►
Description
The 27th Annual CESM Workshop will be a virtual event. Specifically, the Workshop will begin with a full-day schedule on 13 June 2022 with presentations on the state of the CESM; by the award recipients; and two presentations from our invited speakers in the morning, followed by order 15-minute highlight and progress presentations from each of the CESM Working Groups (WG) in the afternoon.
- Introduction: Co-chairs
- The Climate Response of Convective Parcel Modification in CAM: Rich Neale (NCAR)
- Progress towards improving momentum flux parameterization in CAM: Colin Zarzycki
- Fully Unified Boundary Layer and Convection Parameterizations in Global Models: Recent Results in Three-Dimensional Simulations: Joao Teixera
- Atmospheric turbulence modeling using AI: Pierre Gentine
- Implementation of the Secondary Ice Production mechanisms in the Single Column Atmosphere Model Version 6 (SCAM6) Results from MC3E: Arti Jadav
- A renaissance in cloud superparameterization emerging flexibilities for regionalization and machine learning emulation opening new CESM science opportunities: Mike Pritchard
- Discussion
- Understanding the Relationship between Tropical Cyclone Precipitation and SST Utilizing a CAM Hierarchical Framework: Alyssa Stanfield
- Orbital and CO2 effects on tropical cyclones in an ultra-high resolution global climate model: Pavan Harika Raavi
- Investigation of Climate Dynamics Over Euro-Mediterranean Region Using a Variable-Resolution Global CESM Model: Burcu Zehra Boza Karul
- EarthWorks: Progress towards a storm-resolving coupled model: Jim Hurell
- Linking the large-scale processes and the landfalling features of atmospheric rivers in non-hydrostatic CESM simulations: Xingying Huang (NCAR)
- Discussion
- Optional Informal Discussion
To learn more:
https://www.cesm.ucar.edu/events/workshops/2022/
A
B
B
Between
11
45
and
12
30.,
obviously,
the
discussion
is
is
is
optional
for
for
for
everyone,
but
at
this
discussion
I
would
like
to
spend
we.
We
co-chairs
would
like
to
spend
some
time
discussing
results
of
a
recent
survey
that
we
sent
out
on
on
the
possibility
of
removing
some
things
from
cam
to
sort
of
reduce
the
reduce
the
workload
on
our
software
engineering
group.
We've
gotten
lots
of
feedback,
and
I
think
it
it's
it's.
B
B
C
Yeah
so
I'll
take
it
over
and
just
provide
a
a
couple
reminders
for
the
speakers.
So
in
this
first
session
we
have
six
talks
before
our
first
discussion
right
before
the
you
know,
coffee,
slash,
bio,
break,
just
a
reminder
to
the
speakers.
You
have
15
minutes
total,
so
I'm
probably
gonna
turn
on
my
camera
and
interrupt
you
at
14
minutes,
because
we
want
to
stay
on
time
and
it's
going
to
require
a
little
bit
of
there's.
C
Always
the
transition
cost
from
going
to
speaker
speaker
as
well
as
maybe
the
time
for
a
quick
clarification
question
so
so
heads
up
so
try
to
keep
it
to
14
minutes
or
less,
but
preferably
even
less,
to
have
some
initial
questions
so
to
get
us
started.
I
see
that
rich
is
here.
So
our
first
speaker
today
is
rich
neil,
so
rich.
If
you
want
to
start
sharing
your
screen,
so
rich
is
going
to
talk
about
the
climate
response
of
convective
parcel
modification
in
cam.
So
take
it
away
rich.
C
C
D
D
Thanks
kevin
julio,
so
hopefully
this
will
just
serve
as
a
kind
of
wrap
up
to
some
modifications.
We've
recently
been
implementing
in
the
model
and
ostensibly
it's
to
address
some
of
the
problems
we've
seen
when
we've
increased
the
vertical
resolution
and
in
particular
working
towards
the
target
workhorse
model
up
to
58
levels
and
that's
10
levels
in
the
boundary
layer.
D
So
that
will
be
the
first
part
of
there
there's
kind
of
three
purple
lines
there
I'll
talk
about
this
initially
and
hopefully
that'll
kind
of
be
done
for
the
implementation,
a
little
bit
of
a
mention
of
its
performance
in
amip
and
how
we
have
to
make
a
fix
over
land.
And
finally,
if
I
have
time
I
want
to
talk
about
another
problem,
I
think
that
may
need
some
addressing
with
respect
to
deep
convection
oops.
D
Okay.
So
when
we
looked
at
the
implementation
of
these
extra
vertical
levels
is
58
levels.
The
surf,
the
surface
level
is
much
thinner.
Now
I
think
it's
aborted
10
to
15
meters
as
opposed
to
40
or
so
meters,
the
center
of
the
level
with
the
standard,
32
levels.
And
so
when
we
looked
on
the
right
hand
side
here,
we
can
see
see
my
cursor
is
working.
D
Sorry,
these
are
just
simple
single
column
models,
toga
core,
so
west
pacific
convective
region
and
we're
looking
at
the
convective
heating
on
the
right
here
that
results
from
the
deep
convection
scheme,
and
these
are
a
variety
of
vertical
resolutions.
One
extreme
case
here
is
256
levels,
but
that's
just
to
illustrate
this
surface
level
sensitivity.
D
When
we
look
at
this
heating
averaged
over
this
toggle
port
period,
we
can
see
that
you
know
there's
vertical
distributions
of
heating.
You
can
argue
about
the
you
know
the
correctness
of
these
different
vertical
resolutions,
but
one
thing
we
really
did
notice
is
that
surface
level
cooling,
which
was
from
dandruff,
really
does
start
to
get
excessive.
D
The
with
the
the
gradual
thinning
of
these
levels
that
result
with
58
256
and
if
we
and
that's
in
the
average,
but
on
the
left-hand
side,
if
we
had
to
bin
all
the
kind
of
time
periods
of
the
these
single
column
model
runs.
We
can
see
that
with
this
pdf,
you
know
this
256
level
in
particular
we're
getting.
D
You
know:
values
of
minus
60
kelvin
per
day
for
a
you
know,
for
a
20
minute
time,
step
for
the
58
level
which
we're
entered,
which
we're
implementing
for
this
workhorse
we're
also
getting
values
which
we
thought
were
excessive
compared
to
the
blues
and
the
blacks
here,
which
were
the
existing
32
level.
48
levels
is
the
same
within
the
boundary
layer,
so
we
thought,
can
we
implement
something
to
offset
this
until
so
we
don't
get
such
extreme
values
and
and
this
non-convergence
of
the
response
in
the
boundary
layer.
D
You
know
just
to
briefly
summarizes
the
the
scheme:
searches
for
the
maximum
where
the
moist
static
energy
is
the
maximum
level,
and
that
is
almost
always
in
many
many
circumstances,
the
lowest
model
level,
and
so
there
you
see
could
be.
The
problem
is
that,
with
a
gradual
thinning
of
these
levels
that
particularly
with
58
levels,
one
can
see
that
it's
always
that
surface
level,
and
so
there's
there's
a
kind
of
there's
a
problem
as
you'll
see,
there's
a
problem
with
that
and
as
you've
seen
already
with
the
stronger
tendencies
on
the
right
hand,
side.
D
We
propose
this
modification
where
we
introduce
an
a
vertical
scale.
In
this
particular
case,
it's
half
the
boundary
layer
and
that
determines
the
properties,
and
this
is
the
the
moist
static
engine
of
the
pbl
and
the
level
of
which
the
parcel
now
initiates,
which
is
the
top
of
this
vertical
scale.
So
in
in
principle,
this
could
be
insensitive
more
insensitive
to
the
vertical
resolution
changes
just
to
illustrate
how
that
changes.
The
way
the
parcels
are
launched.
We
can
see
here
again
for
the
toga
core
period.
D
The
solid
lines
are
essentially
this
control
of
these
vertical
resolutions,
and
it's
all,
as
I
mentioned,
it's
always
launched
from
the
lowest
model
level,
apart
from
a
little
blip
here.
But
then,
when
we
implement
this
pulsar
modification,
we
can
see
there's
a
much
more
dynamic
level
and
pulse
of
property
of
variation
such
that
during
dry
periods.
The
essentially
the
connected
cloud
base
is
higher.
D
If
we
were
to
now
look
at
those
different
vertical
resolutions
with
this
implementation
of
this
parcel
modification,
we
see
a
the
values
are
much
weaker
than
others
strong
in
this
near
surface
level
which
we're
you
know,
we
kind
of
hope
for
and
they
somewhat
collapse
onto
each
other.
So
now
it's
determined
more
by
the
specification
of
this
mixing
length
scale,
rather
than
the
fact
that
one
of
them
has
a
20
meter
level,
the
near
surface
or
80
meter
near
surface.
So
we
jump
to
applying
that
in
an
amit
simulation.
D
These
20
kelvin
30
kelvin
surface,
and
this
is
a
daily
value
in
this
case.
Sorry,
so
it's
30
kelvin
for
a
day
from
the
d
convection,
which
seems
excessive
and
of
course,
the
the
surface
doesn't
cool
by
that
much
obviously.
Otherwise
you
know
they
get
crazy
mixing.
What
happens
is
this
is
all
the
same
kind
of
tendencies
but
from
the
club
surface
tendencies.
So
there's
this
large
push-pull
between
these
two
different
schemes
here
so
more
broadly
in
the
when
we
think
about
the
amip
simulations.
D
How
is
this
performance?
How
have
we
changed
the
performance
of
the
model?
More
generally?
Oh,
let
me
just
go
to
the
bottom
figure.
First,
actually
so
now,
this
is
a
special
diagram
of
what
that
mean
tendency
of
that
surface
level
is
across
the
globe.
So
you
can
see
this
regional
variation
and
so
yeah,
just
l32
l58,
pretty
similar
of
some
mean
magnitudes
of
order.
