►
Description
NERSC Data Seminars Series: https://github.com/NERSC/data-seminars
Title:
Artificial Design of Porous Materials
Speaker:
Jihan Kim, Department of Chemical and Biomolecular Engineering, KAIST
Abstract:
In this presentation, I will explore the new trend of designing novel porous materials using artificial design principles. I will talk about using our in-house developed generative adversarial network (GAN) software to create (for the first time) porous materials. Moreover, we have successfully implemented inverse-design in our GAN prompting ways to train our AI to create porous materials with user-desired methane adsorption capacity [1]. Next, we incorporate machine learning with genetic algorithm to design optimal metal-organic frameworks suitable for many different applications including methane storage and gas separations [2-3]. Finally, we demonstrate usage of text mining to collect wealth of data from published papers to predict optimal synthesis conditions for porous materials [4]. Overall, machine learning and artificial design can accelerate the materials discovery and expedite the process to deploy new materials for many different applications.
Bio:
Jihan Kim is an associate professor at KAIST (Korea Advanced Institute of Science and Technology). He received his B.S. degree in Electrical Engineering and Computer Science (EECS) at UC Berkeley in 1997 and received his M.S. and Ph.D. degrees in Electrical and Computer Engineering at University of Illinois at Urbana-Champaign in 2004 and 2009, respectively. He worked as a NERSC postdoc in the Petascale Post-doc project from 2009 to 2011 and worked as postdoctoral researcher in UC Berkeley/LBNL with Prof. Berend Smit from 2011 to 2013. His current research at KAIST focuses on using molecular simulations and machine learning methods to design novel porous materials (e.g. zeolites, MOFs, porous polymers) for various energy and environmental related applications (e.g. gas storage, gas separations, catalysis, sensors). He has published over 100 papers and has over 7000 Google Scholar citations.
Host of Seminar:
Brian Austin, Advanced Technologies Group
National Energy Research Scientific Computing Center (NERSC)
Lawrence Berkeley National Laboratory
A
A
B
B
Sorry
so
I
received
my
bs
from
uc
berkeley.
I
was
in
the
x
department
and
I
got
my
ms
and
phd
from
urbana-champaign
and,
as
brian
mentioned,
I
was
a
part
of
the
petascale
postdoc
and
nurse
from
2009
to
2011,
and
that
was
a
great
opportunity
for
me
to
work
with
the
scientists
at
uc
berkeley,
in
particular
professor
baron
smith
and
professor
martin,
head
gordon,
and
that's
where
we
connected
their
scientific
knowledge
to
our
super
computing.
B
I
guess
facilities
as
well
as
our
skills,
and
that's
where
I
learned
a
lot
from
regarding
computing,
as
well
as
the
the
chemistry
and
the
science
and
after
the
pedagogy
postdoc.
I
stayed
with
professor
barron
smith
as
a
postdoc
at
uc,
berkeley
and
berkeley
lab,
and
after
that
I
got
my
position
at
kaist.
B
So
I'm
interested
in
a
lot
of
different
multi-scale
methods
when
it
comes
to
chemistry
and
materials.
So
in
particular
we
look
into
various
quantum
chemical
methods
such
as
gft,
mp2,
monte,
carlo
mp2
and
so
forth,
and
then
a
lot
of
the
classical
molecular
simulation
methods
such
as
gcmc
molecular
dynamics
and
so
forth
and
in
korea.
B
Also
these
days,
ai
and
machine
learning
is
very
trendy,
so
we're
looking
a
lot
into
various
machine
learning
techniques
which
I'll
cover
in
this
talk
and
for
the
materials
that
we're
modeling,
we
model
many
different
materials,
but
for
the
most
part
I
focus
on
porous
materials
such
as
zeolites
and
moss
and
so
forth.
But
we
do
other
2d
materials
as
well,
and
the
applications
include
many
of
the
environmental
and
energy
related
applications.
B
So
my
research
lab
and
research-
I
guess
career
these
days-
can
be
sort
of
divided
into
four
different.
Subcategories
first
is
where
we
look
into
applications,
of
course,
materials
and
in
this
realm,
what
we
do
is
we
work
closely
with
experimentalists
and
we
concentrate
on
just
modeling
one
specific
material
that
has
very
interesting
properties.
B
So
this
is
a
type
where
we
just
focus
on
one
material
and
one
material.
Only
and
the
second
type
of
material
sub
project
is
where
we
do
next
generation
materials
genome.
