►
From YouTube: 2020-09-04 - Auralee Edelen - Steps toward holistic control of particle accelerators with NNs
Description
NERSC Data Seminars Series: https://github.com/NERSC/data-seminars
Title: Steps toward holistic control of particle accelerators with neural networks
Abstract: Particle accelerators are used in a wide array of medical, industrial, and scientific applications, ranging from cancer treatment to understanding fundamental laws of physics. While each of these applications brings with them different operational requirements, a common challenge concerns how to optimally adjust controllable settings of the accelerator to obtain the desired beam characteristics. For example, at highly flexible user facilities like the LCLS and FACET-II, requests for a wide array custom beam configurations must be met in a limited window of time to ensure the success of each experiment — a task which can be difficult both in terms of tuning time and the final achievable solution quality, especially for novel or non-standard setups. At present, the operation of most accelerator facilities relies heavily on manual tuning by highly-skilled human operators, sometimes with the aid of simplified physics models and local optimization algorithms. As a complement to these existing tools, approaches based on machine learning are poised to enhance our ability to achieve higher-quality beams, fulfill requests for custom beam parameters more quickly, and aid the development of novel operating schemes. Focusing on neural network based approaches, I will discuss proof-of-principle studies that point toward the potential of machine learning in this regard, highlight open questions and challenges, and give an outlook on some of the future pathways toward bringing these techniques more fully into operation of accelerators. These improvements could increase the scientific output of user facilities and enable new capabilities by tuning a wider range of machine settings, as well as exploiting subtle sensitivities that may otherwise go unutilized. They could also help us to meet the modeling and tuning challenges that become more acute as we push toward the more difficult-to-achieve beam parameters that are desired for future accelerator applications (e.g. higher beam energies and intensities, higher stability, and extreme adjustments of the beam shape in phase space).
A
All
right,
I
think
we
can
get
started
now.
So,
first
of
all
everyone
for
those
of
you
who
are
unfamiliar
with
our
data
seminar
series
these
are
recorded.
So
I
just
to
remind
everyone
of
that
that
this
is
this
talk
and
the
subsequent
conversation
is
going
to
be
recorded
and
posted
on
our
data
seminar
channel.
A
So
it's
my
great
pleasure
to
welcome
our
speaker
this
week,
aurelie
edelen.
I
first
learned
of
orilli's
work
at
the
monterey
data
conference
in
august
last
year
and
ever
since
then,
we've
been
trying
to
find
a
time
that'll
work
for
her
to
come
and
tell
us
more
about
it.
So
I'm
delighted
that
we've
been
able
to
arrange
her
to
come
and
tell
us
more
about
her
work
with
particle
accelerators
and
the
control
particle
accelerated
the
neural
networks.
A
So
orley
is
a
novsky
fellow
at
slack
national
accelerator
lab.
She
did
her
graduate
work
at
fermi,
national
accelerator
lab,
applying
neural
network-based,
approaches
to
particle
accelerators
and
she's
continued
and
extended
that
work
at
slac
and
developing
machine
learning,
based
approaches
for
the
modeling
and
control
of
particle
accelerators.
So
we're
very
happy
to
have
her
with
us
and
looking
forward
to
hearing
more
about
more
about
your
work
holy.
Please
take
it
away.
B
B
So
there's
a
lot
of
work
that
I
I
won't
cover
here,
but
just
know
that
there's
there's
a
wide
variety
of
work
going
on,
so
for
those
who
aren't
familiar
with
particle
accelerators,
there's
actually
quoted
diverse
different
types
of
accelerators
for
different
use
cases,
and
each
of
these
come
with
fairly
unique
control
challenges
as
well.
So
we
have
some
machines
that
are
quite
flexible.
B
These
are
kind
of
scattered
all
over
the
place
at
universities
in
various
national
labs,
where
the
goal
is
to
really
test
out
new
accelerator
concepts
and
do
things
like
advanced
manipulation,
techniques
for
the
beam,
and
then
we
also
have
some
work.