D
You
know
four
five
six
kelvin
per
day,
but
then
this
in
the
initial
implementation
of
58,
it's
you
know
up
to
15
16
beyond
in
the
west
pacific
they're
killing
per
day,
and
then,
when
we
apply
this
parcel
modification,
we
do
see
yeah
amelioration
of
this
impact
and
again,
if
our
target
is
to
some
extent
the
l32
and
48
lower
level
heating
calling
tendencies
it's
a
little
bit
too
much,
but
that,
but
it's
not
clear
that
that
is
actually
our
target.
So,
and
this
is
changeable
and
we'll
investigate
how
this
changes.
D
So
that's
encouraging
and
you
know
that's
without
any
tuning
whatsoever,
so
that's
pretty
encouraging
and
just
to
iterate
some
of
the
kind
of
regional
improvements
changes
that
we
see.
This
is
winter
time
mean
precip
over
the
maritime
continent,
which
has
always
been
a
problem
anyway
in
terms
of
in
the
control
version
l32
over
beginning
here,
for
example,
and
then
some
of
the
other
islands,
but
bottom
left
here.
D
D
Compared
to
gpcp,
but
even
that
doesn't
show
this
particularly
over
the
top
of
the
island's
precip
maximum
and
then
yeah.
This
is
the
erroneous
implementation
on
the
bottom
here,
but
then,
when
we
make
the
corrected
implementation
on
the
bottom
right,
we're
really
quite
happy
with
the
kind
of
level
of
precip
biases
precept,
more
precip
fields
and
like
smaller
biases
over
the
mountains
and
really.
E
D
D
D
What
we've
noticed
in
developing
the
vertical
grid
is
that
we're
having
trouble
communicating
between
the
fruit
troposphere
and
the
stratosphere
in
terms
of
things
like
the
tech
recorder,
qbo
et
cetera,
and
I
went
back
and
looked
at
what
the
vertical
heating
profiles
of
the
model
over
different
evolutions
between
country
and
chem
6.,
and
these
again
this
is
a
coca-cola
single
column
model
simulation
on
the
left-hand
side.
This
is
the
mean
vertical
heating
profile,
and
you
can
see
that
you
know
in
the
kind
of
crown
jewels
of
having
de-convective,
parametrization
and
heating.
D
Is
that
you
get
a
as
together
shows
on
the
right.
You
get
these
top
heavy
heating
profiles
either
for
like
kind
of
convective
systems
versus
stratiform
systems.
You
know
four
to
five
hundred
millibars
from
can
three
through
the
camp
six.
We
have
made
changes
to
this
convection
which
have
degraded
that
significantly
and
so
now
we're
at
a
convective
average
convective
pipe
from
deep
convection
of
around
600
to
500
millibars.
Now.
D
Upper
troposphere
from
other
moist
processes,
club,
etc,
microphysics.
But
you
know
it
is
not
that
desirable
and
I
think
you
know
the
first
question
is:
oh
my
gosh:
why
did
we
do
that?
It
looks
terrible,
but
of
course
there
are
other
aspects
that
we
you
know
that
were
improvements,
and
this
is
the
precip
distribution,
for
example,
during
that
report
period,
and
you
can
see
the
you
know,
demonstrable
improvements
between
cam3
and
successive
models
and
things
like
the
mgo
and
down
cycle
do
improve.
So
this
is
this.
Is.
G
D
Unknown,
but
this
interplay
between
variability
and
like
mean
forcing
but
one
thing
I've
been
investigating,
is
how
can
we
make
convection
deeper
in
the
model
and
this
criteria
of
deciding
where
the
type
of
convection
is
is
based
on
a
buoyancy,
profile
and
and
the
way
it's
set
up
in
camp
six
as
soon
as
you
hit
a
first
level
of
negative
buoyancy,
that's
the
top
of
convection
and,
of
course,
that
you
know
feels
unrealistic.
It
should
be
a
depth
of
convection.
D
If
I
switch
to
precipitate
at
the
expense
of
the
simulation,
at
least
in
the
single
column
model,
in
fact
it's
quite
a
few
improvements
here.
So
the
red
is
the
existing
and
shallow
connection
and
blue
is
the
deeper.
So
that's,
I
think,
that's
just
a
plan
to
seed
in
people's
minds
that
I
think
we
need
to
think
about
so
I'll.
Yeah
summary
three
parts
we've
implemented
the
parcel
modification
to
address
some
of
these
near
surface
issues.
We've
implemented
this
and
aim.
D
C
B
D
D
D
G
C
Okay,
rich,
I
think,
there's
another
question
in
the
chat.
If
you
don't
mind
answering
it
in
the
chat
yeah,
you
can
go
forward.
So
let's
switch
to
our
next
speaker,
which
is
we're
going
to
switch
to
hearing
a
little
bit
about
the
climate
process
team
so
updates.
So
our
next
speaker
is
colin
zarziki
who's
going
to
talk
about
the
progress
towards
improving
momentum,
flux
primarization
in
cam,
so
looks
yep,
you're
you're
good
to
go
colin.
E
All
right
well
well
thanks
kevin
and
and
thanks
to
the
the
co-chairs
for
for
running
this
mess
so
yeah.
So
I
wanted
to
just
give
a
brief
update
on
our
cpt,
which
again
is,
is
tasked
with
improving
the
representing
representation
of
momentum,
flux,
parameterization
and
we're
doing
so
via
club.
So
you
know
just
a
reminder
that
the
cpt
is
jointly
funded
by
nsf
and
noaa.
So,
thanks
to
thanks
to
them,
let's
see
there
we
go
so
this
is
the
slide.
E
E
So
this
is
this
new
taus
formulation,
we've
brought
in
your
immediate
reaction
might
be.
Oh
no
they've
brought
in
more
of
those
parameters
that
I
still
don't
really
understand.
What's
going
on
with
you
know,
things
like
the
c2
c14,
but
I
kind
of
want
to
emphasize
that
you
know
in
doing
so.
We've
really
designed-
or
you
know,
vince
and
engine
guo
and
some
other
folks
have
been
working
on
this
code.
E
I've
probably
presented
on
some
of
this
is
that
we
actually
now
have
this
this
club,
what
we
call
club
x
or
this
code
with
those
two
improvements,
are
now
on
cam
development,
so
they
were
merged
in
as
of
the
tag
63059,
which
was
I
think
about
three
weeks
ago,
and
so
big
thanks
to
kate,
they
are
called
her,
who
did
really
the
heavy
lifting
on
this.
This
was
much
more
challenging
than
any
code.
E
I've
previously
been
working
with
in
terms
of
getting
on
onto
cam
development,
but
also
adam
ben
steve,
cheryl,
gunther,
vince,
jesse
courtney,
john,
who
were
all
people
that
you
know
at
least
commented
on
the
pull
request.
So
I
apologize
if
I
missed
your
name.
So
this
was
really
like
a
a
big
kind
of
aspect
of
of
you
know
getting
this
into
cam,
so
we
can
start
really
evaluating
what
what
we're
getting
in
terms
of
global
climate.
E
In
the
late
summer
we've
been
talking
about
separating,
maybe
the
damping
of
scale
or
fluxes
from
momentum
fluxes,
currently
they're
they're
treated
with
their
handled
with
the
same
subroutine
if
we
partition
them,
that
does
increase
club's,
cost
so
kind
of
we're
thinking
about
the
practicality
of
that.
But
some
initial
results
show
that
maybe
there
could
be
additional
improvements
gained
from
kind
of
almost
having
essentially
an
if
statement
in
the
code
on
that,
but
but
not
in
cam.
E
But,
for
example,
we've
introduced
this
new
parameter
that
really
controls
turbulence
and
sheared
flow,
and
what
we
found
is
that
you
know
obviously
a
term
that
impacts,
turbulence
and
sheared
flow
seems
to
have
a
relatively
large
impact.
So
if
you
look
kind
of
horizontally
on
the
stripe,
this
is
the
planetary
boundary
layer,
height,
10,
meter,
wind
speed.
Anything
that
you
see
is
a
a
darker
stripe
means
that
that
club
sensitive
perturbing,
that
club
parameter
drives
sensitivity
in
that
output
metric
in
the
global
cam
solution.
E
and
then,
with
this
information
about
sensitivity,
you
know
which
parameters
are
kind
of
controlling
which
aspects
of
the
the
climate
system.
This
kind
of
allows
us
to
now
start
to
think
about
kind
of
understanding
and
evaluating
tuning.
So
this
is
one
thing
that
we've
been
looking
at
mainly
led
by
the
uwm
group
and
vince
and
his
postdocs
and
grad
students,
but
really
a
simple
kind
of
almost
like
linear
model
of
the
sensitivity
of
these
different
fields
like
things
like
stratocumulus,
cumulus
clouds.
E
Things
like
that
to
these
different
parameters,
so
just
an
example
here
of
you
know
how
your
stratocumulus
might
change.
If
you
change
c8
and
then
you
can
actually
use
this
to
almost
do
this
really
simple
matrix,
optimization
to
try
to
understand
before
actually
running
a
simulation.
This
is
the
type
of
thing
that
you
could
kind
of
almost
like
code
up
in
matlab
run
and
then
try
to
predict.
Maybe
what
the
the
cam
output
would
be.
E
The
the
kind
of
tuning
that
I
want
to
briefly
touch
on
here,
which
we've
been
a
little
more
focused
on
at
penn
state,
is
a
little
more
of
a
kind
of
physically
based
optimization
strategy.
We've
been
using
this
idea
of
what's
called
nelder
mead,
and
this
has
been
sporadically
applied
for
climate
models
previously.
E
But
what
it's
essentially
showing
is
what
we
do
is
we
kind
of
run
a
simulation
of
cam
where
we
hold
these
two
club
parameters
that
we
know
are
really
important
for
the
boundary
layer,
height,
let's
say
c14
and
ck10.
We
run
this
baseline
configuration
with
our
default
values,
and
then
we
essentially
start.
You
know
we
blind
this
algorithm
to
what
those
default
values
are,
and
we
say:
okay,
algorithm
you're
on
your
own.
We
have
no
expert
judgment.