Where,
instead
of
doing
one
material,
we
do
thousands
or
sometimes
millions
of
materials
and
we
find
ways
to
screen
for
the
best
properties
and
we
collaborate
with
experimentalists
on
sort
of
trying
to
synthesize
the
best
materials
that
we
covered
from
our
screening
procedure.
B
So
this
is
the
picture
we
took
and
currently
I'm
on
a
sabbatical
and
I'm
in
uc
san
diego
right
now.
But
this
was
a
picture
that
we
took,
and
this
is
my
research
lab
just
before
I
left.
So
things
were
very
sort
of
unique,
as
you
all
know,
during
the
kova
stages,
but
luckily
for
computational
researchers.
B
So
here's
the
outline
of
the
talk
briefly
I'll
go
over
what
force
materials
are
and
then
I'll
go
over
different
stages,
of
course,
materials.
Sorry
this!
I
think
this
is
sort
of
on
auto
sort
of
going
through
pages,
so
it
seems
like
it's
going
on
its
own,
so
I'll
try
to
control
it,
but
it
might
be
confusing
so
I'll,
introduce
the
porous
materials
first,
so
introduction
to
porous
materials.
B
So
nano
coarse
materials
have
pores
on
the
order
of
few
angstroms
to
nanometers
and
different
classes
of
these
forced
materials
exist,
such
as
zeolites
metal,
organic
frameworks,
covalent
organic
frameworks
and
so
forth.
So
the
one
commonality
amongst
these
porous
materials
is,
as
the
name
suggests
they
have
these
pores
that
are
on
that,
can
sort
of
capture
different
types
of
gas
molecules,
store
them
and
use
them
for
different
applications,
and
because
they
have
these
millions
of
these
small
pores,
they
have
very
large
surface
area,
and
these
pores
can
be
highly
tunable.
B
So
some
of
the
applications
of
nano
force
materials
are
sensors.
Where
you
have
these
force
materials,
and
these
gas
molecules
can
approach
and
absorb
onto
these
materials,
and
as
it
does,
you
can
have
a
chemical
resistive
response
such
that
you
can
have
sensors
for
toxic
gas
molecules
and
so
forth,
and
second
is
catalysis,
where
you
create
different
chemical
compounds
from
your
inputs,
and
these
types
of
reactions
typically
happen
inside
the
pore,
and
these
types
of
materials
can
facilitate
these
reactions
such
that
you
get
the
chemical
products
that
you
want.
B
Even
before
all
the
sort
of
the
machine
learning
and
data
science
was
a
fad.
The
number
of
course
materials
which
extends
in
this
graph
up
to
2010,
has
been
increasing
rapidly,
so
these
are
sort
of
thousands
of
structures,
new
structures
that
are
synthesized,
and
this
is
just
for
one
class
of
forced
materials
called
moss.
B
So
currently,
the
number
of
experimentally,
synthesized
moss
is
over
a
hundred
thousand
and
then
there's
millions
of
hypothetical
moths,
hypothetical
force
materials
that
people
have
been
sort
of
creating
in
silico
inside
the
computer.
So
you
know,
even
before
the
data
science
thing
sort
of
took
off
in
its
sort
of
brand,
there
were
a
lot
of
data,
at
least
in
terms
of
material
science.
With
regards
to
these
course
materials.
B
So
one
research
goal
in
our
group
is
to
sort
of
create
new
course
materials
in
silico
such
that
we
can
facilitate
the
materials
development
relay
this
information
to
the
experimentalist
and
then
hopefully
we
can
synthesize
useful
materials
fast.
So
what
we
did
was
we
sort
of
looked
into
different
classes,
of
course,
materials,
and
we
wanted
to
sort
of
concentrate
on
groups
of
materials
that
have
yet
to
be
synthesized.
So
this
diagram
is
might
be
a
little
bit
outdated.
B
B
So
this
is
sort
of
the
process
that
procedure
that
most
of
us
have
been
doing
for
the
last
decades
or
so,
but
now
there's
a
parallel
approach
where,
instead
of
sort
of
thinking
about
how
to
design
new
algorithm
for
new
materials,
we
just
give
it
a
lot
of
data
because
there's
a
lot
of
data
of
both
experimental
and
hypothetical
materials
and
then
see
if
the
neural
network
can
create
new
materials.
So
I
think
this
is
sort
of
a
parallel
path.
B
That
looks
like
a
lot
of
the
previous
big
data
as
opposed
to
creating
something
that's
just
sort
of
totally
new
and
out
there.
