Verilog
language
is
used
to
complete
the
RTL
modeling
of
high
efficiency
LSTM
accelerator
and
reconfigurable
CNN-LSTM
on
FPGA.
Through
comparing
calculation
results
hardware
software,
functional
correctness
designed
confirmed.
The
experimental
show
that
proposed
has
16
times
acceleration
ratio
CPU,
19.12%
power
consumption
GPU,
85.68
GOPS
throughput,
22.4
GOPS/W
energy
efficiency,
which
superior
other
designs
same
type.
Compared
with
can
achieve
12
ratio,
while
only
approximately
10.02%
GPU;
throughput
rate
reaches
77.5
GOPS,
42.9
GOPS/W.
In
application
background,
compared
efficient
accelerator,
on-chip
resource
reduced
decreasing
time
consumed
process
a
set
data
by
65%.
Scientific Reports,
Год журнала:
2025,
Номер
15(1)
Опубликована: Янв. 6, 2025
This
study
investigates
the
application
of
various
neural
network-based
models
for
predicting
temperature
distribution
in
freeze
drying
process
biopharmaceuticals.
For
heat-sensitive
biopharmaceutical
products,
is
preferred
to
prevent
degradation
pharmaceutical
compounds.
The
modeling
framework
based
on
CFD
(Computational
Fluid
Dynamics)
and
machine
learning
(ML).
ML
explored
include
Single-Layer
Perceptron
(SLP),
Multi-Layer
(MLP),
Fully
Connected
Neural
Network
(FCNN),
Deep
(DNN).
Model
optimization
achieved
through
Fireworks
Algorithm
(FWA).
Results
reveal
promising
performance
across
all
models,
with
MLP
demonstrating
highest
accuracy
both
test
training
datasets,
achieving
an
R2
score
0.99713
0.99717
respectively.
SLP
also
exhibits
strong
performance,
0.88903
dataset.
FCNN
DNN
perform
admirably,
scores
0.99158
0.99639
dataset
These
results
highlight
efficiency
network-driven
specifically
MLP,
precisely
forecasting
values
spatial
coordinates.
Additionally,
integration
model
refinement
yields
advantages
improving
predictive
these
models.
Energy & Fuels,
Год журнала:
2025,
Номер
39(9), С. 4549 - 4564
Опубликована: Фев. 19, 2025
The
axial
mixing/segregation
behavior
of
single
plastic
particles
in
a
bubbling
fluidized
bed
reactor
has
been
investigated
by
noninvasive
X-ray
imaging
techniques
the
temperature
range
500–650
°C
and
under
pyrolysis
conditions.
Experimental
results
showed
that
extent
mixing
between
particle
increases
as
both
fluidization
velocity
increase.
Three
modeling
approaches
were
proposed
to
describe
particle,
i.e.,
purely
mechanistic
model,
physics-informed
neural
network
(PINN),
an
augmented
PINN
(augPINN).
former
model
is
based
on
second
law
motion.
standard
PINN,
built
simply
embedding
motion
loss
function.
third
approach
involves
introduction
new
interphase
distribution
parameter,
P,
into
model.
This
parameter
represents
relative
importance
effects
emulsion
bubble
phases
particle.
was
obtained
training
using
displacement
data.
augPINN
shown
outperform
models
describing
polypropylene
particles.
Moreover,
P
found
be
physically
interpretable.
main
novelty
this
work
show
how
different
frameworks
concept
machine
learning
can
successfully
applied
complex
real-world
hydrodynamic
data
sets.
Micromachines,
Год журнала:
2025,
Номер
16(3), С. 350 - 350
Опубликована: Март 19, 2025
Thermal
analysis
is
an
indispensable
aspect
of
semiconductor
packaging.
Excessive
operating
temperatures
in
integrated
circuit
(IC)
packages
can
degrade
component
performance
and
even
cause
failure.
Therefore,
thermal
resistance
characteristics
are
critical
to
the
reliability
electronic
components.
Machine
learning
modeling
offers
effective
way
predict
IC
packages.
In
this
study,
data
from
finite
element
(FEA)
utilized
by
machine
models
during
package
testing.
For
two
types,
namely
Quad
Flat
No-lead
(QFN)
Thin
Fine-pitch
Ball
Grid
Array
(TFBGA),
derived
analysis,
employed
resistance.
The
values
include
θJA,
θJB,
θJC,
ΨJT,
ΨJB.
Five
models,
light
gradient
boosting
(LGBM),
random
forest
(RF),
XGBoost
(XGB),
support
vector
regression
(SVR),
multilayer
perceptron
(MLP),
applied
as
forecasting
study.
Numerical
results
indicate
that
model
outperforms
other
terms
accuracy
for
almost
all
cases.
Furthermore,
achieved
highly
satisfactory.
conclusion,
shows
significant
promise
a
reliable
tool
predicting
packaging
design.
application
techniques
these
parameters
could
enhance
efficiency
designs.
