Landslides,
Год журнала:
2023,
Номер
20(9), С. 1853 - 1863
Опубликована: Май 27, 2023
Abstract
In
this
study,
a
new
paradigm
compared
to
traditional
numerical
approaches
solve
the
partial
differential
equation
(PDE)
that
governs
thermo-poro-mechanical
behavior
of
shear
band
deep-seated
landslides
is
presented.
particular,
paper
shows
projections
temperature
inside
as
proxy
estimate
catastrophic
failure
landslides.
A
deep
neural
network
trained
find
temperature,
by
using
loss
function
defined
underlying
PDE
and
field
data
three
To
validate
network,
we
have
applied
following
cases:
Vaiont,
Shuping,
Mud
Creek
The
results
show
that,
creating
training
with
synthetic
data,
landslide
can
be
reproduced
allows
forecast
basal
case
studies.
Hence,
providing
real-time
estimation
stability
landslide,
other
solutions
whose
study
has
calculated
individually
for
each
scenario.
Moreover,
offers
novel
procedure
design
architecture,
considering
stability,
accuracy,
over-fitting.
This
approach
could
useful
also
applications
beyond
Computer Methods in Applied Mechanics and Engineering,
Год журнала:
2024,
Номер
428, С. 117063 - 117063
Опубликована: Июнь 4, 2024
The
two
fundamental
concepts
of
materials
theory,
pseudo
potentials
and
the
assumption
a
multiplicative
decomposition,
allow
general
description
inelastic
material
behavior.
increase
in
computer
performance
enabled
us
to
thoroughly
investigate
predictive
capabilities
ever
more
complex
choices
for
potential
Helmholtz
free
energy.
Today,
however,
we
have
reached
point
where
their
models
are
becoming
increasingly
sophisticated.
This
raises
question:
How
do
find
best
model
that
includes
all
effects
explain
our
data?
Constitutive
Artificial
Neural
Networks
(CANN)
may
answer
this
question.
Here,
extend
CANNs
(iCANN).
Rigorous
considerations
objectivity,
rigid
motion
reference
configuration,
decomposition
its
inherent
non-uniqueness,
choice
appropriate
stretch
tensors,
restrictions
energy
potential,
consistent
evolution
guide
towards
architecture
iCANN
satisfying
thermodynamics
per
design.
We
combine
feed-forward
networks
with
recurrent
neural
network
approach
take
time
dependencies
into
account.
Specializing
visco-elasticity,
demonstrate
is
capable
autonomously
discovering
artificially
generated
data,
response
polymers
at
different
rates
cyclic
loading
as
well
relaxation
behavior
muscle
data.
Since
design
not
limited
iCANNs
might
help
identify
phenomena
subsequently
select
most
model.
focus
on
providing
thermodynamically
framework
behaviors
how
incorporate
an
architecture-based
manner.
Our
source
code,
examples
available
Holthusen
et
al.
(2023a)
(
https://doi.org/10.5281/zenodo.10066805).
Batteries,
Год журнала:
2023,
Номер
9(6), С. 301 - 301
Опубликована: Май 30, 2023
Accurate
forecasting
of
the
lifetime
and
degradation
mechanisms
lithium-ion
batteries
is
crucial
for
their
optimization,
management,
safety
while
preventing
latent
failures.
However,
typical
state
estimations
are
challenging
due
to
complex
dynamic
cell
parameters
wide
variations
in
usage
conditions.
Physics-based
models
need
a
tradeoff
between
accuracy
complexity
vast
parameter
requirements,
machine-learning
require
large
training
datasets
may
fail
when
generalized
unseen
scenarios.
To
address
this
issue,
paper
aims
integrate
physics-based
battery
model
machine
learning
leverage
respective
strengths.
This
achieved
by
applying
deep
framework
called
physics-informed
neural
networks
(PINN)
electrochemical
modeling.
The
charge
health
cells
predicted
integrating
partial
differential
equation
Fick’s
law
diffusion
from
single
particle
into
network
process.
results
indicate
that
PINN
can
estimate
with
root
mean
square
error
range
0.014%
0.2%,
has
1.1%
2.3%,
even
limited
data.
Compared
conventional
approaches,
less
still
incorporating
laws
physics
process,
resulting
adequate
predictions,
situations.
