►
Description
Ralph Kube from PPPL talks about machine learning in fusion research
A
Okay,
beautiful
so
my
name
is
ralph
cougar,
I'm
a
research
physicist
at
the
princeton
plasma
physics
laboratory
and
I
will
talk
a
little
bit
about
machine
learning
in
fusion
energy
sciences.
A
So
when
I
was
asked
to
do
this
presentation,
I
really
had
to
go
back
and
think
what
I
was
going
to
present,
because
I
certainly
am
a
machine
learning
practitioner
in
fusion
energy
sciences,
but
it
is
quite
a
large
field.
A
So
to
illustrate
the
the
breadth
of
the
field,
I
took
all
the
abstracts
from
that
meeting
and
generated
a
word
cloud
and
here
stripping
away
all
filler
words
like
and
with
etc.
The
how
large
a
word
appears
in
this
plot
gives
you
some
sense
of
the
relative
of
the
frequency.
These
words
appear
so
first,
let
me
say
that
fusion
energy
sciences
can
be
roughly
separated
into
magnetic
confinement
and
inertial
confinement.
Fusion
and
this
mini
conference
covers
both,
so
this
workout
gives
you
oh.
A
Certainly
neural
network
appears
very
large
that
reflects
the
fact
that
most
machine
learning
practitioners
in
fusion
energy
sciences
do
deep
learning
in
all
kinds
of
varieties
that
can
be
multi-layer.
Perceptrons.
You
see
the
phrase
convolutional
neural
network
popping
up.
You
see
deep
neural
network.
There
are
some
people
who
do
time
series
modeling
with
a
recurrent
neural
network,
neural
networks
and
then
more
of
coming
from
the
from
the
physics
side.
You
see
the
big
phrase
simulation
and
that
reflects
the
the
research
thrust
in
the
field
to
replace
expensive
numerical
simulations
with
machine
learning
surrogate
models.
A
That's
certainly
an
aspect,
many
people
look
into,
and
then
you
see
many
smaller.
You
see
many
smaller
terms.
For
example,
anomalous
transport
coefficients
beam
emission
spectroscopy
disruption
prediction,
inner
shell
target
and
those
are
all
terms
that
reflect
the
different
physics
modalities
so
to
say
that
come
that
that
are
important
from
the
from
the
exp
from
the
experiments,
and
you
also
see
some
smaller
words,
for
example,
gaussian
process
regression.
A
A
Okay,
I'm
personally
working
on
magnetic
confinement
fusion,
so
I
thought
I
it
was
coming
from
the
experimental
side.
I'm
going.
I
thought
I
motivate
a
little
bit
what
kind
of
data
we
are
working
with,
because
when
you
say
machine
learning,
you
certainly
look
at
the
data
sources.
A
So
here
on
the
top
left,
you
see
a
token
mac,
a
machine
for
use
for
magnetic
confinement,
fusion-
and
here
this
is
an
illustration
of
the
upcoming
spark
token
mac,
one
of
three
token
mags
that
is
currently
under
construction
in
the
us
and
then
in
the
top
right.
I
just
took
some
random
visualizations
of
data
that
are
measured
in
plasma
discharges
and,
as
you
can
see,
all
these
five
plots
look
nothing
alike,
and
that
is
because
the
data
we,
the
measurement
data
we
sample
infusion
plasma,
varies
by
a
lot
it
can.
A
There
are
different
diagnostics,
which
are
all
sensitive
to
different
physics
and,
for
example,
we
have
magnetic
fluctuations
here
shown
a
spectrogram.
We
have.
A
There
are
diagnostics
and
sort
that
are
sensitive
to
electromagnetic
radiation
in
all
kinds
of
the
spectrum,
for
example,
visible
light
infrared
which
targets
the
material
walls
cyclotron
radiation,
which
is
kind
of
micrometer
range
and
they're.
