Physics of Fluids,
Journal Year:
2024,
Volume and Issue:
36(12)
Published: Dec. 1, 2024
The
rapid
and
accurate
prediction
of
the
flow
field
during
supersonic
isolator
operation
is
crucial.
Deep
learning-based
pressure
monitoring
an
effective
method
for
prediction.
A
dataset
was
produced
a
ground-based
experiment
with
variable
incoming
Mach
number
back
pressure.
An
approach
predicting
future
based
on
proposed.
model
incorporating
long
short-term
memory,
temporal
convolutional
network,
block
attention
module
structures
has
been
performance
proposed
analyzed
compared
those
other
time-series
neural
networks
location
shock
train
leading
edge
introduced
as
priori
information
to
enhance
performance.
impact
weights
associated
in
network
training
discussed.
This
study
presents
optimization
scheme
models
specifically
tailored
problem.
Journal of Applied Physics,
Journal Year:
2025,
Volume and Issue:
137(2)
Published: Jan. 10, 2025
Piezoelectric
semiconductors
(PSCs)
are
crucial
in
micro-electromechanical
systems,
but
analyzing
their
size
effects
and
accurately
determining
flexoelectric
parameters
is
challenging
due
to
the
complexity
of
multi-scale
multi-field
coupling.
Physics-informed
neural
networks
(PINNs),
which
merge
physical
laws
with
machine
learning,
provide
a
promising
approach
for
solving
partial
differential
equations
parameter
inversion.
In
this
paper,
we
develop
PINN
model
solve
system
fourth-order
PSC
nanowires,
accounting
strain
gradient
effects.
Predictions
by
closely
match
results
from
traditional
numerical
methods.
Additionally,
minimal
labeled
data,
can
predict
both
solutions
material
parameters,
such
as
coefficient.
It
expected
that
PINNs
offer
an
effective
method
nanowires
inverting
key
properties.
Physics of Fluids,
Journal Year:
2025,
Volume and Issue:
37(1)
Published: Jan. 1, 2025
The
seepage
equation
is
essential
for
understanding
fluid
flow
in
porous
media,
crucial
analyzing
behavior
various
pore
structures
and
supporting
reservoir
engineering.
However,
solving
this
under
complex
conditions,
such
as
variable
well
rates,
poses
significant
challenges.
Although
physics-informed
neural
networks
have
been
effective
addressing
partial
differential
equations,
they
often
struggle
with
the
complexities
of
physical
phenomena.
This
paper
presents
an
improved
method
using
asymptotic
solution
nets
combined
scaling
before
activation
(SBA)
gradient
constraints
to
solve
media
varying
rates
without
labeled
data.
model
consists
two
networks:
one
that
approximates
another
corrects
approximation
errors
ensure
both
mathematical
accuracy.
When
rate
changes,
network
may
fail
fully
satisfy
due
pressure
distribution
variations,
resulting
sub-optimal
outcomes.
To
address
this,
we
incorporate
information
into
loss
function
reinforce
utilize
SBA
enhance
approximation.
derived
from
at
previous
rate,
regulates
weight
adjustments
through
adjustment
coefficient
constrained
by
function,
preventing
local
minima
during
optimization.
Experimental
results
show
our
achieves
accuracy
range
10−4
10−2.