E
So
this
is
I'm
going
to
just
show
one
one
quick
top-level
example
of
doing
this
with
club,
so
one
error
we've
had
with
cam6
and
club
is:
we've
had
this
issue
where,
if
we
improve
the
sea
level
pressure
fields
via
tuning,
it
degrades
the
surface,
stress
fields
and
vice
versa.
It's
been
tough
to
really
get
an
optimization
of
both
of
them.
E
So
we're
going
to
use
this
nelder
mead
algorithm
and
we're
actually
going
to
try
to
minimize
essentially
nudging
errors
in
the
model,
and
what
we
find
is
that
if
we
use
this
new
kind
of
formulation
of
club
with
the
prognostic
momentum,
fluxes
and
the
taus
code,
this
is
the
out-of-the-box
cam6.
So
I
think
cesm2
and
we
use
this
nelder
mead.
E
Anything
that's
highlighted
in
orange
shows
a
better
global
skill
just
using
the
atmospheric
model,
working
group
diagnostics
packages
and
again,
I'm
not
really
going
to
spend
a
lot
of
time
on
this
and
get
into
the
weeds.
But
essentially
we
see
a
lot
more
orange
than
we
see
a
lot
more
yellow
over
here,
so
that
at
least
club
x
with
this
tuning
shows
some
positive
potential
and
in
terms
of
improving
some
of
the
biases
and
the
climate
model,
we
can
also
use
this
tuning
for
interpretability.
E
The
nice
thing
about
this
is
with
the
prognostic
momentum
is
we
can
essentially
look
at
you
know,
for
example,
what
is
club
x
doing
differently
than
cam6
that
provides
better
global
skill
metrics.
So
one
thing
we
find
is
that
this
red
here
is
kind
of
what
club
x
wants
to
do
with
this
optimization
blue.
Is
this
cam
6
out
of
the
box?
E
So
this
is
a
slide
from
yo,
kim
and
ganila
at
stockholm
university,
but
essentially
this
black
curve
here
is
where
we
want
to
be
in
terms
of
observations
of
wind,
turning
distributions
and
this
red
curve
is
essentially
what
we
have
for
for
cam,
2.
or
c
cam.
Sorry,
cam,
6
csm2.
So
we
think
there's
plenty
of
room
for
improvement
with
regard
to
wind,
turning
ekman
spiral
associated
with
momentum,
fluxes
and
then
last
slide.
E
I
promise-
and
this
is
one
thing
I
do
think
we
need
to
talk
about
in
the
amwg,
but
we've
noticed
that
with
the
higher
vertical
resolution
as
well
as
these
less
diffusive
configurations
of
club,
we
started
to
see
a
lot
of
noise
crop
up,
especially
over
land.
In
terms
of
things
like
the
10
meter,
wind
speed
field,
you
know
julio
showed
a
slide.
E
E
This
code
is
now
all
merged
onto
cam
development.
So
I
guess
the
big
thing
I
want
to
say
is:
neither
option
is
on
by
default,
but
it's
pretty
easy
to
toggle
for
tuning
and
evaluation.
So
if
it's
something
you're
interested
in,
please
let
me
know,
I'm
I'm
happy
to
kind
of
work
with
you
on
maybe
exploring
this
the
more
eyes
we
have
on
it
the
better
and
then
also.
I
will
say
that
while
we've
been
developing
this,
we
have
found
that
this
is
more
stressful.
E
C
All
right,
thanks,
colin,
I
have
maybe
time
for
a
quick
question
if
there
is
any
go
ahead
and
raise
your
hand,
but
I
think
I
don't
see
anybody
jumping
up,
but
we'll
continue
the
discussion
in
the
the
discussion
at
the
end
of
the
talks
here.
So
thank
you,
colin,
so
to
kind
of
continue
with
the
climate
process
teams.
Our
next
speaker
is
jiao
te
shera.
It's
going
to
talk
about
the
fully
unified
boundary
layer
and
convection
primarization
in
global
models.
C
C
G
C
J
C
I
C
C
K
I
I
Much
sorry
for
the
delay,
no
problem
I'll
I'll,
be
on
time
I'll
stay
on
time.
Thank
you
again
for
the
the
opportunity
to
do
this
today.
We're
gonna
have
update
you
on
some
recent
results
from
what
we
call
a
fully
unified
boundary
layer
and
convection
pressurization
in
cam,
which
means
a
boundary
layer.
Shallow
convection
ended
convection
parameterization,
so
no
need
for
an
explicit
convection
ceramization
in
our
simulations.
I
I
I
So
how
do
we
do
this?
I
will
obviously
be
brief
in
this
respect,
but
the
essence
of
the
story
and
we'll
show
you
a
schematic
that
highlights
that
is
that
club
ends
up
representing
sort
of
this
double
double
gaussian
mixing,
while
this
multiple
plumes
represent
additional
discrete
skewness
of
the
sub
grid,
pdf
and
and
you'll
see
that
in
two
slides
from
next
the
multiple
mf
and
I'll.
I
have
a
very
quick
slide
on
that
two
things
to
keep
in
mind.
I
The
way
this
is
actually
implemented,
code,
wise
and
algorithmically
wise
is
with
with
club-
is
that
we
do
it
via
the
five
diagonal
prognostic
solver
that
exists
for
solving
min
fields
and
turbulent
fluxes
that
are
solved
simultaneously
within
club,
and
so
what
we
do
is
that
we
add
this
extra
explicit
right
hand,
side
term,
that
you
can
see
in
this
equation,
based
on
the
multiple
plumes,
so
directly
affects
the
equation.
For
the
for
the
mean
flow
for
the
mean
field,
it
indirectly
affects
the
the
fluxes
as
well,
because
this
is
solved
simultaneously
now.
I
I
We
sample
that
pdf,
so
some
of
these
plumes
initially
when
they
initialize
already
are
more
energetic
than
others
and
at
the
same
time
they
will
go
and
encounter
different
levels
of
entrainment
and
different
numbers
of
entraining
events,
and-
and
this
is
basically
partly
inspired
by
a
paper
by
rom
but
has
a
bunch
of
modifications
recently
and
the
two
main
messages
here.
I
think
the
main
one
is
that,
because
we
have
multiple
plumes,
we
are
able
to
have
different
types
of
convection
coexist
in
the
same
model
breed
box,
so
you
will
have
dry
convective
plumes.
I
You
will
have
shallow
convection,
you
have
deep
convection
depending
on
the
environment.
There
are
no
trigger
functions
or
sort
of
partitions
or
anything
they
will
can
all
consist
in
a
friendly
way
within
a
green
box.
Now.
Another
aspect
which
is
secondary
but
but
relevant
is
that
the
total
updraft
area
and
the
properties,
the
abstract
properties,
are
basically
the
sum
of
the
individual
updraft
areas
at
a
certain
level
right,
so
some
plumes
will
live
up
to
that
others
will
not.
Now.
I
Let
me
just
show
you
there's
a
a
very
nice
paper
led
by
michael
white.
I
think
it
either
is
about
to,
or
was
already
published
in
monthly
but
review.
The
review
process
is
finished.
It
covers
essentially
single
column
model,
single
call
modeling
with
cam,
with
scam
for
shallow
convection
events.
This
is
just
one
example.
This
is
the
bomex
case
on
the
left.
You
have
the
liquid
water
on
the
right.
I
But
you
know,
essentially
what
you
can
see
is
that
there
is
an
an
extra
vertical
mixing
that
is
provided
by
the
plumes
and
that
allows
the
red
line
to
be
closer
to
the
black
line,
so
to
for
club.
Mf
is
close
to
the
lias
results
and
one
sort
of
simple
way
to
see
that
is
by
analyzing
this
schematic.
That
shows
basically
the
perturbations
at
a
certain
level,
in
this
case
at
730
meters
above
the
ground.
That's
about
250,
about
above
cloud-based.
I
This
is
just
the
perturbations
in
liquid
water
potential
temperature
versus
perturbations
in
total
water.
You
have
three
components
in
this
image.
You
have
the
les
joint
pdf,
it's
a
histogram
of
a
bunch
of
lds
points
in
that
level
across
the
entire
domain,
every
10
by
10
meters,
or
so
you
can
see
the
club
joint
pdf.
That
is
the
sort
of
parameterized
club
joint
pdf
from
from
cam,
and
then
what
you
can
see
is
that
that
pdf
really
does
not
go
across
the
space.
I
I
You
can
see
the
stratocumulus
regions
off
the
west
coast
and
then
slowly
the
amount
of
clouds
just
decreases
as
you
move
towards
towards
the
itcz
and
at
the
bottom.
You
can
see
the
june
july
august
cloud
fraction
at
different
levels
from
both
club
control
and
club
mf
on
the
right.
The
club
control
by
the
way,
as
well
as
club
mf,
is
based
on
the
l58
version.
I
Not
the
l32
I'll
show
some
l32
in
the
next
box,
but
this
one
is
the
l58
and
that's
basically
the
control
we're
comparing
to
at
the
bottom.
You
see
two
different
sort
of
remote
sensing,
the
action
slightly
different
characteristics
and
and
resolution,
but
indeed
what
you
can
see
is
that
in
in
this
case,
both
the
club
control,
l58,
not
l32,
would
not
be
able
to
do
this
as
well
as
you
all
well
know,
and
club
mf
seem
to
reproduce
the
behavior
that
you
would
expect.
I
There
are
slight
differences,
but
both
of
them
seem
to
create
stratocumulus
in
the
right
regions,
and
then
the
boundary
layer
gets
deeper
from
about
500
meters
off
the
coast
to
about
two
kilometers
across
hawaii
and
with
much
less
cloud
cover,
as
you
would
expect
since
you're
entering
the
shallow
convection
region.
So
it
does
a
good
job
in
pvl
clouds.
I'll
show
you
maps
of
that.
This
is
the
annual
mean
low
cloud
cover
for
this
20-year
ami
plan.
The
observation
so
I'll
show
you
two
or
three
plots
that
are
very
similar.