So
I
guess
in
that
sense
the
old
techniques
better,
but
perhaps
in
10
years,
or
so
it
might
be
different
so
before
going
into
artificial
design,
I'll
just
sort
of
introduce
sort
of
one
natural
design
approach
and
one
case
study
from
our
group.
B
Okay,
to
contrast
with
the
artificial
design-
and
this
is
what
we
dubbed
as
the
nano
porous
materials
mashup
project
and
for
those
of
you
who
are
familiar
with
the
word
mashup.
It
used
in
music
when
it
comes
to
production,
where
you
have
two
different
songs
or
two
different
music
or
two
or
more,
and
you
integrate
them
or
sort
of
overlap
them
on
top
of
one
another
and
then
sort
of
do
some
editing
and
play
it
sort
of
simultaneously
such
that
the
newly
recorded
music
sounds
very
sort
of
unique
and
new.
B
And
similarly,
what
we
did
was
we
have
a
sort
of
a
large
full
of
experimentally,
synthesized,
forced
materials,
for
example,
material,
a
here
which
is
small,
five
material
b
here,
which
is
some
other
moth,
and
we
try
to
overlap
them
on
top
of
one
another.
And
usually
this
leads
to
a
lot
of
collisions
between
the
atoms.
B
So
you
don't
get
something:
that's
thermodynamically
stable,
but
if
we
have
millions
of
materials-
and
we
do
all
these
sort
of
paralyzed
pairings,
then
accidentally,
you
might
stumble
onto
a
pair
of
materials
that
can
sort
of
fit
in
very
well
such
that
it
would
be
thermodynamically
stable
and
we're
thinking
we.
So
the
algorithm
was
to
sort
of
identify
these
matched
up
materials
and
to
see
whether
we
can
synthesize
them.
B
So
we
started
with
over
80
000
materials
and
we
used
our
mashup
algorithm
to
sort
of
pairwise
connect
all
these
materials
and
see
which
one
of
these
can
hetero
interpenetrate
and
the
sort
of
the
sample
image
of
hetero
interpenetrated
unit
cell
looks
something
like
this.
Where
blue
is
one
material
and
red
is
the
second
material,
and
we
provided
like
a
couple
of
these
heterointerpentry
materials
to
experimental
group
and
unfortunately,
they
couldn't
synthesize
them.
B
So
instead
we
we
sort
of
looked
into
other
types
of
materials,
we're
thinking
that
with
some
tweaking,
we
can
go
into
these
hetero
interpenetrated
force,
moths
and
then
just
sort
of
move
on
to
core
shell
and
core
shell.
The
algorithm
is
kind
of
similar
in
the
sense
that
the
two
materials
need
to
have
similar
lattice
constants
such
that
they
can
sort
of
match
well,
and
then
you
can
have
you
need
to
have
sort
of
a
good
sort
of
connecting
points
between
material
and
material
b
such
that
they
can
chemically
bond.
B
The
good
thing
about
these
core
shell
materials
is
that
you
don't
have
to
worry
about
sort
of
this
comfort
on
this
sort
of
task
of
having
diffusion
of
the
second
moth
into
the
first
month,
which
was
very
difficult
in
the
experimental
synthesis,
because
here
you
can
do
a
sequential
reaction
where
you
just
sort
of
synthesize
the
core
moth
and
then
after
that,
just
synthesize
the
shell
move
on
the
top.
So
it's
easier!
B
So
that's
what
we
did.
What
we
did
was
with
these
sort
of
80
000
moths.
We
sort
of
did
a
sort
of
a
manual
fingerprint
on
where
the
chemical
connection
points
would
be.
So
we
took
a
unicell
a
first
moth.
We
cut
them
in
a
plane,
identify
these
atoms,
where
the
second
moth
would
connect.
For
example,
this
is
a
carboxylate
linker
with
the
oxygen,
and
so
this
is
where
the
oxygen
atoms
will
be
located.
B
So
this
is
a
sample
moth,
where
we
have
one
mothmaf5.
The
ligand
oxygen
is
located
here.
The
second
moth
called
h,
1,
the
methyl
atoms
are
located
here.
So
if
you
overlap
this
unit
cell
on
top
of
one
another,
they
would
fit
nicely
such
that
they're
on
top
of
one
another,
and
if
you
look
at
that
from
another
view,
they're
sort
of
clicking,
like
sort
of
like
a
magnets
clicking
on
sort
of
on
the
other
part
of
the
magnet
and
just
fitting
perfectly
so
inside
the
computer.