That's
focused
on
sort
of
fundamentally
more
challenging
to
control
accelerator
technology,
so
this
includes
things
like
trying
to
use
plasmas
to
accelerate
accelerate
beams
and
then
finally,
there
are
a
variety
of
industrial
and
medical
use
cases.
B
In
contrast
to
something
like
x-rays,
which
tend
to
scatter
scatter,
the
deposited
energy
around,
it's
a
little
bit
harder
to
really
accurately
target
a
tumor.
So
in
that
case,
if
you
can
then
get
really
fine
control
over
the
beam
position
and
energy,
you
could
potentially
do
a
better
job
of
cancer
therapy.
B
We're
really
trying
to
support
a
huge
variety
of
experiments,
ranging
from
trying
to
understand
matter
and
extreme
conditions
to
trying
to
understand
natural
processes
like
photosynthesis,
and
we
support
a
lot
of
different
user
experiments.
So
in
2016
it
was
around
a
thousand
and
then
typically,
every
single
experiment
has
a
different
need.
As
far
as
the
x-ray
wavelength,
the
beam
pulse,
duration,
etc,
and
the
way
that
we
achieve
those
target
beam
characteristics
for
the
users
is
by
adjusting
a
variety
of
machine
settings.
B
So
I'm
only
showing
a
very
small
subset
here,
but
ultimately
it
ends
up
being
dozens
to
hundreds
of
parameters
that
can
potentially
be
tuned
and
typically
we're
trying
to
produce
quite
different
electron
beam
shapes
in
phase
space.
So
I'm
showing
three
examples
of
different
time.
Energy
correlations
that
you
might
get
on
the
beam
at
lcls
and
to
do
this
we
actually
spend
quite
a
lot
of
time
adjusting
the
accelerator
settings
to
to
facilitate
the
needs
for
these
experiments
and
to
put
some
numbers
on
this.
B
If
you
try
to
attach
some
value
to
how
much
the
number
of
hours
that
we
spend
tuning
is
you
can
see
that
if
we
are
running
about
5000
hours
of
experiment
delivery
per
year
and
we
have
an
annual
budget
of
145
million
dollars,
this
means
it's
about
30
thousand
dollars
per
experiment
hour
to
run
the
machine,
and
so
the
400
hours
that
we
spent
hand
tuning
has
about
a
12
million
dollar.
Equivalent
value
is
one
way
of
looking
at
it.
B
B
B
B
Presumably
we
could
do
offline
optimization
of
the
model
and
then
put
those
settings
onto
the
machine,
and
we
would
get
the
beam
that
we
wanted,
but
unfortunately
nothing's
ever
that
simple.
We
have
lots
of
sources
of
fluctuations
on
the
machine,
so,
for
example,
the
initial
laser
that
produces
the
electron
beam
can
fluctuate.
B
B
In
effect,
these
heuristic
control
policies
that
they
then
use
to
help
tune,
but
this
is
an
exceptionally
hard
problem
overall,
so
you're,
typically
only
focusing
on
one
small
part
of
the
accelerator
at
a
time.
For
example,
we
tend
to
tune
the
injector
upstream
part
of
the
accelerator
separately
from
the
main
linac
and
separately
from
at
lcls,
the
the
undulator
that
produces
the
photon
beam
and
we
kind
of
know
from
experience
that
this
is
actually
a
sub-optimal
way
to
tune.
B
Then
you
would
get
if
you
first
tune
one
section
in
the
machine
and
then
use
that
optimal
beam
as
input
to
the
next
section
of
the
machine
shown
in
green
here.
So
this
hints
that
you
know
we
might
be
able
to
actually
get
fundamentally
higher
beam
quality
if
we're
tuning
on
parameters
across
the
entire
accelerator.
B
B
Okay,
so
back
to
the
the
overall
tuning
problem,
there's
sort
of
a
range
of
ways
that
one
can
start
to
to
think
about
controlling
a
large
machine
like
this.