I
We'll
have
three
maps
at
the
top
three
observations:
level
30,
12.
l32
is
the
second
one,
then
l58
both
control
of
mf
at
the
bottom,
and
you
can
see
that
the
sort
of
the
the
benefits
that
that
were
present
in
the
previous
plot
are
here
as
well.
The
observations
clearly
show
regions
of
a
lot
of
low
cloud,
pvl
cloud
off
the
west
coast
of
continents
in
the
north,
specific
and
north
atlantic,
closer
to
the
east
coast
and
also
all
across
the
southern
ocean.
I
It's
clear
that
l32
is
not
able
to
produce
strata
cumulus
as
it
should.
Both
l58
and
club
mf
are
able
to
produce
stratic
cumulus
regions
with
slight
differences,
but
but
fairly
effectively.
I
would
say-
and
we
did
sort
of
the
added
benefit
for
club
mf-
that
it's
able
to
produce
clouds
in
the
southern
ocean
and
in
the
northeast
pacific.
I
Sorry,
yeah,
the
northwest,
pacific
and
atlantic
with
cloud
cover
that
is
less
than
l58
control
and
closer
to
the
observations.
Now,
because
we
didn't
have
the
explicit
deep
convection.
Of
course,
our
main
interest
was
recently
to
find
out
how
does
club
mf
respond?
To
sort
of
you
know
the
the
the
the
forcing
of
of
of
the
earth's
climate
in
order
to
sort
of
be
able
to
reproduce
or
not
the
characteristics
of
the.
I
So
so,
if
you
look
at
precipitation,
for
example,
so
this
is
annual
imprecipitation
minutes
today,
again
the
observations
levels
well,
l,
32,
l58
and
mf.
All
of
them
seem
to
show
patterns
that
are
not
too
dissimilar
and
that's
very
encouraging.
From
our
perspective,
because
again
we
don't
have
explicit
deep
convection
parameterization
we're
allowing
club
mf
to
produce
deep
convective
clouds
when
the
conditions
arise,
but
not
only
is
club
mf
able
to
reproduce
a
climatology.
That
is
not.
I
You
know
worse
than
the
one
from
the
control
version
is
actually
able
to
produce
rain
in
regions
where
it
appeared
that
the
previous
versions
had
a
little
bit
of
difficulty
in
producing
that
sort
of
amount
of
rain.
Now
you
know
this
is
not
as
conclusive
as
you
know.
You
know
this
is
a
20-year
ami
run,
but
it's
a
very
good
sign
and
then,
if
you,
if
we
look
down
and
I'll
finish
with
two
very
quick
plots,
this
is
high
cloud
cover.
I
I
C
All
right,
thank
you
very
much.
I
think,
in
an
effort
of
time
to
stand
scheduled
we're
going
to
move
on
to
the
the
next
talk,
but
we'll
be
able
to
have
a
discussion
more
broadly
here
shortly.
So
our
next
speaker
is
pierre
gente,
perfect
who's,
going
to
talk
about
atmospheric
turbulence,
modeling
using
machine
learning.
L
So
really,
what
we
want
to
tackle
here
is
trying
to
better
understand
and
better
also
model
the
planetary
boundary
there,
and
we
know
that
there's
actually
very
intense
turbulence
in
the
boundary
layer
due
to
up
drops
and
down
tracks
and
also
the
shear
and
really
what
we
want
to
do
is
trying
to
understand
what
is
kind
of
the
bulk
aperture
representation
of
transport
and
fluxes
throughout
the
boundary
layer.
And
throughout
the
talk
I
mean,
I
will
use
standard
notation.
So
the
average
here
would
be
the
slab
average
and
then
prime
would
be
the
turbulent
fluctuation.
L
L
Velocity
could
be
potential
temperature
or
tribunal,
kinetic
energy
here,
and
that's
really
kind
of
the
the
art
of
the
powerization
game
and
there's
been
quite
a
bit
of
work
on
that
in
that
space
for
many
many
decades
now,
starting
with
the
add
diffusion
model,
which
is
the
simplest
way,
we
can
actually
model
turbulent
flux,
especially
the
vertical
transport
of
turbulent
flux,
but
very
quickly.
People
realize
that
it
doesn't
really
work
so
well
in
the
boundary
layer,
because
it's
actually
very
well
mixed.
L
One
of
the
issue
when
we
start
thinking
about
led
diffusion
and
any
diffusion
mass
effect
is
that
even
with
the
best
tuning.
So
that's
one
example
here.
Looking
at
the
larger
distribution,
where
we
took
an
example
where
we
were
trying
to
find
the
best
sets
of
parameters
that
would
work
across
many
different
simulations
with
various
sheer
levels
and
various
convective
regimes,
and
typically
what
you
will
find
is
it
can
work
in
a
particular
regime.
But
it's
very
difficult
to
actually
generalize
to
regimes
that
you
haven't
seen.
L
L
So
that's.
Basically,
the
objective
is
trying
to
use
machine
learning
to
parameterize
boundary
layer,
tubulin
fluxes,
and
we
didn't
want
to
use
just
a
brute
force
perspective,
but
also
trying
to
see
what
did
we
miss,
for
instance,
when
we
are
looking
at
francis
and
eddie
diffusion
mass
flux,
taking
that
as
the
reference
mark
and
really
trying
to
get
into
the
process
level
understanding
and
trying
to
think
what
is
what
are
the
key
components
of
boundary
layer
turbulence
across
regimes
from
more
shear
driven
to
more
convective
or
buoyancy-driven
regimes?
L
So
what
we
did
is
we
ran
a
bunch
of
larger
simulations.
Actually
now
we
extended
the
data
sets
from
very
sheer,
driven
to
almost
channel
flow,
all
the
way
to
very
very
buoyant
conditions,
and
that
created
a
very
large
data
set
to
to
work
with
to
try
and
see
whether
we
could
actually
predict
the
fluxes
using
machine
learning
and
as
input.
We
will
play
the
typical
game
of
parametrization,
but
you
will
have
the
large-scale
variable
so
potential
temperature
tuberculin,
kinetic
energy,
but
some
other
tracers.
L
L
Now
what
we
wanted
to
do
is
not
brute
force
that
so
in
the
les
we
have
many
vertical
levels
and
we
thought
we
actually
don't
want
to
actually
replicate
that,
because
that
would
be
a
really
very
much
very
much
like
a
brute
force
forcing
and
we
want
to
actually
get
something
a
bit
more
more
a
bit
simpler
and
something
that
would
be
like
a
lower
dimensional
representation
of
turbulence.
So
how
can
we
kind
of
step
back
and
avoid
those
many
different
levels
of
having
this
massive
and
huge
neural
network
in
particular?
L
L
So
what
we
thought
is
we
could
actually
approach
that
using
some
sort
of
latent
space,
where
we
could
try
and
reduce
the
dimension
of
the
system
and
starting
on
the
turbulent
kinetic
energy
profile
and
then
going
into
the
fluxes
and
the
different
variables
and
then
trying
to
combine
that.
Okay,
so
I'll
try
to
walk
you
through
that.
L
So
the
first
step
is
you
take
the
turbulent
kinetic
energy
profile
again
many
many
different
levels,
but
you
don't
want
to
reproduce
that
necessarily
right.
I
mean
you
want
to
have
some
sort
of
embedding
of
that.
Something
that's
way
simpler
than
that,
and
so
that's
what
sarah
did
she
actually
created
this
latent
space
and
I'd
explain
in
a
minute
what
it
represents,
but
it
only
has
four
dimensions.
L
So
it's
a
very
compact
representation
of
turbulence
and
turbulent
kinetic
energy
and
you
have
what
we
call
an
encoder
and
then
a
decoder
so
think
of
that
as
a
non-linear
pca.
It's
very
similar
to
that
and
we
are
trying
to
basically
recreate
the
turbulent
kinetic
energy
again
at
the
large
scale,
and
you
can
do
kind
of
the
same
for
any
other
scalar
you're
looking
at.
So
we
also
did
it
for
momentum,
but
you
can
do
it
for
potential
temperature
and
any
other
skater
as
you
wish,
and
you
can
actually
plug
that
in
there
as
well.
L
You
have
an
encoder,
so
you're
reducing
the
dimension
of
the
large
scale
feature,
and
then
you
combine
the
tk
latent
space,
with
the
variable
latent
space,
decode
that
and
then
you're
getting
your
flux.
So
what
this
is
telling
you
is
that
you
reduce
the
dimensional,
the
dimension
of
the
of
the
tk.
You
reduce
the
dimension
of
your
skater,
combine
them,
and
then
you
can
predict
your
flux.
L
And
you
can
actually
do
that
in
a
relatively
physically
consistent
way,
so
we
can
actually
map
that.
So
we
first
map
the
tke.
Then
you
could
also
also
make
sure
that
you
have,
because
you
will
have
this
latent
dimension
and
you
don't
know
what
those
dimensions
actually
mean.
So
we
try
to
actually
project
that
across
different
modes
of
one
that
would
be
more
shear
driven
and
one
that
would
be
more
convectively
driven
so
that
we
can
actually
disentangle
those
modes.
L
L
So
then
we
can
always
get
the
transport
and
the
main
point
being
that
here
the
mapping
is
always
the
same,
so
the
decoder
is
always
the
same.
The
only
thing
that's
changing
is
really
the
latent
space,
so
how
we
are
actually
reducing
the
dimensionality,
but
the
decoding
part
is
actually
consistent
and
we
could
use
it
for
whatever
we
wish,
you
know
could
be
water
vapor.
It
could
be
some
other
trace
or,
if
you're
looking
at
chemistry
in
particular.
L
So
it's
generic
okay.
So
how
does
it
work
so
in
practice
when
we
look
at
tk,
so
we
look
at.
Basically,
we
took
the
chord
screen
version
of
the
les
and
we
try
to
assess
that
at
the
kilometer
scale.
We
could
do
that
at
multiple
scale,
so
that's
actually
something
we
are
trying
to
do
now
and
the
main
point
being
that
across
those
different
scales
we
can
and
those
different
samples.
We
can
actually
do
a
very
good
job.
L
So
when
we
look
at
all
of
those
samples,
so
those
are
just
examples,
but
they
are
very
typical
of
what
we
are
seeing.