B
We
can
see
that
these
two
mosques
would,
in
principle,
connect
very
well,
so
we
try
synthesizing
some
of
these
pairs.
This
is
what
they
look
like
inside
the
computer,
with
blue
being
the
first
moth
and
the
yellow
being
the
second
moth,
and
then
we
see
that
we
get
these
really
nice,
looking
experimentally
on
coil,
shell
maf
or
you
have
the
for
example.
B
So
all
some
of
this
during
the
research
process,
it
worked
so
well
that
we
thought
that
maybe
everything
will
work
well,
but
that
wasn't
what
was
reported.
So
what
we
did
was
we
purposefully
took
two
materials
that
have
bad
connection
points.
So
you
see
that
when
we
sort
of
overlap
these
two
mobs
on
top
of
one
another,
they
don't
match
well,
and
then
we
tried
experimentally,
synthesizing
them
and
they
look
messy.
B
So
it
wasn't
the
case
that
we
get
these
nice
clean
structures
all
the
time
we
need
to
judiciously
choose
the
two
sort
of
core
and
the
shell
mods.
That
would
connect
with
one
another
very
well,
so
that
was
sort
of
sort
of
very
nice
and
we
sort
of
moved
on
to
this
type
of
work
where,
instead
of
doing
3d
three
mod,
3d
moth
connection,
we
do
3d
mop
and
2d
moth
connection
sort
of
like
this,
and
these
2d
mobs
are
sensor
moths.
B
So
right
now
we
have
some
of
these
materials
that
we
sort
of
predict
inside
our
computer
and
they
can
be
used
as
part
of
the
chemi
resistive
sensor.
So
we
work
with
these
people
who
are
sort
of
creating
these
modules,
such
that
the
materials
that
we
just
design
is
inside
the
sensor
interface.
This
whole
thing,
the
smart
cart
sensor.
Module
goes
inside
your
cell
phone,
such
that.
B
If
it
detects
certain
gases
like
no2
or
co
or
something,
then
it
will
give
a
signal
so
we're
working
on
this
with
collaboration
with
samsung
right
now,
it's
pretty
exciting.
How
sort
of
we
can
build
these
materials
inside
the
computer
and
have
them
sort
of
be
ported
inside
these
sort
of
electronic
devices?
B
So
that's
very
interesting,
so
that
was
an
example
of
natural
design
where
we
came
up
with
an
algorithm
to
connect
two
materials,
but
we
want
to
sort
of
in
certain
we're
kind
of,
I
guess
we're
all
kind
of
lazy
in
certain
sense.
So
we
want
to
to
let
the
machine
do
all
the
work
and
creating
new
materials,
because
it's
hard
to
come
up
with
a
good
design.
B
B
Like
pictures
and
all
sort
of
other
data,
where
you
have
a
lot
of
sort
of
these
input
images
and
then
it
can
create
new
images,
but
this
is
sort
of
easy.
This
wasn't
that
sort
of
easy
for
a
lot
of
these
materials
initially,
and
we
had
a
lot
of
problems
with
that
and
it
wasn't
due
to
lack
of
data
in
my
opinion,
because
there's
millions
of
these
materials
available.
So
that
was,
I
think,
sufficient
enough.
B
B
So
all
you
have
to
do
is
sort
of
concentrate
on
these
two
atom
types
and
the
the
way
that
this
sort
of
tetrahedral
sio4
structure
is
bonded
to
one
another
that
will
the
topology
sort
of
dictates,
how
many
different
zeolites
there
would
be
so
essentially
chemically
these
pure
silica.
Zeolites
are
all
the
same.
It's
just
that
how
they're
linked
together
in
topology
is
what
sort
of
creates
a
wide
variety
of
them.
So
there's
computationally,
discovered
uls
is
very
large,
so
we,
this
was
something
that's
available
to
us.
B
So
we
focused
on
methane
assertion,
as
this
is
sort
of
one
of
the
easiest
ones
to
model,
and
this
has
its
application
value,
because
with
regards
to
energy
storage,
because
this
is
somewhat
outdated
but
usdoe-
they
were
interested
in
the
methane
storage
technology
for
these
vehicles,
because
the
cng
vehicles
that
run
on
natural
gas,
they
have
to
be
stored
in
very
sort
of
high
pressure
at
around.
I
think
700
bar
or
so,
and
that
we
wanted
to
sort
of
make
this
pressure
mild
conditions.