So
so
one
is
to
simply
assume
that
you,
you
aren't
going
to
be
able
to
learn
anything
systematic
about
the
machine,
in
which
case
you
can
do
some
model
free,
optimization
and
hopefully
find
a
method
to
do
this
for
many
parameters
simultaneously.
B
So
that's
on
the
one
extreme
on
the
other
extreme,
we
do
have
good
accelerator
simulations
and
with
some
work
these
can
be
made
to
match
the
accelerator
well
in
some
cases,
but
generally,
this
is
extremely
computationally
intensive
to
do
so.
For
example,
this
is
just
one
run
at
lcls
and
it
took
tens
of
hours
on
thousands
of
cores
at
nurse,
and
this
is
this.
B
There
are
some
approaches
that
still
use
the
physics
simulation
to
try
to
get
get
towards
faster
execution
speeds.
So,
for
example,
there
are
folks
who
have
taken
particle
accelerator
simulations
and
made
them
into
gpu
accelerated
versions
menus
to
those
online,
but
in
general
the
execution
speed
is
still
somewhat
limited.
You're
talking
about
seconds
to
minutes.
B
So
this
is
an
example
where
we're
predicting
the
longitudinal
phase
space
measured
at
this
transverse
deflecting
cavity,
while
changing
a
really
wide
variety
of
upstream
machine
settings
over
basically
their
entire
operating
range.
So
this
this
is
still
somewhat
limited.
To
do
to
do
this
for
the
full
simulation
you
would
need
to
include
many
more
parameters,
but
this
is
a
nice
step,
and
what
this
really
means
is
that
you
know.
B
Instead
of
having
to
run
on
nurse,
we
can
now
run
on
a
laptop
or
put
this
onto
the
control
system,
computers
and
run
it
there,
which
we're
also
starting
to
do
so.
We're
incorporating
some
of
these
models
into
gui's
that
can
actually
talk
to
the
overarching
control
system
and
read
back
parameters
from
the
control
system
and
give
an
operator
a
you
know,
an
interface
with
which
to
play
around
with
different
potential
settings
so
that
that
sort
of
addresses
this
question
of.
B
Can
we
speed
up
some
some
of
our
models
and
put
them
on
the
machine,
but
it
doesn't
answer
the
question
of.
Is
it
a
good
idea
to
use
those
models
in
optimization
and
fundamentally,
the
problem
here
is
that
as
soon
as
you're
optimizing,
your
movie
you're,
potentially
moving
outside
of
the
training
distribution
that
you
had
for
that
model.
B
So
we
wanted
to
see
how
reliable
we,
how
reliable
our
models
were
in
practice
for
for
using
with
optimization.
So
we
tried
this
on
an
injector
setup,
so
this
is
a
pretty
non-linear
part
of
the
accelerator.
It's
at
low
energy,
so
the
beam
cell
fields
are
pretty
important
and
we
ran
just
sort
of
a
genetic
algorithm
on
the
the
the
accelerator
physics
simulation,
and
this
is
commonly
done
for
for
like
design
and
for
finding
good
machine
working
points.
B
So
it
was
a
natural,
optimization
algorithm
to
pick
to
compare
against,
and
then
we
trained
a
machine
learning
model
based
on
a
random
sample
of
the
input
variables
and
ran
a
genetic
algorithm
on
that
model
as
well,
and
then
compared
the
resulting
pareto
front
trade-offs,
which
show
just
the
optimal
trade-off
between
these
two
parameters.
And
then
we
took
the
prio
front
that
we
obtained
from
the
neural
net
model
and
put
it
back
through
the
physics
simulation
to
see
how
good
the
prediction
was.
B
B
In
addition
to
being
a
way
to
just
encapsulate
a
fast
executing
physics
simulation,
I'm
going
to
skip
this
just
because
I
think
I
want
to
make
good
use
of
time,
but
we
also
have
tried
this
on
on
a
different,
a
different
system
as
well
and
there's
a
paper
that
I've
pointed
to
that
you
can
check
out
if
you're
interested.