We
can
capture
a
lot
of
the
mode
availability
of
the
tk
sample
in
the
boundary
layer,
and
here
is
a
summary
of
the
result
from
the
from
the
encoder
and
the
spread
here
is
showing
the
variance
and
you
can
see
that
the
model
is
actually
able
to
also
capture
not
just
the
mean,
but
also
the
variance
across
those
many
different
samples.
L
So
it's
kind
of
good
news
because
it
means
it's
not
that
the
model
is
not
very
static.
It
actually
can
represent
a
lot
of
the
variability
we
are
seeing
in
the
boundary
layer
and
across
the
different
samples
that
we
are
seeing
and
the
basically
the
spatial
heterogeneity
that
we
are
seeing.
So
that's
good
news
with
them.
L
We
can
look
at
the
heat
flux
on
the
left
hand
side.
So
what
is
the
quality
of
the
reconstruction
again
only
with
the
dimensions
of
four?
So
we
only
have
four
modes.
You
know
that
are
trying
to
reconstruct
that
and
again
we
can
see
that
we
are
doing
a
very
good
job
here
across
the
different
circles,
and
I
want
to
emphasize
here
that
those
samples
actually
span
a
really
huge
diversity
in
terms
of
sheer
conditions
and
also
buoyancy
and
convective
conditions.
So
this
is
again
good
news.
L
It
means
we
can
really
replicate
what's
going
on
even
at
a
very
small
scale,
and
if
we
look
at
the
r
squared
here,
that's
actually
the
r
squared
across
all
of
the
samples
in
space
and
also
across
all
of
the
simulations.
So
the
the
accuracy
is
actually
really
really
high
and
it
means
that
you
can
really
depict
the
kind
of
the
essence
of
turbulence
in
the
boundary
layer
and
that's
for
heat
flux.
L
But
we
could
do
the
same
for
tracer
and
we
could
do
the
same
for
the
tk
flux
and
again
we
have
very
good
results
across
those
many
different
samples.
And
again,
we
feel
that
we
can
really
kind
of
reproduce
the
heterogeneity
that
we
are
seeing
across
simulations
across
shear
and
buoyancy
conditions,
but
also
within
a
simulation
itself
in
terms
of
the
spatial
heterogeneity.
So
agree,
that's
something
that
for
weather
forecasting
with
the
thing
would
work
relatively
well.
L
So,
like
a
sorry,
a
middle
part
of
the
boundary
layer,
heavy
profile
and
then
a
bottom
heavy
part
on
the
right
hand,
side,
and
what
we
thought
is
that
basically,
this
mode
here
that
we
have
here,
which
is
the
first
latent
space,
is
really
trying
to
explain
biomes.
So
that
would
be
the
equivalent
of
your
mass
flux
component
in
the
edmpf.
L
L
L
So
again,
this
latent
dimension
is
very
physically
based
and
it
can
really
tell
us
a
little
bit
more
about
the
physics
and
just
to
conclude,
we
can
actually
start
investigating
a
little
bit.
What
is
that
telling
us
you
know,
and
so,
when
we
start
thinking
again
about
ad
diffusion,
so
it's
again
you
have
a
gradient,
so
the
vertical
gradient
and
you
have
a
coefficient,
which
is
this
case
or
the
diffusion
coefficient,
and
if
we
were
to
actually
lump
all
of
that
together.
L
So
if
we
wanted
to
actually
map
that
diffusion,
we
would
basically
get
those
lines
right.
So
that's
the
projection
of
the
total
flux
on
the
different
added
diffusion
coefficient
and
you
can
see
it's
kind
of
all
over
the
map
here.
It
doesn't
really
tell
you
much
and
there's
quite
a
bit
of
discrepancy
between
the
sheer
case,
the
very
convective
case,
and
really
we
don't
see
any
any
any
convergence
between
those
things.
L
Now,
if
we
use
the
modes
that
I
just
told
you
before,
like
the
first
mode
to
the
first
laden
space,
and
now
we
project
onto
that
date
and
mode,
we
can
see
that
there
now
there's
a
collapse
actually
of
all
of
the
diffusion
effects,
and
what
this
is
telling
us
is
really
a
diffusion
is
really
need
to
shear
and
that's
why
there's
actually
this
collapse
across
all
of
those
modes,
and
it
means
that
we
can
really
start
understanding.
What
is
the
physical
process
in
at
play
in
every
single
one
of
those
latent
dimensions?
L
So
very
quick
summary.
So
it
seems
that
machine
learning
can
really
help
first,
better
understanding
physics,
but
also
predicting
with
a
very,
very
low
dimension
turbulence
in
the
boundary
layer,
and
one
thing
we
we
came
that
became
very
clear
is
that
we
really
need
to
also
create
this
dialogue
between
domain
knowledge
and
machine
learning,
because
informing
the
retrieval
with
physics,
guided
such
as
projecting
on
the
shear
mode
or
convective
mode,
was
really
really
important
as
a
way
to
actually
represent
those
different
things,
and
we
believe
that
it
really.
L
C
All
right,
thank
you,
pierre!
We
have,
if
there's
any
quick
question
or
so
feel
free
to
raise
your
hand
if
not
can
continue
to
interact
on
in
the
chat.
So
just
as
a
reminder
to
speakers,
as
you
finish,
your
talk,
there
are
questions
coming
across
in
the
chat
if
you
could
refer
to
them,
but
maybe
to
stay
on
time.
C
M
M
So
this
hybrid
bundle
bin
bulk
scheme
means
that
we
during
this
processes,
we
represent
the
the
total
number
and
mass.
But
when,
when
this,
when
this,
the
the
number
and
the
the
number
is
discretized
into
the
size
bins,
so
that
in
the
ice
processes
we
have
added,
we
get
a
better
result
and
we
used
the
mid-latitude
continental
convective
cloud.
Experiment
observation
data
for
our
validation.
M
So
I'll
start
with
the
cloud
eyes,
the
in
the
deep
convective
clouds.
So
the
ice
is
present
in
the
cloud
by
by
initiation
and
multiplication
processes,
and
this
initiatin
initiative
initiation
can
be
by
the
aerosols
which
are
present
in
the
atmosphere
which
is,
and
this
process
but
and
by
which
the
cloud
is
formed
is
called
heterogeneous
ice
nucleation,
in
which
the
cloud
droplets
are,
and
they
become
cloud
eyes.
M
You
have
this
cloud
droplets,
which
just
instantly
freeze
to
become
cloud
eyes,
but
when,
when
we
talk
about
the
cloud
ice
well
from
the
observations,
we
have
seen
that
these
processes
are
not
enough
to
explain
the
total
ice
we
see
in
the
cloud.
So
this
means
that
there
are
other
mechanisms
which
are
active
to
produce
the
ice
and
so
to
represent
this
mechanisms.
We
have
introduced
this
the
ice
multiplication
processes
and
there
are
three
three
well
recognized
secondary
ice
production
mechanisms.
M
The
first
one
is
the
halite
mossep
in
which
the
supercooled
cloud
droplets
you
have
above
and
all
this
is
happening
above
the
zero
zero
degree
level,
which
is
the
freezing
level.
So
so,
then
you
have
the
super
cool
crowd
cloud
droplets
they
rhyme
on
cloud
eyes
and
produce
ice
splinters,
and
this
process
is
active
in
the
region,
minus
three
degrees
celsius
to
minus
eight
degrees
celsius
and
then
the
other.
M
The
next
process
is
the
I
size,
collisional
breakup,
which
is
the
collisions
between
the
different
eyes,
particle
habits,
and
this
theoretical
formulation
is
based
on
the
paper
philips
at
all
2017,
and
then
the
next
process
we
have
introduced
is
the
raindrop
freezing
fragmentation
and
there
are
two
modes
in
this.
The
first
mode
is
where
the
collision
is
between
the
ice
particle
and
the
cloud
droplet.
M
But
here
the
cloud
the
ice
particle
will
is,
has
less
mass
than
the
cloud
droplet
and
the
second
mode
is
where
the
ice
particle
has
a
higher
mass
than
the
raindrop
or
the
the
cloud
drop
and
bo,
and
this
formulation
is
based
on
the
philips
at
all
2018
paper.
So
when
we
added
this
process,
we
saw
a
good
validation
with
the
aircraft
observations
during
this
mc3e
campaign
and
the
black
line
here
here
in
this
graph.
M
I
have
done
a
budget
analysis
using
the
tagging
tracers
and
the
tagging
tracers
are.
I
have
used
for
the
primary
eyes
and
which,
which
comes
from
the
prime,
the
heterogeneous
ice
nucleation
and
the
homogeneous
freezing
and
the
three
sip
mechanisms,
and
I
we
can
see
that
the
halit
muscle
process,
which
is
represented
by
this
magenta
line,
it's
like
between
the
minus
three
and
eight
minus
eight
degrees
and
the
the
most
of
the
eyes
below
the
minus
36
degrees.
M
It
comes
from
the
from
the
from
the
sip
mechanism
we
have
introduced,
and
I
I
I'm
sorry
I
I
failed
to
mention
this
before
the
homogenous
freezing,
which
is
like,
where
the
cloud
droplets
instantly
freeze,
that
is,
it
occurs
above
the
minus
36
degree
levels,
but
here
we
see
that
representation
below
the
minus
36
degree.
That
is
because
of
the
happening
in
the
cloud
so
and
that
is
represented
in
the
tagging
traces
we
have
introduced
in
this
model,
and
this
is
a
better
budget
analysis.
M
M
So
I
think
so,
then
it
was
important
to
have
this
representation
of
this
ice
multiplication
processes
and
the
deep
convective
scheme
in
the
scam
model,
and
our
next
step
would
be
to
have
this
mechanisms
also
in
the
stratified
microphysics
of
the
model
and
do
a
global
run
instead
of
just
having
a
like
with
the
single
column
run.
So
we
have.
We
have
the
results
from
the
single
column
round,
so
it
would
be
good
to
have
from
the
global
run
yeah.