B
B
B
It
largely
has
two
neural
networks,
the
discriminator
and
the
generator-
and
I
think
the
analogy
or
the
case
study-
that
a
lot
of
people
use
is
for
these
sort
of
fake
versus
real
money,
and
the
idea
is
that
the
police
want
to
catch
the
thief.
That's
making
this
fake
money
and
the
thief
want
to
sort
of
create
more
realistic
money.
So
if
you
have
sort
of
a
neural
network
that
competes
with
one
another
through
adversarial
learning
after
a
while,
the
thief
will
be
very
good
at
sort
of
making
fake
money.
B
That
looks
just
like
realistic
money.
Okay,
so
that's
sort
of
the
model
that
we
look
for.
This
creation
is
that
we
wanted
to
sort
of
give
a
lot
of
these
sort
of
real
zeolites
to
our
neural
network
and
then
they
would
create
something.
But
initially
it
wouldn't
look
that
great,
so
discrimination
would
be
very
high
in
the
surface
and
then
so
that
neural
network
would
have
to
sort
of
get
better
at
creating
more
realistic
material
and
then
you'll
see
that
how
that
might
lead
to
better
zeolites
okay.
B
So
the
objective
of
this
project
was
sort
of
twofold.
We
wanted
to
sort
of
just
create
new
zeolites,
because
back
in
four
four
five
years
ago,
there
wasn't
anyone
who
managed
to
create
these
porous
materials
using
ai.
So
we
wanted
to
be
the
first
group
to
do
it
and
the
second
is
not
just
creating
just
any
old
zeolites,
but
we
wanted
the
zls
to
have
certain
user
desired.
Capacitor
property
such
that
for
methane
storage.
B
You
wanted
to
user
to
say
I
want
this
material
that
have
certain
range
of
methane
heat
of
distortion
and
then
the
zeolite
would
sort
of
give
you
that
range.
So
we
wanted
that
user
design
capability,
because
that's
sort
of
the
hallmark
of
this
sort
of
using
neural
network
is
the
inverse
design
where
you
give
this
property
that
you
want
the
material
to
have
and
the
neural
network
could
create
that
kind
of
material.
B
B
So
you
need
to
sort
of
enforce
this
criterion
somehow
and
so
forth.
So
initially
we
had
a
lot
of
difficulties
and
what
we
decided
to
do
was
we
decided
to
use
the
3d
shape
representation
where
we
would
sort
of
just
take
the
unit
cell
of
zeolites.
These
are
crystalline
materials,
so
these
small
unit
cells,
they
sort
of
expand
and
repeat
themselves
across
the
xyz
dimensions
and
which-
and
the
idea
is
with
the
z
neural
network-
can
create
the
unit
cell.
Then
we
can
just
use
that
for
expansion.
B
Sorry
about
the
laser
point,
the
property
shape
was
represented
as
green.
So
essentially
we
have
these
colored
images,
so
we
can
represent
all
of
these
sort
of
zeolites
in
that
kind
of
manner
where
the
red
yellow
represents
the
silicon
and
the
oxygen
and
the
green
represents
the
methane
assortment
sort
of
locations
where
the
methane
likes
to
be
so.
B
So
this
is
sort
of
a
sample
run
initially
the
generator
you
create
something
that
just
looks
like
noise
and
as
sort
of
the
this
emd
is
the
earth
mover
distance,
which
is
the
distance
between
real
and
generated
data
distributions.
So
we
sort
of
have
a
metric
to
separate
between
the
real
distributions
of
the
zeolite
and
the
fake
distribution
and
then
after
a
while,
as
the
number
of
steps
sort
of
increase,
you
see
that,
after
a
while
it
sort
of
starts
looking
like
a
realistic
zeolite.
B
So
this
is
the
generated
zeolites,
and
this
x
axis
is
silicon
to
o
oxygen
ratio
and
all
of
these
still
zeolites,
the
silicone
to
oxygen
ratio
should
be
0.5,
and
so
it's
peaked
here.
But
you
see
that
it's
creating
a
lot
of
these
zeolites
or
pseudo
zeolites
that
are
not
abiding
by
this
si2o
ratio.
And
if
you
look
at
some
sample
pictures,
you
see
that
the
ai
is
creating
something
that
looks
nothing
like
zeolites
and
then
sort
of
in
this
point.