So
we
had
this
question,
can
we
can
we
do
the
same
sort
of
thing
on
a
live
accelerator?
B
So
we
went
down
to
ucla
to
their
pegasus
beam
line
and
we
tried
to
do
this
kind
of
approach
just
with
measured
data
from
the
machine,
and
we
decided
to
tackle
this
round
of
flat
beam
transform
problems.
So
the
goal
is
to
take
a
round
beam
coming
off
of
the
cathode
and
to
turn
it
into
a
flat
beam,
and
this
is
actually
quite
a
challenging
problem
because
of
all
of
the
non-linearities
in
the
system.
B
When
you
do
this
procedure-
and
we
actually
did
this
for
for
a
day
after
the
training
training
data,
when
there
was
a
little
bit
of
distribution
drift
in
some
of
the
inputs.
And
so
the
real
punch
line
here
is
that,
instead
of
relying
on
hand
tuning
which
takes
around
10
to
20
minutes,
the
neural
net
very
quickly
in
less
than
a
minute,
is
able
to
give
some
suggested
settings
given
drifting
upstream
settings
upstream
inputs.
Rather
that
that
get
you
close
to
this
flat
beam
solution.
B
B
So
then,
even
though
the
solution
isn't
perfect,
you
can
still
fine-tune
with
other
optimizers.
So
we
we
started
with
the
the
guests
from
our
neural
net
model,
and
then
we
used
a
local
tuning
algorithm.
A
and
process
based,
bayesian,
optimization,
algorithm
and
then
hand
tuning,
and
this
sped
up,
starting
with
the
neural
net
guest
sped
up
the
convergence
for
all
of
these
other
methods.
B
So,
where
we're
headed
now
is
actually
sort
of
trying
to
combine
these
two
things.
So
this
I'm
showing
a
slide
from
older
work
where
we
tried
to
take
some
simulation
data,
train
and
neural
net
and
then
just
update
that
with
kind
of
a
smaller
amount
of
measured
data,
and
we
were
able
to
get
pretty
good
agreement
between
measurements
do
doing
it
this
way,
and
the
appeal
here
is
really
just
that
you
would
be
able
to
get
away
with
less
measured
data
potentially
and
also
you
would
still
have
a
model.
B
B
The
orbit
feedback
settings
to
compensate
for
this.
So
what
they
did
was
train
neural
net
over
a
few
years
of
archive
data
and
then
looked
at
the
sensitivities
to
various
inputs,
and
they
found
that
there
was
actually
a
really
strong
correlation
with
a
ground
temperature
sensor
in
the
parking
lot
outside
of
spear
3..
B
And
in
retrospect
this
kind
of
makes
sense,
because
the
the
actual
temperature
control
inside
the
accelerator
enclosure
isn't
actually
that
good
and
so
you'll
get
expansion
and
contraction
of
the
entire
machine,
which
then
you
know
would
lead
you
to
need
to
readjust
the
steering
set
points
to
to
maintain
good
performance,
so
that
the
idea
here
now
is
that
this
could
potentially
be
a
way
to
do.
An
initial.
B
So
so
far,
everything
that
I
talked
about
involved
forward,
modeling,
where
we
have
some
settings
and
possibly
some
input-
diagnostics
like
a
some
information
about
the
laser
upstream,
predicting
some
output
diagnostics
that
are
measurements
of
the
beam
that
we
care
about
and
then
using
that
you
know
with
an
optimizer
to
to
try
to
get
get
a
good
set
of
settings.
But
in
fact
you
can
kind
of
skip
these
steps
a
little
bit
if
you
instead
try
to
develop
an
inverse
model.
B
So
the
idea
here
is
that
if
you
have
some
desired
output
being
characteristics-
and
you
have
some
other
things
that
indicate
what
the
machine's
state
is,
otherwise
that
you
don't
want
to
control.
So
this
could
be
settings
that
you
don't
want
to
touch
or
input
diagnostics
that
you
have
no
control
over.
