledge,
we
employed
retrieval-augmented
generation
pipeline,
demonstrating
how
preexisting
simulation
setups
can
further
specialize
subdomains
such
energy
aerospace.
Despite
great
agent,
human
oversight
remains
crucial
ensuring
accuracy
adapting
shifting
contexts.
Fluctuations
over
time
suggest
need
monitoring
mission-critical
applications.
Although
our
demonstrations
focus
on
OpenFOAM,
adaptable
nature
this
framework
opens
door
developing
LLM-driven
into
wide
range
solvers
codes.
By
streamlining
CFD
approach
has
potential
accelerate
both
fundamental
research
industrial
advancements.
Research Square (Research Square),
Journal Year:
2024,
Volume and Issue:
unknown
Published: May 29, 2024
Abstract
Designing
active-flow-control
(AFC)
strategies
for
three-dimensional
(3D)
bluff
bodies
is
a
challenging
task
with
critical
industrial
implications.
In
this
study
we
explore
the
potential
of
discovering
novel
control
drag
reduction
using
deep
reinforcement
learning.
We
introduce
high-dimensional
AFC
setup
on
3D
cylinder,
considering
Reynolds
numbers
(Re_D)
from
100
to
400,
which
range
including
transition
wake
instabilities.
The
involves
multiple
zero-net-mass-flux
jets
positioned
top
and
bottom
surfaces,
aligned
into
two
slots.
method
relies
coupling
computational-fluid-dynamics
solver
multi-agent
reinforcement-learning
(MARL)
framework
based
proximal-policy-optimization
algorithm.
MARL
offers
several
advantages:
it
exploits
local
invariance,
adaptable
across
geometries,
facilitates
transfer
learning
cross-application
agents,
results
in
significant
training
speedup.
For
instance,
our
demonstrate
21%
Re_D=300,
outperforming
classical
periodic
control,
yields
up
6%
reduction.
To
authors'
knowledge,
present
MARL-based
represents
first
time
where
conducted
cylinders.
This
breakthrough
paves
way
conducting
progressively
more
complex
turbulent-flow
configurations.
Physics of Fluids,
Journal Year:
2025,
Volume and Issue:
37(1)
Published: Jan. 1, 2025
We
study
large
eddy
simulation
(LES)
in
the
form
of
vorticity
transport
equations
(VTE),
employing
six
subgrid-scale
(SGS)
models,
including
dynamic
Smagorinsky
model,
mixed
velocity
gradient
scale
similarity
approximate
deconvolution
and
iterative
(DIAD)
model.
In
a
priori
study,
correlation
coefficient
SGS
stress
given
by
DIAD
is
significantly
higher
than
those
other
structural
relative
error
lowest.
posteriori
validation,
models
outperform
functional
predicting
energy
spectra,
enstrophy
probability
density
functions
(PDFs)
vorticity,
strain-rate
tensor,
flux,
production
term.
These
results
confirm
feasibility
VTE-based
LES.
They
also
indicate
that
classic
modeling
approaches
for
terms
filtered
Navier–Stokes
(NSE)
are
applicable
to
counterparts
VTE.
Journal of Computational Physics,
Journal Year:
2024,
Volume and Issue:
522, P. 113577 - 113577
Published: Nov. 15, 2024
We
propose
a
new
neural
network
based
large
eddy
simulation
framework
for
the
incompressible
Navier-Stokes
equations
on
paradigm
"discretize
first,
filter
and
close
next".This
leads
to
full
model-data
consistency
allows
employing
closure
models
in
same
environment
as
where
they
have
been
trained.Since
LES
discretization
error
is
included
learning
process,
can
learn
account
discretization.Furthermore,
we
employ
divergence-consistent
discrete
defined
through
face-averaging
provide
novel
theoretical
numerical
analysis.This
preserves
divergence-free
constraint
by
construction,
unlike
general
filters
such
volume-averaging
filters.We
show
that
using
formulation
coupled
with
convolutional
model
produces
stable
accurate
results
both
a-priori
a-posteriori
training,
while
(divergence-inconsistent)
requires
training
or
other
stabilityenforcing
measures.