C
N
C
G
M
C
Cool
there
are
no
questions
rt,
please
just
keep
a
look
on
the
chat
unless
any
come
in,
and
then
we'll
move
on
to
our
final
speaker
before
our
discussion
for
the
first
half,
which
is
mike
pritchard
who's,
going
to
talk
about
renaissance
in
cloud,
super
parametrization,
emerging
flexibilities
for
regionalization
and
machine
learning,
emulation,
opening
new
csm
science
opportunities.
Very.
I,
like
the
title,
sorry
for
the
mouthful
yeah.
O
O
The
appeal
is
just
to
reduce
the
number
of
parameters,
the
number
of
interacting
parameterizations,
to
avoid
making
assumptions
about
things
that
are
hard
to
think
about.
Like
convective
organization
to
allow
microphysics
to
interact
more
naturally
with
the
explicit
updrafts
that
it
tends
to
be
formulated
to
to
interact
with
and
a
consequence,
is
a
lot
of
freedom
to
vary
the
gcm
vertical
resolution
in
ways
that
don't
produce
major
major
effects
and
vary
the
horizontal
resolution.
So
more
skill,
awareness.
O
So
this
approach
has
been
used
over
15
years
and
in
the
last
decade
it
mapped
out
a
lot
of
issues
about
the
storm,
resolving
era
ahead
of
schedule.
So
you
know
insights
in
how
the
modern,
julian
oscillation
works
and
its
climate
sensitivities
and
insights
into
how
explicitly
resolve
deep
convection
is
sensitive
to
warming,
including
regional
extremes
and
mesoscale
convective
systems.
O
Of
course,
now
global
storm
resolving
models
handle
that
more
elegantly,
but
only
for
those
who
have
access
to
huge
computational
resources,
and
I
want
to
suggest
that
this
next
decade,
more
flexible
mmfs,
can
democratize
that
global
storm
resolving
simulation
and
also
especially,
allow
us
to
begin
to
penetrate
the
sub-kilometer
turbulence,
frontier
ahead
of
schedule.
So
on
the
first
point
of
flexibility,
I'd
like
to
show
an
example,
this
is
from
a
postdoc
in
our
group
has
been
experimenting
with
with
making
a
regionally
flexible
super
parameterization.
O
So
normally
super
parametrization
is
one
size
fits
all
use.
The
same
cloud,
resolving
model
grid
structure
in
every
grid
cell,
but
liron
here
has
decided
to
use
the
normal
cloud.
Resolving
structure
that
resolves
storms
of
one
to
four
kilometer
resolution
in
all
the
white
regions
of
the
planet
and
in
these
blue
regions
he's
embedding
much
higher
resolution.
So
200
meter
horizontal
20
meter
vertical
in
order
to
resolve
marine
stratocumulus
better.
O
And
so
you
know,
if
you
try
to
do
that
globally,
it's
very
expensive.
Your
simulations
run
very
slow,
and
this
is
why
loren
worked
hard
to
try
to
fix
this
problem.
On
the
left-hand
panel
he's
showing
a
theory
he
developed
to
predict
the
optimal
processing
scale
as
a
function
of
the
workload
intensity.
O
You
can
unclog
this
load
imbalance
and
get
simulations
that
have
a
lot
faster,
throughput
and
less
expense
with
regionalized
physics.
So
the
right-hand
part
is
just
sort
of
there
on
validating
his
prediction
of
the
optimal
scale
against
measurements
of
the
csm
that
that
showed
that
it
works.
Now,
you
should
be
aware
that
there
could
be
problems
at
the
artificial
boundaries
you've
introduced
between
regions
with
different
grid
resolutions
in
the
embedded
model
and
the
right
hand.
Plots
here
are
meant
to
convince
you
that
actually,
these
are
not
large
issues.
O
So
you
know
this
is
just
an
example
to
point
out
that
our
regionally
flexible
super
parametrization
is
possible.
That
I
think,
opens
the
door
for
lots
of
niche
impacts
or
dynamic
studies.
You
know
suppose
you
only
want
to
study
the
tropical
boundary
layer
in
high
fidelity
and
its
interactions
with
climate
dynamics.
You
could
super
parametrize,
just
the
tropical
rainforests
and
save
a
lot
of
computing
power
if
you'd
like
to
you
know,
study
the
effects
of
atlantic
cyclogenesis
on
the
eastern
coast
real
estate.
O
You
could
just
embed
very
high
resolution
large
storm
permitting
domains
along
the
corridor
of
cyclogenesis
yeah,
and
you
know
insert
your
interesting
question
here
that
deserves
regionally
variable
resolution.
So
my
own
personal
obsession
is
the
boundary
layer
and
the
sub
kilometer
physics,
which
should
keep
us
night
right
we
up
at
night.
We
know
that
mixing
between
the
ppl
and
free
troposphere
is
really
important
to
equilibrium,
climate
sensitivity
and
that
difficult
to
parametrize
processes
in
the
boundary
layer
mediate
aerosol
cloud
formation.
That
is
at
the
heart
of
our
largest
uncertainties
in
climate
prediction.
O
F
O
You
can
see
a
pretty
big.
You
know:
five,
like
five
grams
per
meter,
squared
on
average
increase
with
not
much
seasonal
variation
and
the
blue
dots
show
you
what
happens
when
you
refine
the
resolution
at
great
computational
expense
of
each
of
the
models,
and
it
looks
on
the
one
hand
like
not
that
dissimilar
a
liquid
response,
but
behind
the
circles,
if
you
look
at
the
triangles,
you
can
see
a
regime
dependence
emerging.
That's
quite
important.
O
Actually,
the
upward
facing
triangles
separate
the
non-reading
clouds
from
the
reigning
clouds
and
they
have
an
opposite
sign
response,
so
they
actually
decrease
their
liquid
in
response
to
aerosol
loading,
and
it
makes
sense
that
that
you
know
non-reading
clouds
don't
participate
in
lifetime
effects,
and
this
is
a
feature
that's
observed
in
in
observations
and
in
large
eddy
simulations.
So
it's
nice
to
see
you
can
make
a
global
a
global
model
produce
when
you
add
explicit
boundary
layer
turbulence.
O
Now,
of
course,
there's
problems
with
these
approaches.
There's
biases
that
have
been
difficult
to
control
and
one
of
our
long-standing
ones
has
been
a
dim
bias
in
the
stratocumulus
zones,
which
is
not
what
you
intend
to
do
when
you
add
high
resolution
to
an
mmf.
But
I
think
a
lot
of
these
problems
are
just
associated
with
not
understanding
how
to
tune
multi-scale
models
well
and
there's
progress
on
that
front.
So
here's
an
example
of
some
work
where
we've
just
added
some
hyper
viscosity
to
the
cloud
resolving
model.
O
It's
a
predictable
consequence
of
the
leapfrog
momentum
solver
in
the
cloud
resolving
model
that
we
use
and
so
for
the
same
reasons
that
we
apply:
skill,
selective,
hyper
viscosity
to
gcm,
to
gcms,
to
avoid
good
scale
pilot
of
noise.
There's
there's
philosophical
reasons
to
do
the
same
thing
in
the
cloud
or
solving
models.
Once
you
begin
to
not
rely
on
grid
scale,
eddies
to
provide
your
your
enthalpy
transport-
and
you
know
various
tunings
like
these
can
fix
the
dim
biases
in
the
stratocumulus.
O
Here's
an
example
showing
the
brightening
of
the
clouds
that
happens
in
response
to
this
hyper
viscosity
tuning,
and
there
are
other
insights
being
drawn
about
how
to
control
mmf
rainfall
biases
in
the
department
of
energy's
latest
experiments
with
with
mmfs.
Of
course,
their
resolution
internally
and
higher
resolution.
Externally.
O
O
You
know
we'd
rather
do
things
that
approach
real
less,
that
are,
you
know,
three-dimensional
and
closer
to
50
or
sub
50
meter,
horizontal
resolution
to
get
increasingly
explicit
and
credible
with
boundary
layer
turbulence
and
low
cloud
feedback,
and
I'd
like
to
suggest
that
raw
computing
power
isn't
the
main
issue.
There
are
incredible
clusters
proliferating
nationally.
This
is
the
latest
one.
I
love
at
texas,
frontera.
You
know
over
8,
000,
really
high
quality
intel
nodes
and
journalists.
There
will
tell
you
they're
surprised
by
how
few
climate
applications
apply.
O
What
fraction
of
this
machine
is
occupied
by
non-climate
science
research,
so
there's
plenty
of
compute
cycles
to
exploit,
but
the
actual
potential
of
turbulence
permitting
mmf
has
never
been
explored
at
parity
with
global
cloud,
resolving
models
which
can
run
easily
on
a
hundred
thousand
course.
You
know
the
two
degree
mmf
might
have
14
000
embedded
las.
The
maximum
cores
I
can
use
is
fourteen
thousand
and
there's
inefficiencies
that
mean
I
frequently
will
use
one
to
six
thousand.
O
So
the
other
thing
to
look
ahead
to
that
is
why
I'm
still
interested
in
mmfs
is
the
exciting
promise
of
machine
learning.
Why
is
it
exciting?
Because
well
we
have
this
trade
off
with
super
computers
and
convection
permitting
resolution,
and
it
normally
assumes
the
job
is
really
hard
that
the
mission
is
stimulating
the
whole
atmosphere
for
decades.
O
So,
of
course,
we
can't
have
satisfying
3d
turbulence
calculations,
and
I
have
to
make
approximations
like
this
in
an
mmf,
but
in
the
machine
learning
era
the
job
could
change
they're,
just
making
short
simulations
for
the
purpose
of
mining.
Their
essence,
like
pier
shinteen,
was
showing
us
as
possible
to
train
machine
learning
emulators
and
if
that's
the
case,
then
we
can
already
do
much
more
justice
to
turbulence
physics
today.
So
that's
why
there's
breakthrough
potential
possible
there?