You
can
then
map
those
to
predicted
optimal
settings
that
should
get
you
close
to
your
desired
beam
output,
and
this
is
pretty
appealing
for
accelerators.
B
For
a
lot
of
reasons
I
mean
one
is
just
the
high
dimensionality,
but
another
is
that
we
also
have
a
lot
of
these
image,
based
where
it
would
be
really
nice
to
use
something
like
a
neural
net,
to
look
directly
at
this
image.
And
then
you
know
either
use
that
in
use
that
to
give
an
initial
setting
guess
to
to
get
to
that
target,
face
space
or
to
use
it
actually,
just
in
a
control
policy
which
I'll
talk
about
later.
B
But
just
in
this
first
case
we
actually
tested
this
out
a
couple
of
years
ago
in
in
kind
of
a
preliminary
study
at
lcls.
So
there
we
have
a
specific
target,
longitudinal
phase
space
that
that
we
want
to
achieve
at
the
transverse
deflecting
cavity,
and
so
what
we
did
was
just
took
some
some
data
from
the
machine
and
then
trained
an
inverse
model
to
give
the
suggested
settings
for
two
settings
on
the
machine
that
speak
very
strongly
to
the
longitudinal
phase
base.
B
And
then
we
used
a
local
optimizer
to
kind
of
fine-tune
that
solution,
afterward
and
so
the
the
result
essentially
was
that
we
messed
up
the
beam
phase
space.
We
gave
it
this
target
and
when
we
tried
to
use
just
the
local
optimizer,
it
was
not
able
to
converge
in
the
end,
but
when
we
gave
it
the
initial
guess
from
the
neural
network,
it
was
able
to
converge
pretty
closely
to
the
target
longitudinal
phase
space.
B
So
this
is
kind
of
a
nice
example
of
combining
sort
of
the
two
opposite
ends
of
this
spectrum
of
ways
of
tackling
the
problem
that
I
suggested
earlier.
We're
combining
global
modeling
with
model
free,
optimization,
moving
more
towards
this,
this
middle
middle
approach,
where
you're
kind
of
trying
to
learn
a
model
as
you
go
or
learn
a
control
policy.
As
you
go
I'll
I'll
speak
more
about
this,
and
I
think
this
is
a
major
area
that
we
want
to
keep
pushing
on
in
the
future.
B
So
a
method
that
we've
used
pretty
successfully
in
accelerators
now
at
a
couple
of
places,
is
bayesian,
optimization
and
in
particular,
bayesian
optimization
with
gaussian
processes
oops.
So
the
the
idea
here
is
that
you
learn
a
model
that
has
some
associated
uncertainty,
estimate
with
it,
and
then
you
use
that
uncertainty
along
with
the
model
prediction
to
help.
You
select
the
next
point
to
sample
and
this
way
you're
iteratively
updating
the
model
as
you
search
and
so
at
lcls.
B
This
has
been
demonstrated
on
the
problem
of
trying
to
adjust,
focusing
magnus
to
maximize
x-ray
pulse
energy.
This
was
an
effort
led
by
joe
durris,
and
the
the
main
result
is
that
when
we
we
apply
gp
based
bayesian
optimization,
we
get
substantially
faster
and
higher
quality
x-ray
pulse
energy
output
than
when
we
use
a
standard
optimizer
like
nelder,
mead,
simplex
and
then
further.
B
There
are
actually
kind
of
expected
correlations
between
adjacent
input
parameters
and
the
x-ray
pulse
energy,
and
one
nice
thing
about
gas
and
process
models
is
that
you
can
actually
kind
of
bake.
These
expected
correlations
into
the
kernel,
and
when
you
do
this
for
the
same
data
points,
you
actually
learn
a
much
better
representation
of
the
underlying
ground
truth,
and
so
this
was
done
for
for
lcls
on
this
problem,
and
indeed
you
get.
You
get
substantially
faster
convergence.
This
way
and
in
both
cases
it's
still
faster
than
the
standard
optimizer.