Also,
for
example,
diagnostics
that
are
based
on
particle
flux,
so
yeah,
a
broad
array
of
diagnostics,
is
used
to
sample
the
various
aspects
of
the
plasma
and
we
use
those
to
try
and
reconstruct
the
plasma
state.
Now
these
diagnostics
also
have
different
output.
Some
diagnostics
just
give
a
zero
zero
dimensional
quantity,
for
example,
a
pressure
sensor.
A
Some
diagnostics
give
us
one
dimensional
data
which
would
be
of
profile
or
two
dimensional
data
and
in
addition,
many
of
these
diagnostics
either
due
to
the
physics
they
are
sensitive
to
or
electronic
limitations
will
sample
on
different
time
scales.
A
So
on
the
plot
on
the
lower
right
here,
this
kind
of
illustrates
what
kind
of
length
and
time
scales
we
are
dealing
with
in
fusion
plus
mass
time
scales
range
from
less
than
nanoseconds
for
wave
physics
to
seconds
for
microscopic
transport
and
the
length
scales.
A
So
that's
that's
over
10
orders
of
magnitude
on
the
x
axis
and
then
on
the
y
axis.
We
have
time
scale
differences
between
seconds
and
10
to
the
minus
seconds.
So
it's
a
lot
of
variability
and
from
this
variability
comes
also
the
fact
that
machine
learning
tasks
vary
a
lot
in
fusion
plasmas.
A
So
let
me
just
quickly
illustrate
three
kind
of
machine
learning,
research
topics
that
are
popular
at
pppl
and
since
we
are
associated
with
princeton,
also
at
princeton
university
on
the
left,
you
see
an
illustration
that's
supposed
to
be
plasma
control
and
here-
and
that
is
exactly
what
the
name
says.
If
we
have
a
discharge,
we
have
a
real-time
feedback
system
that
tries
to
steer
the
plasma
away
from
unstable
configurations,
kind
of
kind
of
just
keep
it
confined,
and
there
are
some
there
are
some.
A
There
are
some
research
going
on
how
machine
learning
can
be
used
to
optimize
the
feedback
system
so
that
the
plasma
plasma
control
system
is
able
to
operate
more
efficiently?
Keep
the
plasma
better
on
time.
A
And
the
we
are
running
some
simulations
on
yeah,
the
the
big
computers
like
corey
and.
B
A
Daddy
has
to
talk
to
us
guys,
I'm
sorry,
I'm
having
a
daycare
situation
today.
Okay,
right,
so
we
we,
we
have
the
xgc
code,
which
is
very
compute
intensive
and
in
preparation
for
running
more
physically
accurate
models.
We
would
like
to
replace
some
compute
intensive
parts
of
the
codes
with
machine
learn
with
machine
learning
surrogate
models.
A
A
A
So
when
it
comes
to
using
nurse
infrastructures,
nurse
provides
a
great
infrastructure
and
many
of
the
machine
learning
practitioners.
I've
talked
to
use
no
use
the
infrastructure
and
are
very
happy
with
that.
A
Some
things
we
are
very
happy
with
is
that
tensorflow
and
pi
charge
modules
are
readily
available,
also
in
modern
versions
and
a
typical
job
for
machine
learning.
Practitioners
are
smallish
problems;
they
don't
require
much
compute
time.
So
for
this
we
rely
heavily
on
jupiter
notebook
and
we
really
appreciate
the
gpu
support
we
have
in
there.
A
Five
minutes
to
go
five
minutes:
okay
for
logging,
people
use
tensorboard
and
for
larger
jobs.
I
know
people
use
ray
tune,
horowat,
cray,
hpo
and
other
tools
that
are
available
at
a
nurse.
A
Some
bottlenecks
I
personally
have
experienced
were
that
when
we
train
large
training
jobs
which
require
multiple
terabytes
of
data,
data,
loading
becomes
a
bottleneck
and
we
have
inspected
this
with
debuggers
such
as
tau
and
then
so.