B
Five
area
is
look
creating
something
that
looks
like
it
and
then
at
higher
ratio
again,
nothing
like
the
geodes,
so
we
generated
one
million
zeolites
using
the
ai
and
unfortunately,
only
around
120
or
20
of
them
were
real
zeolites
and
the
other
is
sort
of
all
bad,
poorly
created
materials.
So
there
needs
to
be
a
lot
of
work
done
still,
and
these
are
sort
of
sample,
zeolites
or
sample
materials
that
were
successful
zeolites.
B
B
B
Also
sorry-
and
this
is
sort
of
a
fake
zeolite-
this
isn't
really
a
zeolite,
but
it
created
something
like
this
and
it
matched
something
on
the
hypothetical
zeolite
database.
So
it
turns
out
that
even
in
the
hypothetical
zeolite
database
that
people
created
in
the
past
there's
something
that
warned
zeolites,
but
they
were
sort
of
mistakenly
added
there
and
the
neural
network
sort
of
saw
this
and
they
created
something
similar.
B
We
can
quickly
compute
the
heat
of
insertion
and
because
we
can
compute
the
heat
of
assertion,
we
can
use
a
penalty
function
such
that
if
the
user
desired
zeolite
has
a
heat
of
distortion
within
a
certain
range,
then
we
bias
the
creation
just
within
that
range
and
give
high
penalties
for
anything
that's
outside
of
it.
So
once
we
did
that
these
are
the
profiles
of
the
zls
created
before
this
penalty
function
and
after
the
penalty
function.
B
So
before
the
penalty
function,
you
see
that
the
purple
distribution
is
very
wide,
whereas
after
implementation
of
this
loss
function
is
peaked
at
around
18
to
22
kilojoules.
So
it's
taking
this
sort
of
in
the
creation
mode
and
then
just
sort
of
creating
the
zeolites
that
have
just
this
sort
of
narrow
range.
B
So
we
identify
we
looked
into
some
of
these
structures
and
eight,
I
think
around
eighty
percent
of
the
materials
with
this
user
defined
sort
of
criterion
had
this
heat
of
distortion
within
that
range.
So
this
demonstrated
that
we
cannot
only
just
create
zls,
but
we
can
create
user
desired
zeolites
as
well.
B
So
this
is
something
that
we
worked
on
for
zeolites
and
then
we
moved
on
to
moths,
which
are
chemically
more
complicated
but
they're
in
some
sense,
they're
modular,
and
they
have
high
tunability
okay
compared
to
zeolite
they're,
both
personal
in
both
parts
but
they're
chemically
complicated.
So
for
moss.
We
decided
that
we're
not
going
to
use
the
gaussian
representation
to
represent
the
position
of
each
of
the
atoms
that'll
be
too
cumbersome
and
that's
sort
of
a
sort
of
we
we're
sort
of.
B
We
found
this
to
be
so
difficult
that
we
thought
that
it
would
be
sort
of
good
to
use
the
modular
nature
of
these
materials
in
the
sense
that
these
mops
can
be
decomposed
into
metal
clusters
and
organic
linkers
that
are
stitched
together.
So,
instead
of
representing
all
these
are
metal
cluster
atoms,
one
by
one,
we
can
just
sort
of
lump
them
all
together
and
just
sort
of
represent
this
as
some
type
a
and
this
linker
as
some
type
b.
B
As
for
the
machine
learning
to
create
new
materials,
so
they
wouldn't
be
creating
these
metal
clusters.
These
linkers
from
scratch,
they'll
just
be
using
them
and
then
sort
of
stitching
them
in
a
different
manner
to
sort
of
build
a
wider
range
of
these
mouse
structures
and
we
sort
of
accumulated
about
3000
topologies
from
a
database
such
that
we
have
a
wide
range
of
metal
nodes.
B
Organic
linkers
and
these
mathematical
topologies
and
sort
of
the
schematics
to
create
new
mods
is
something
like
this.
Initially,
we
generate
a
bunch
of
hypothetical
moths
and
then
we
compute
a
property
that
we
want
for
mc
simulation.
So
we
did
this
for
methane.
We
did
this
for
hydrogen.
We
did
this
for
xenon
krypton
separation,
but
I'll
just
concentrate
on
methane
for
now.
B
So
once
you
have
this,
you
want
to
sort
of
go
over
a
very
large
space
of
moss
to
create
the
materials
that
have
very
good
methane
storage
capacities
and
to
sort
of
compute
the
methane
storage
capacities.