B
You
can
also
start
to
think
about
taking
a
similar
kind
of
approach,
but
with
a
neural
network.
So
if
you
have
a
neural
network
control
policy,
basically
what
you're
doing
is
taking
some
measurement
of
the
system
state
and
then
making
sudden
changes
on
the
machine
and
then
you'll
get
some
feedback,
and
this
is
a
common
paradigm
that
that
this
is
applicable
to
is
reinforcement,
learning.
B
So
I'm
actually
in
this
video
just
showing
seven
iterations
of
the
algorithm
and
it
converged
to
a
similar
solution
that
the
gaussian
process
optimizer
did
in
this
case.
This
again
is
a
low
dimensional
problem,
so
this
was
kind
of
a
preliminary
test,
but
there's
a
lot
of
a
lot
of
work
going
on
throughout
the
accelerator
community
to
try
to
apply
reinforcement,
learning
with
neural
nets,
to
a
variety
of
different
kinds
of
accelerator
systems.
It's
something
that
we're
interested
in
pursuing
more
lcls.
B
B
It
might
update
significantly
more
slowly
than
that
and
then
at
smaller
accelerators
there's.
Sometimes
this
issue,
where
you
you
actually
might
only
have
one
one
of
these
measurement
devices
that
you
actually
have
to
physically
move
to
different
parts
of
the
beam
line,
in
which
case
you've
lost
your
ability
to
measure
at
both
at
at
both
places.
B
This
is
showing
in
a
beam
line
at
fermilab,
where
you
actually
have
to
physically
insert
a
mask
into
the
beam
line,
and
then
you,
you
can
measure
the
transverse
phase
space
by
seeing
how
the
beam
spreads
out
and
doing
some
fits
to
that
as
measured
on
a
on
a
screen.
That's
put
downstream
in
the
beam.
So
if
everything
upstream
is
sufficiently
stable
enough
and
you've
trained
a
model
on
a
wide
range
of
settings
here,
you
could
potentially
just
use
the
machine
learning
model
to
predict.
B
B
A
related
approach
that
was
demonstrated
at
facet
a
few
years
ago
was
to
actually
take
a
very
simple
simulation
that
we
don't
expect
to
match
the
machine
very
well
and
then
adaptively
tune
that
with
based
on
measurements
from
a
diagnostic,
that's
always
available.
So
in
this
case
it
was
the
seag
diagnostic
that
was
then
used
to
help
tune.
This
model
online
to
then
predict
the
the
longitudinal
face
space
directly.
B
So
this
also
has
been
demonstrated
experimentally
at
this
time.
We
had
done
it
on
lcls,
but
the
plan
is
to
keep
working
on
this
and
get
it
fully
implemented
for
for
fast
it
too.
So
in
each
of
these
cases
these
are
not
really
used
in
regular
operation.
Yet
there's
still
some
development,
that's
needed
to
to
make
sure
that
they're
robust
enough
and
I'll
talk
a
little
bit
more
about
about
that.
B
In
a
few
minutes,
another
piece
of
the
puzzle
here
is
that
we
we
also
want
to
be
able
to
actually
extract
useful
information
from
these
really
complicated,
high
dimensional
signals,
and
this
is
an
area
where
there's
been
some
work
on
trying
to
use
machine
learning
in
various
ways
to
do
this.
So
I'll
just
go
through
one
example.
B
So
the
the
present
method
relies
on
a
number
of
assumptions,
including
that
you're
only
looking
at
vertical
slices
at
once,
which
starts
to
break
down
as
you
get
into
the
so-called
saturation
regime,
where
some
of
these
spikes
start
to
to
have
an
angle
to
them,
and
what
a
grad
student
that
we
were
working
with
found
was
that
you
actually
can
predict
the
x-ray
power
profile.
Quite
well,
if
you
use
a
neural
net
instead
of
using
the
iterative
reconstruction
method
and
it's
much
faster
as
well.