O
There's
no
no
trade-off
with
the
fit
accuracy
and
then
lately
we've
been
looking
at
climate
invariant
renormalizations
to
shield
emulators
from
out
out-of-sample
extrapolation,
those
papers-
you
can
read
about
there,
but
this
is
a
really
rapidly
moving
research
frontier,
and
I
want
to
point
out
that
you
know
there
are
some
shiny
results
that
have
been
amplified
in
the
literature
like
this
one
in
the
middle
panel.
That's
a
neural
network
powered
version
of
cam,
making
the
same
equatorial
wave
spectrum
as
the
sp
cam,
that's
much
more
expensive
to
run
on
the
left.
O
But
more
often
the
case.
What
you
see
are
explosions
like
this,
that
the
emulators
are.
They
have
imperfect
emulation
causes
blow
ups
with
the
large
degrees
of
freedom
available
once
coupled
to
a
host
model
and
controlling
these
explosions
is
an
active
research
frontier
so
and
again,
mmfs
provide
this
perfect
testbed
for
that
and
there's
lots
of
lessons
emerging,
I'm
confident
that
there's
going
to
be
a
step
change
in
this.
C
O
Yeah
for
sure
yeah,
so
the
the
mmf
was
really
investigated
during
an
nsf
science
and
technology
center
called
cmap
the
center
for
multi-scale
modeling
of
atmospheric
processes
that
went
from
2004
to
2014
and
as
part
of
its
legacy.
Super
parameterization
was
ported
into
the
csm
and
has
been
maintained
as
a
physics
option
ever
since
so
anybody
can
turn
on
the
super.
Parametrization
switch
in
modern
versions
of
the
csm
and
that's
part
of
why
the
community
can
can
do
machine
learning,
emulation
of
super
parameterization,
quite
quite
naturally,
with
an
open
source
code.
K
Well,
I
guess
I'm
I'm
curious,
because
the
the
sp
cam
prioritization,
while
it's
been
maintained,
hasn't
been.
We
have
not
run
any
scientific
checks
on
it,
since
esm,
1.1.1
and
the
code
hasn't
changed.
So
I
don't
know
what
it
does
scientifically
right
now
when
you
run
it
because
the
model
underneath
it
has
changed
quite
a
bit,
and
so
I'm
just
wondering
how
is
that
sp
cam
being
validated.
O
Yeah,
that's
a
great
question
and-
and
it's
good
to
know
it's
not
being
internally
validated,
you
know.
So
I
guess
it's
it's
the
distributed
community,
that's
running
it,
that's
validating
it
against!
You
know
I
I
can
say
that
I've
run
multiple
versions
of
it
and
we
have
to
do
microphysical
tunings
of
the
sort
that
are
done
internally
for
csm
releases
in
order
to
get
good
validation
against
top
of
atmosphere
and
radiative
constraints,
etc.
O
C
So
I
think,
maybe
related
to
that
as
peter
had
a
question
actually
in
the
chat
which
is
related
to
what
die
car.
What
die
core
are
you
specifically
using
in
the
in
your
experience
of
csm
sp,
but,
but
I
think
probably
because
that
I
think
now
you
can
see
the
questions
are
starting
to
get
tuned
towards
our
discussion
later
in
after
the
second
set,
which
is
really
focused
on
the
the
survey
results.
O
Sure
yeah
I've
used
I've.
You
know
he's
been
using
sp
for
so
many
years
that
I've
experienced
it
with
the
old
semi
lagrangian
spectral
dicor
and
with
the
finite
volume
dicor
and
with
the
spectral
element
dykor
to
my
view
that
the
dicor
doesn't
matter
so
much
compared
to
the
sp
interaction.
I
know
that
there
have
been
concerns
about
some
grid
scale.
Speckle
that's
been
emerged,
especially
with
the
with
the
semi,
with
the
spectral
element
dipole.
O
If
you
don't
use
a
physics
grid
remapping
that
can
be
amplified
by
the
mmf,
but
the
emergent
phenomena
don't
seem
to
care
about
that
as
much
yeah.
I
don't
know
if
I'm
speaking
to
your
interest
peter,
but
for
a
while,
I
was
really
worried
that
our
over-entrainment
problem
in
the
in
the
high-resolution
mmf
came
from
the
dicor,
but
you
know
even
large.
Variations
in
the
hyper
viscosity
of
the
fe
dicor
disproves
that
but
yeah
it
should
be.
Since
it's
a
physics
option,
it
should
play
nice
with
any
dicor.
C
Yeah,
I
think
that
that's
partially,
that
probably
answers
his
question
in
part
because
yeah
he
was
wondering
about
the
the
issues
with
the
sc
grid,
but
also
with
the
and
using
the
physics
grid,
but
also
the
implication
of
if
there
are
dicors
that
are
going
to
be
sunset
within
the
development.
You
know,
with
the
continuing
going
forward,
version
doesn't
mean
that
you
can't
download
cesm,
1
or
csm
2.1
and
and
use
them
if
that
would
be
an
issue
for
for
the
super
famorization.
B
Well
sure,
I'll
just
ask
real
quickly:
you
you,
you
must,
for
many
applications,
still
need
a
boundary
layer
scheme
right.
You
need
a
vertical,
a
vertical
diffusion
scheme
of
some
kind.
What
what?
What
is
what
is
sp
cam
using
now,
and
can
you
know
if
you're,
if
you're
doing
these
2d
lanes,
can
you
really
ever
get
away
from
needing
to
parameterize
kind
of
turbulent
processes?
Somehow.
O
That's
a
great
question
yeah.
So
when
we
use
course
resolution
sp,
the
boundary
layer
scheme
is
partly
subsumed,
so
the
vertical
diffusion
of
scalars
is
bypassed
and
a
spike.
The
the
land
model
instead
passes
a
spike
of
dry,
static,
energy
and
moisture
to
the
actual
battery
solving
model
which
naturally
diffuses
them
in
part
through
its
own
subgrid
scale
scheme,
which
is
a
you
know
like
a
one
and
a
half
order:
tke
closures,
magritzki
style,
or
if
you
choose
a
club
based
scheme,
there
are
other
ways
to
cope
with
the
two-dimensionality
problem.
O
Although
I
don't
think
that
features
exist
in
the
csm
version
of
sp,
I
personally
think
that
just
high
resolution
mmf
is
the
way
to
go
to
to
challenge
ourselves
to
regionalize
the
high
resolution
to
where
we
really
care
about
it,
for
small
process
studies
and
to
to
just
bypass
these
assumptions
by
using
three-dimensional
domains
that
have
enough
vertical
and
horizontal
resolution
to
do
quasi-explicit
boundary
layer
dynamics
that
aren't
relying
as
much
on
their
circuit
skills
team.
But
you
know
I'm
biased,
that's
my
interest
and
you
can't
do
that
globally,
but
then
did.
B
O
I
mean
there's
csm
as
a
grassroots
community,
with
all
sorts
of
different
interests
right
and
yeah,
but
that's
one
limit.
That's
really
interesting
that
I
think
is
hard
to
do
with
any
other
model
modeling
framework,
but
you
can
really
bypass
boundary
layer
schemes
now,
for
instance,
if
you
want
to
study
the
boundary
layer
responds
to
tropical
deforestation
and
its
global
climate.
Feedbacks,
including
ocean
coupled
feedbacks-
and
this
is
where
having
sp
within
a
model-
that's
got
an
actively
developing
ocean
component
is
so
vital.
C
Thank
you.
So
I'm
wondering
if,
if
well
mike,
if
you
don't
mind,
stop
sharing
your
screen
and
then
we
have
about
10
minutes
left
of
the
full
discussion
and
of
course
we
can
continue
this
discussion.
But
I
wanted
to
open
it
up
also
to
to
questions
more
broadly
related
to
the
the
talks
and-
and
some
of
this
discussion
is
setting
up
our
discussion
later,
which
is
which
is
useful
as
well,
but
springs
up.
C
Well,
brian
medeiros
brought
up
a
good
question
in
the
chat
which
isn't
just
true
for
the
maybe
mmf
frameworks.
That's
also
true
for
any
type
of
regional
refinement
approaches.
Does
anybody
have
thoughts
of
whether
there's
a
good
kind
of,
or
should
we
be
creating
some
type
of
guidelines
for
load
balancing
when
you're
setting
up
different
domains
with
with
high
resolution,
whether
it's
mmf
or
whether
it's
actually
using
the
vr
capabilities
of,
for
example,
the
spectral
element.core
of
how
to
handle
you?
I
mean
right.
C
So
one
of
the
things
here
is
when
we
start
to
develop
a
workhorse
model
that
that
maybe
has
a
lot
more
functionality
in
terms
of
focusing
domains
on
your
science
question,
which
I
think
is
partially
what
mike
was
advocating
towards
there.
It
also
means
that
the
the
balance
or
the
the
needs
from
the
the
support
team
and
the
software
engineering
team
might
come
in
because
there
is
these
big
questions
about
potentially
load
balancing,
and
so
I
wonder
if
anybody
wants
to
talk
to
see
peter
you've
raised
your
hand.
C
G
G
O
That's
great
to
know
about
yeah.
I've
wondered
about
this
and
I'd
love
to
talk
to
you
offline
more
about
it,
but
I
guess
it
speaks
to
the
fact
that
a
lot
too
much
of
this
sp
development
work,
including
my
own,
has
happened
in
silos
that
aren't
integrated
the
way
they
should
be
with
the
main
development
branches
of
csm.
O
Yes,
I
think
there
are
you
know
and
they
would
probably
sort
themselves
on
an
axis
of
computational
intensity
and
I
wonder
if
one
should
prioritize
the
lower
intensity
ones
that
would
be
more
accessible
to
a
broader
community
or
not
kill
the
end
card
compute
budget,
even
with
directio
yeah
but
yeah.
I
think
there's
natural
ways
to
sort
by
by
microphysical
complexity
by
by
resolution
there's
been,
you
know,
validation
of
one
to
four
kilometer
physics,
validation
of
200,
meter
physics.
O
They
have
different
tuning
demands,
but
these
are
separate
tuning
exercises
and
none
of
these
models
have
been
exhaustively
tuned
and
it's
probably
beyond
our
capability
to
tune
them
exhaustively,
and
so
they
should
be.