We
need
to
do
grand
canonical,
monte,
carlo
simulations
and
they
might
take
minutes
or
so,
but
within
this
framework
of
keep
on
generating
new
materials,
doing
a
monte
carlo
simulation
and
then
sort
of
going
back
and
forth
even
that
small
sort
of
wall
time
of
15
to
20
minutes
was
long.
B
So
we
had
a
we
added
a
genetic
algorithm
flow
to
it,
such
that
during
the
genetic
algorithm
stages.
You
sort
of
swap
a
lot
of
these
metal
nodes
linkers
with
one
another
and
then
use
this
sort
of
neural
network
to
compute
the
methane
working
capacity,
and
the
idea
is
that
the
ones
that
are
sort
of
high
gets
passed
on
to
the
next
generation
and
then
you
sort
of
keep
on
doing
this
sort
of
iteratively.
B
So
if
you
do
all
these
combinations
and
if
you
can
allow
for
multiple
metal
nodes
and
multiple
edges
such
that,
you
can
have
two
different
kinds
of
metals
and
two
different
kinds
of
linkers
within
a
single
material
which
is
commonly
found
in
these
materials,
then
I
think
the
combination
was
very
large
10
to
14
or
something
so
with
the
brut
for
screening
for
methane
storage.
The
idea
that
you
go
through
all
these
200
trillion
hypothetical
mass
structures
was
very
cumbersome.
B
So
we
wanted
to
sort
of
use
a
genetic
algorithm
for
this
so
to
access
or
assess,
evaluate
methane
storage.
We
use
monte
carlo
simulation
and
we
use
a
gpu
monte
carlo
simulation
tool,
the
model
and
the
parametry
we
used
for
it
was
the
classical
leonard
jones
potential
universal
force
field
for
moth
trap
b
for
methane
and
the
target
met.
Property
is
methane.
B
Working
capacity,
which
is
for
uptake
at
65
bar
uptake
is
5.8
bar,
and
that
is
the
condition
for
the
methane
storage
for
these
ang
vehicles
and
the
temperature
is
room
temperature
at
298k
and
sort
of
that.
One
component
was
to
build
a
neural
network
that
can
predict
the
methane
working
capacity
such
that
when
it's
part
of
the
genetic
algorithm
and
part
of
the
component
even
minutes
is
very
long
time.
So
we
wanted
this
to
sort
of
output
in
seconds,
so
we
train
the
about.
B
I
think
50
000
moths,
computed
gcmt
simulations
with
the
working
capacity.
So
that's
the
working
capacities
on
the
x-axis
and
our
prediction
of
the
machine
learning
model
is
the
y-axis.
So
you
see
that
it's
relatively
well.
It
predicts
sort
of
at
the
y
equals
x.
So
with
this
tool,
then
we
can
use
that
to
just
output
what
the
methane
working
capacity
would
be
for
an
input
month.
B
B
B
So
the
final
sort
of
phase
is
the
synthesis
part.
Okay,
so
a
lot
of
times
with
porous
materials.
The
biggest
stumbling
block
when
we
work
with
experimentalists
is
that
we
have
this
very
beautiful
and
nice
structure
that
theoretically,
should
be
sort
of
thermodynamically,
stable,
everything's,
fine,
but
we
can't
synthesize
them
or
the
experimentalists
there's
many
different
reasons.
They
say
something
like
they're,
not
getting
crystalline
materials.
There's
these
ligands
are
difficult
to
sort
of
synthesize
and
they're,
getting
different,
topologies
and
so
forth.
B
So
this
is
sort
of
a
common
stumbling
block
these
days
amongst
all
sort
of
computational
materials.
Research
where
you
get
to
the
end
in
identifying
the
materials
that
have
great
properties,
but
we
can't
synthesize
them
so
and
conventionally
there
aren't
too
many
or
sort
of
conventional
molecular
simulation
techniques
that
model
synthesis.
It's
complicated
using
various
sort
of
various
sort
of
sort
of
these
precursors
and
all
that
stuff
with
different
temperature
and
there's
a
lot
of
complicated
dynamics
going
on,
so
that
it's
very
difficult
to
model
sort
of
synthesis.
B
So
the
general
idea
is
this:
I
mentioned
earlier
that
there
are
a
lot
of
mops
that
have
been
synthesized
so
for
every
synthesis.
Successful
synthesis,
there's
a
sort
of
a
corresponding
paper
so
that,
if
we
jot
through
the
paper,
there's
a
synthesis
section
and
it
talks
about
the
moth
name.