B
B
Optimization
we're
trying
to
get
faster,
more
accurate
machine
models
that
we
could
potentially
use
online
to
aid
in
aid
tuning
and
then
we're
also
trying
to
extract
more
information
from
the
accelerator,
whether
it's
predicting
an
output
diagnostic
or
trying
to
improve
kind
of
on-the-fly
analysis
of
these
diagnostics
and
then,
lastly,
we're
trying
to
use
machine
learning
to
try
to
facilitate
some
improved
understanding
of
the
machines.
So
this
includes
things
like
trying
to
pull
out
sensitivities
that
we
might
then
be
able
to
exploit
and
control
that
could
otherwise
go
unnoticed.
B
B
You
can
ensemble
sets
of
neural
nets
together,
we're
looking
into
beijing
neural
nets
at
the
moment
and
trying
to
see,
if
that,
that's
a
potential
route
for
us.
There's
also
related
to
that
this.
This
question
of
how
do
you
retrain
these
models
online
in
a
reliable
fashion,
so
for
machines
where,
where
things
aren't
changing
very
often
are
changing
very
quickly?
This
is
less
of
an
issue,
but
for
particularly
for
machines
where
you're
constantly
trying
to
change
the
operating
conditions.
This
is
a
big
problem.
B
Another
one
is
how
best
to
make
use
of
diverse
data
sources,
so
the
biggest
one
for
us
is
trying
to
combine
data
from
simulation
and
the
machine.
So
how
do
you?
How
do
you
evaluate
you
know
how
to
how
to
which
one
to
trust
when
they
give
you
conflicting
answers,
and
how
do
you
incorporate
those
in
a
seamless
way
into
some
of
these
models
and
another
big
one
is
scaling
to
higher
dimension.
B
As
I
said,
one
of
the
the
goals
for
all
of
this
is
to
do
kind
of
larger
scale,
start
to
end
system
control
and
so
far
we've
been
demoing
stuff
on
much
smaller
scale
systems
and
accelerator
subsystems.
So
this
is
a
big
open
area
and
even
for
things
like
generating
surrogate
models
from
simulation,
how
to
adaptively
sample
the
parameter
space
to
maximize
information,
gain
per
point
as
a
big
open
question,
and
then
we're
also
working
on
combining
the
strengths
of
different
methods,
so
gp
based
bayesian,
optimization,
really
excels.
B
When
you
don't
have
like
a
lot
of
previous
data
that
you
can
use,
whereas
neural
nets
perform
really
well
in
this
kind
of
high
data
regime.
But
really
it's
really
difficult
to
get
an
accurate
model
when
you,
when
you
don't
have
very
much
data
and
then
it's
also
hard
to
get
a
reliable
uncertainty
estimate.
So
combining
these
things
and
doing
neural
network
based
bayesian
optimization
this
kind
of
things.
B
B
Yet
I
would
say
in
this
community
and
aside
from
that,
everything
that
I
talked
about
so
far
involves
basically
single
time,
step,
input
single
time,
step,
output,
data,
but
really
there
are
some
systems
in
accelerators
where
you
have
to
take
into
account
the
time
and
evolution
of
the
system
as
you're
doing
control.
So
this
includes
things
like
cryogenic
system
control
and
control
over
things
like
the
resonant
frequency
of
rf
cavities.
B
B
So
it's
it's
not
just
a
it's,
not
just
one
of
these
cases
where
we
we
can
take
methods
that
are
fully
developed
in
the
ml
community
and
apply
them
to
the
accelerator.
This
actually
could
be
an
interesting
test
bed
to
push
forward
some
of
those
methods,
and
then
we
also
have
pretty
strong
incentives
for
trying
to
improve
system-wide
control
of
the
accelerator.
B
We
want
to
switch
really
quickly
oftentimes
between
highly
customized
beam
setups.
We
actually
want
to
push
beam
quality,
to
sort
of
unprecedented
levels
that
we
haven't
been
able
to
achieve
before
by
exploiting
a
wider
variety
of
controls
on
the
machine,
and
then
this
also
should
in
principle,
enable
us
to
to
actually
tune
more.
