The Journal of Chemical Physics,
Год журнала:
2024,
Номер
161(24)
Опубликована: Дек. 23, 2024
Graph
neural
network
interatomic
potentials
(GNN-IPs)
are
gaining
significant
attention
due
to
their
capability
of
learning
from
large
datasets.
Specifically,
universal
based
on
GNN,
usually
trained
with
crystalline
geometries,
often
exhibit
remarkable
extrapolative
behavior
toward
untrained
domains,
such
as
surfaces
and
amorphous
configurations.
However,
the
origin
this
extrapolation
is
not
well
understood.
This
work
provides
a
theoretical
explanation
how
GNN-IPs
extrapolate
geometries.
First,
we
demonstrate
that
can
capture
non-local
electrostatic
interactions
through
message-passing
algorithm,
evidenced
by
tests
toy
models
density-functional
theory
data.
We
find
GNN-IP
models,
SevenNet
MACE,
accurately
predict
forces
in
indicating
they
have
learned
exact
functional
form
Coulomb
interaction.
Based
these
results,
suggest
ability
learn
interactions,
coupled
embedding
nature
GNN-IPs,
explains
ability.
GNN-IP,
SevenNet-0,
effectively
infers
domains
but
fails
arising
kinetic
term,
which
supports
suggested
theory.
Finally,
address
impact
hyperparameters
performance
potentials,
SevenNet-0
MACE-MP-0,
discuss
limitations
capabilities.
Science and Technology of Advanced Materials Methods,
Год журнала:
2023,
Номер
3(1)
Опубликована: Окт. 12, 2023
Recently,
machine
learning
potentials
(MLPs)
have
been
attracting
interest
as
an
alternative
to
the
computationally
expensive
density-functional
theory
(DFT)
calculations.
The
data-driven
approach
in
MLPs
requires
carefully
curated
training
datasets,
which
define
valid
domain
of
simulations.
Therefore,
acquiring
datasets
that
comprehensively
span
desired
simulations
is
important.
In
this
review,
we
attempt
set
guidelines
for
systematic
construction
according
target
To
end,
extensively
analyze
sets
previous
literature
four
application
types:
thermal
properties,
diffusion
structure
prediction,
and
chemical
reactions.
each
application,
summarize
characteristic
reference
structures
discuss
specific
parameters
DFT
calculations
such
MD
conditions.
We
hope
review
serves
a
comprehensive
guide
researchers
practitioners
aiming
harness
capabilities
material
The Journal of Chemical Physics,
Год журнала:
2024,
Номер
161(24)
Опубликована: Дек. 23, 2024
Graph
neural
network
interatomic
potentials
(GNN-IPs)
are
gaining
significant
attention
due
to
their
capability
of
learning
from
large
datasets.
Specifically,
universal
based
on
GNN,
usually
trained
with
crystalline
geometries,
often
exhibit
remarkable
extrapolative
behavior
toward
untrained
domains,
such
as
surfaces
and
amorphous
configurations.
However,
the
origin
this
extrapolation
is
not
well
understood.
This
work
provides
a
theoretical
explanation
how
GNN-IPs
extrapolate
geometries.
First,
we
demonstrate
that
can
capture
non-local
electrostatic
interactions
through
message-passing
algorithm,
evidenced
by
tests
toy
models
density-functional
theory
data.
We
find
GNN-IP
models,
SevenNet
MACE,
accurately
predict
forces
in
indicating
they
have
learned
exact
functional
form
Coulomb
interaction.
Based
these
results,
suggest
ability
learn
interactions,
coupled
embedding
nature
GNN-IPs,
explains
ability.
GNN-IP,
SevenNet-0,
effectively
infers
domains
but
fails
arising
kinetic
term,
which
supports
suggested
theory.
Finally,
address
impact
hyperparameters
performance
potentials,
SevenNet-0
MACE-MP-0,
discuss
limitations
capabilities.