What
kind
of
metal
precursors
are
used?
What
kind
of
organic
precursors
are
used
with
solvents?
B
B
B
So
the
first
part
of
this
was
the
text
mining,
and
this
is
very
dirty
work
in
certain
sense,
and
I
think
few
people
are
sort
of
starting
to
do.
Text
mining
and
the
workflow
goes
something
like
this.
We
download
a
lot
of
papers
with
the
permission
from
these
sort
of
journals
and
publishers
and
then
based
on
the
format
we
sort
of
try
to
identify
the
synthesis
paragraph
and
for
moth
papers.
B
And
then
we
had
a
test
set.
We
developed
a
text
mining
code
to
get
all
this
information
regarding
the
metal,
precursor,
organic
precursor
conditions
and
so
forth,
and
we
checked
the
f1
score
on
a
test
set
of
100,
mods
and
they're
pretty
well
point
in
this
text:
mining
materials
world,
I
think
0.7.8-
is
pretty
decent
performance.
We
can't
get
all
of
them.
Unfortunately,
so
we
got
this
information
and
then
we
looked
into
statistics
on
what
kind
of
metal
precursors
are
often
used.
What
kind
of
organic
precursors
are
used?
B
What
kind
of
solvents
are
used
for
these?
All
these
are
structures,
and
we
checked
with
the
experimentalists
on
the
sanity
of
this
text,
mind
data
and
they
were
sufficient
enough
such
that
we
can
use
this
to
feed
the
machine
learning
algorithm
and
with
the
machine
learning
algorithm.
With
regards
to
synthesis.
There's
always
this
sort
of
issue
where
only
the
successful
synthesis
is
reported
and
the
unsuccessful
synthesis
is
not
reported.
B
So
in
this
case
the
input
is
the
metal,
precursor
type,
organic
precursor
type,
composition,
temperature,
time
and
solvents.
So
we
have
this
information
for
the
positive
data
for
the
in
silico
data.
We
just
randomly
choose
the
combinations
of
metals,
composition,
temperature
and
so
forth.
So
that's
our
unlabeled
data
and
the
idea
is
that
the
pu
learning
algorithm
can
distinguish
between
the
positive
and
the
unlabeled
such
that
it
can
give
a
crystalline
score
which
is
sort
of
equivalent
to
whether
or
not
it
can
be
synthesized
or
not.
B
And
we
saw
that
with
the
positive
data
which
is
green.
It
gives
high
scores
and
because
we
didn't
have
negative
data,
what
we
did
was
instead
of
using
negative
data.
We
had
something
called
the
amorphous
data
in
the
sense
that
certain
number
of
these
mosques
were
reported
as
being
amorphous
in
the
literature,
so
we're
thinking
that
these
are
sort
of
akin
to
not
being
successful.
B
So
we
took
this
amorphous
data,
synthesis
conditions
for
selected
moss
and
we
saw
that
the
pu
algorithm
it
sort
of
reports,
a
very
low
negative
score
in
the
red
for
these
mods.
So
we're
thinking
that
this
is
sort
of
one
way
where
we
can
distinguish
between
this
type
of
a
model
can
be
used
to
sort
of
input,
a
particular
synthesis
condition,
and
it
can
give
a
crystalline
score
based
on
whether
or
not
it's
high
or
low
and
the
ones
are
low.
B
So
these
are
some
of
the
works,
that's
being
done
in
our
lab,
we're
doing
natural,
sorry,
sorry,
natural
sort
of
design
which
is
cumbersome
but
kind
of
fun,
and
then
we're
doing
a
lot
of
artificial
design
where
we
look
into
different
input,
representations
with
user
desired
properties
and,
finally,
we're
doing
a
lot
of
text.
Mining
and
sort
of
I
forgot
to
include
it
here,
but
we
do
a
lot
of
image,
mining
and
figure
mining.
B
B
So
we
try
to
sort
of
we're
trying
to
gain
as
much
information
from
the
previously
published
papers
to
see
whether
or
not
we
can
use
them
to
get
synthesis
conditions
and
ultimately
sort
of
in
four
or
five
years.
What
we
want
to
do
is
we
want
to
use
these
for
text
and
image
data,
to
write
new
papers
and
to
sort
of
write
new
projects
and
to
combine
different
concepts
such
that
even
the
research
idea
generation
and
the
paper
writing
process
can
be
automated
using
aon.
Hopefully,
so.
Thank
you
for
your
time.