Finally-
and
lastly,
this
is
not
just
for
large
scientific
user
facilities.
B
A
Great,
thank
you
very
much
aurelie.
That
was
really
interesting.
So
we
have
time
for
a
few
questions.
If
you
have
a
question,
feel
free
to
unmute
yourself
and
ask
or
raise
your
hand,
I
see
that
we
have
one
talk
in
the
chat
already
hungry.
Do
you
want
to
read
read
your
question.
Ask
your
question
or.
B
This
is
something
that
we've
we've
tried
looking
at,
so
there
are
actually
approaches
where
you
can
try
to
increase
the
the
ability
of
gaussian
process
models
to
handle
like
high
dimensional
cases
like
this,
so
some
of
them
combine
bayesian,
neural
net
methods
with
bayesian
optimization.
Some
of
them
simply
try
to
learn
kind
of
a
higher
dimensional
kernel
by
attacking
a
neural
net
on
top
of
the
gaussian
process
model.
A
C
Yeah
early,
this
is
a
really
really
awesome
talk.
So
thank
you
and
I
I
wanna.
I
have
a
question
about
custom
beams,
but
I
wanted
to
first
say
that
you
know
nurse
really
speaks
the
language
of
science
per
dollar
too,
and
so,
wherever
you
know
wherever,
if,
if
simulations
rather
than
reduced
order,
models
can
be,
are
part
of
that.
C
You
know,
let's
do
the
calculus
of
of
where,
where
it
fits
and
the
the
the
notion
of
of
custom
beams,
I'm
wondering
whether
or
not
you
could
basically
put
a
you
know
a
price
estimate.
You
know
or
a
you
know,
whether
that's
in
in
time
or
in
other
ways
whereby
you
know
a
science
team
could
propose
that
they'd
like
light
or
electrons,
to
behave
in
a
certain
way
and
that
you
could
describe
for
them
how
how
difficult
that
is,
if
there's
a
kind
of
a
metric
for
that.
B
B
And
in
in
contrast,
if
you
have
instead
used
some
of
these
ml
methods
to
just
run
a
bunch
of
simulations
on
nurse,
for
example,
I
think
we
actually
did
some
calculation
for
one
of
our
surrogate
models
where
we
tried
to
get
an
estimate
of
the
equivalent
time
that
it
took
us
to
run
the
number
of
simulations
we
needed
in
order
to
actually
produced
the
surrogate
model
that
we
could
then
use
more
directly
in
optimization
that
I
think
it
was.
It
was
something
ridiculous,
ridiculously
small.
B
It
was
like
200
or
something
in
equivalent
operating
time.
As
far
as
you
know,
just
trying
to
run
those
simulations
in
parallel
on
mars
and
then
it's
kind
of
there's.
No,
I
think,
there's
no
direct
case
that
we
have
right
now
where
we
can
say
like.
Oh,
in
this
case,
we
used
a
neural
net
surrogate
model
that
we
generated
from
massively
parallel
simulations
on
nurse
and
that
sped
up
the
commissioning
time
for
this
new
operating
mode
so
that
we
saved
you
know
this.
B
This
other
amount
of
machine
time
and
here's
the
difference
in
dollar
amount.
We
don't
really
have
that,
but
it's
it's
certainly
interesting
to
think
about.
I
think
that,
fundamentally
trying
to
to
develop
better
tools
for
the
offline
planning
of
these
setups
and
then
coupling
those
with
better
online
optimization
is,
is
going
to
save
like
quite
a
lot
of
time
in
the
in
the
operating
budget.
B
C
A
Great
thank
you,
so
we've
actually
come
to
one
o'clock
at
the
top
of
the
hour
now.
So
I
think
we
need
to
adjourn,
but
perhaps
already
people
can
follow
up
with
you
with
you
offline
with
questions.
If
people
have
additional
questions,
yep,
certainly
all
right
great.
Well,
let's
thank
you
very
much
again
already.
That
was
a
great
talk
and
thanks
everyone
for
joining
us.