We,
we
are
actually
very
happy
with
the
software
that
is
provided.
So
I'm
actually
going
to
start
wrapping
up
my
talk
now
and
talk
about
some
new
contenders
and
there
are
the
ideas
that
that
is
basically
using
automatic
differentiation
in
combination
with
scientific
simulation.
A
Some
that
paradigm
is
also
known
under
programming
2.0
and
there.
The
idea
is
basically
to
use
automatic
differentiation
to
make
arbitrary
code
amenable
to
gradient-based
optimization,
not
only
neural
networks,
for
example.
A
So
there
are
two
tools
that
I
gather
are
popular
among:
machine
learning,
practitioners
and
those
are
julia
and
jax.
So
julia
is
a
language
that
is
developed
from
the
scratch.
It's
very
young
automatic
differentiation
as
a
first
class
citizen,
and
the
entire
language
is
just
in
time
compiled,
so
it
runs
very
fast
and
jax
is
used
in
conjunction
with
python.
A
It
can
differentiate
native
python
and
numpy,
and
it's
also
just
in
time
compiled,
but
right
now
those
languages
are
only
used
for
a
very
at
a
very
experimental
stage
and
not
at
a
production
stage,
and
it
is
still
unclear
how
fusion
energy
sizes
will
incorporate
this
automatic
differentiation
with
traditional
hpc
simulations
and
then
another
trend
I
would
just
like
to
shortly
bring
up
here
is
that
we
fusion
will
most
likely
with
the
next
generation
of
fusion
experiments,
generate
very
large
data
sets.
A
We
are
talking
about
crossing
petabytes
of
data
per
day
and
it
is,
and
we
are
looking
into
how
machine
learning
can
be
used
with
the
with
these
kinds
of
data
sets.
So
some
technologies
we
are
discuss.
I
have
discussed
internally
with
colleagues,
are
first
custom
machine
learning
hardware,
and
this
goes
under
the
name,
wafer
scale
engines.
Sierra
brass
recently
released
one
of
these.
It's
basically
chips
designed
for
machine
learning,
tasks
specialized
hardware
and
combine
them
into.
B
A
That
are
like
physically
large,
hence
the
name
way
for
scale
engines.
Tesla
also
talked
about
this
when
they
talked
about
their
dojo
cluster
and
it's
it's
basically
tensor
processing
units.
A
What
what
they're
doing
from
which
google
released,
I
think
in
2017,
but
those
I
believe
will
be
commercially
available,
also
in
conjunction
with
the
big
data
age
of
fusion,
if
you
so
will,
is
a
move
towards
transformer
neural
networks
and
those
those
are
by
design
more
general,
but
require
more
data
to
to
to
work
on
the
tasks.
So,
there's
some
research
going
on
how
to
use
these
networks
for
fusion
energy
right
now,
yeah.
A
Great
timing,
in
that
case,
we
have
one
minute
left.
If
anybody
has
a
question
to
ask.
B
I
guess
I
can
ask
a
question
so
thanks
ralph
for
the
nice
shout
out
of
the
machine
learning
software,
but
is,
is
there
anything
you
feel
is
missing
from
the
stack
or.
A
Actually,
no-
and
I
I
have
I
to
the
colleagues
I
talked
in
preparation
for
this
talk,
I
really
haven't
heard
anything
the
anything
that
they
feel
is
missing
or
would
really
increase
their
productivity
when
they
do
machine
learning
at
nurse.
So
I
get
the
impression
that
the
facilities
are
quite
yeah,
where
they
need
to
be
right
now.
Well,
we'll
see
later
when,
when
we
really
do
machine
learning
with
the
with
these
big
data
that
are
supposed
to
be
generated
from
ether.
B
Okay,
yeah
feel
free
to
immediately
reach
out.
If
there
is-
and
I
don't
know
if
you've
had
a
chance
to
try
permanently
yet
but.