They
should
come
with
a
warning
and
they
should
yeah,
but
I
think
guidelines
would
be
would
be
helpful.
F
C
C
You
know,
maybe
a
narrow
set
of
workhorse
models
that
we
or
you
know
the
next
versions
that
are
supported.
I
think,
having
more
information
on
how
individuals
can
do
some
high-level
tuning,
I
think,
could
be
helpful
and
I
think
that
gets
at
the
question
that
it
might
just
answer.
So
brian
did
you
wanna
yeah,.
F
G
I
I
This
is
not
what
we
have
right
now,
because
we
have
had
to
sort
of
establish
priorities.
So
there's
a
lot
of
stuff
we
haven't
done
yet,
but
our
priority
has
been
in
the
last
six
months
to
actually
be
able
to
show
that
you
can
produce
a
climate
that
is
realistic
compared
to
control
versions.
Without
ever
calling
an
explicit,
deep,
convection
formulation,
now
we're
going
to
add
a
lot
of
these
extra
complexities.
I
So
so
that
sort
of
formulations
we've
tested
as
well
in
single
column
model
and
we
haven't
implemented
any
of
that,
but
from
a
theoretical
perspective,
that's
actually
fairly
plausible
and
defendable.
So
so
we
hope
to
do
that
and
we
plan
to
do
that
when
when
it
gets
to
that,
but
we
still
need
to
add
other
things
as
well
as
as
I
mentioned,
so
I
I
hope
it
wasn't
too
long
of
a
of
an
answer,
but
the
answer
is
yes:
we
can
adapt
to
high
resolution.
C
So
I
I
we
don't
have
a
much
of
a
break,
so
I'm
going
to
cut
off
the
discussion.
Sorry
sorry
adam,
I
saw
you,
maybe
came
out
to
add
something,
but
maybe
what
I
I'd
like
to
just
suggest
that,
as
we
go
into
the
discussion
after
the
next
series
of
talks.
So
after
that
we
have
a
10-minute
break
until
12
30,
eastern
10,
30
mountain
time
and
we'll
come
back
and
we'll
hear
a
little
bit
more
about
some
of
the
science
applications
of
the
atmospheric
model.
C
But
we
can
come
back
to
some
of
these
discussions,
particularly
related
to
the
development
and
future
development
of
cam.
In
our
open
discussion
later
on,
as
well
as
maybe
have
a
discussion
about
right,
we've
seen
some
promising
advances
and
I
think
it
would
be
nice
to
have
a
discussion
about.
You
know
how
to
how
do
some
of
these
advances
make
their
way
into
whether
it's
cam,
seven
or
or
maybe
the
cam
seven
point
x.
I
think
that's
a
good
discussion
to
have
in
our
in
our
fuller
breakout
later
so
with
that.
G
P
P
Q
Great
all
right,
thank
you
so
much
yeah,
so
today,
I'll
be
presenting
on
some
work.
I've
done
alongside
my
advisor
kevin,
that's
kind
of
the
part
of
the
last
part
of
my
phd
dissertation,
which
I
just
finished
up
this
past
may
so.
First,
when
you're
thinking
about
the
relationship
between
extreme
precipitation
and
temperature,
you
can
kind
of
think
of
think
about
it
on
like
two
different
time
scales.
So
there's
like
the
short
kind
of
like
weather
time
scale
where
the
relationship
it
can
depend
on
storm
type
season,
localized
convective
feedbacks.
Q
I
think
this
is
was
from
a
location
in
the
netherlands
or
something
the
the
rate
of
of
increase
of
precip
with
temperature
change
is
well
above
the
seven
percent
per
kelvin
baseline,
which
is
shown
as
that
dotted
like
black
line
and
then
there's
also
besides
the
short-term
relationship,
which
is
also
called
the
apparent
scaling
of
precipitation
and
temperature,
then
you
can
also
think
on
like
the
longer
term
climate
scale,
so
this
is
also
called
the
climate
scaling.
So
on
the
long
term,
as
the
atmosphere
and
the
sea
surface,
temperatures
are
warming.
Q
How
is
precipitation
responding
to
that,
and
these
two
scalings
are
not
necessarily
the
same.
So
then
there's
a
whole
other
aspect
we
bring
in
here.
So
looking
at
tropical
cyclone,
extreme
precipitation,
and
so
one
problem
in
look
looking
at
observations
is
that
in
the
past
over
tropical
oceans,
you
know,
observations
have
been
kind
of
spatially,
sparse
and
not
very
frequent.
So
therefore,
it's
tough
to
estimate
the
apparent
scaling
and
then
we
also
only
have
like
you
know,
20
years
at
most
of
more
high
resolution
observations.
Q
So
therefore,
it's
also
tough
to
estimate
climate
scaling
from
this
kind
of
short
time
frame
and
the
climate
scaling
is
probably
what
I'd
say
you
know
most.
The
most
interest
has
been
in
understanding
how
tc
precip
will
will
respond
to
climate
warming,
so
most
studies
just
kind
of
compare
a
historical
and
a
future
gcm
simulation.
Q
So
these
idealized
simulations
are
again
cam5
simulations.
They
were
run
with
the
official
radiative
convective
equilibrium
model
intercomparison
project
comp
set,
which
was
detailed
in
kevin
and
co-author's
paper
from
late
last
year,
and
so
we
use
the
protocols
of
the
this
rce
mip
except
one
change
is,
we
add
rotation,
so
the
model
is
rotating
at
the
same
rate
as
the
real
earth,
so
we
have
11
simulations
with
global,
uniform
sst,
varying
between
295
and
305
kelvin
in
one
kelvin
increments.
Q
The
horizontal
grid
spacing
is
about
28
kilometers.
We
use
the
sc
dynamical
core,
and
so
since
these
are
rc
simulations,
there's
no
seasonal
or
diurnal
cycles,
and
if
you
want
more
information
about
these
simulations
or
the
tcs
and
the
assimilations,
we
I
published
this
paper
back
in
2021
about
these
simulations.
Q
So
then,
for
more,
like
the
realistic
model
simulations,
we
also
use
a
cam,
some
cam5
simulations,
their
amip
simulations
for
1980
to
2012..
They
have
the
same
horizontal
grid,
spacing
of
about
28
kilometers.
Also,
that's
the
dynamical
core,
and
so
these
simulations
were
also
used
to
study
tcs
in
these
couple
of
of
previous
papers.
So
they
have
been
used
to
study
tcs
before
and
I'll
just
mention
that
to
track
these
tc's
and
extract
their
precipitation.
We
use
the
tempest
extreme
software,
so
finally
observations
so
I'll.
Q
Q
And
I'll
just
mention
quickly,
so
we
only
look
at
tcs
that
are
between
5
and
20
degrees
latitude
in
both
hemispheres.
Just
because
we
wanted
to
try
to
focus
on
the
region
where
the
idealized
simulation
tcs
would
most
likely
to
be
are
most
likely
to
be
comparable
to
real
world
tropical
cyclones
and
so
to
get
into
a
couple
of
results.
So
looking
at
tc
intensity,
we
have
the
rce
simulations.
The
idealized
simulations
are
black.
Q
Observations
are
blue
cam,
5
amip
simulations
are
this
purple
color,
and
so
we
see
that
actually,
the
intensity
distributions
for
the
models
are
pretty
comparable
to
observations
not
too
far
off.
The
medians
are
shown
on
this,
the
x-axis
and
the
diamonds,
and
so
for
both
measures
of
tc
intensity,
the
maximum
wind
speed
and
the
minimum
sea
level
pressure.
The
models
underestimate
the
intensities
a
little
bit,
but
overall
really
not
too
bad.
Q
Next,
we
look
at
tc,
precipitation,
radial
profiles
and
so
on.
The
x-axis
here
just
shows
the
distance
from
the
tc
center
and
the
y-axis
is
showing
the
precipitation
rate
in
millimeters
per
day
for
rce
cam
5
observations,
and
so
we
see
when
we're
looking
at
the
values
of
the
precipitation
in
these
radial
profiles,
they're,
actually
pretty
comparable
between
all
the
data
sets.
Q
Km5
is
actually
higher
than
observations
in
some
cases,
so
I
forgot
to
mention
these
are
split
up
by
sea
surface
temperature
going
from
299
kelvin
to
303
kelvin,
and
so
again
we
see
the
values
between
the
data
sets
are
actually
fairly
similar.
However,
the
response
of
the
profiles
to
sst
increases
does
vary
a
bit
so
for
the
rc
simulations.
Q
We
see
that
most
of
the
increase
with
sst
is
in
the
inner
core
of
the
storm
for
the
cam5
simulations
in
the
middle
there's,
an
increase
in
the
inner
core
and
actually
not
so
much
in
kind
of
the
middle
region
and
a
little
more
in
the
outer
region
of
the
storm
and
then
for
observations.
There's
a
little
bit
more
of
like
a
blanket
increase
over
most
of
the
radii
as
the
sst
increases.
Q
So
next
I
kind
of
tried
to
recreate
that
plot.
I
showed
in
this
the
second
slide,
so
looking
at
the
relationship
between
sst
on
the
x-axis
and
precip
on
the
y-axis.
So
here
I'm
showing
I
take
each
tc
at
each
time
step
and
I
calculate
its
99th
percentile
precipitation,
so
an
extreme
precipitation
metric
and
then
I
bin
the
storms
based
on
the
sst
that
is
underneath
them,
and
so
the
color
scheme
is
consistent.
Q
Q
Q
And
so
what
we
found
was
that,
looking
in
the
in
the
black,
the
black
column
is
the
total
kind
of
change
in
99th,
percentile
tc,
precipitation
per
kelvin.
So
again
we
see
consistent
with
the
figure.
I
showed
two
slides
ago:
the
rce
only
shows
a
6.1
percentage
increase
per
per
kelvin,
whereas
cam
5
and
observations
are
a
bit
higher
with
observations
being
the
highest
at
about
15
and
so
for
all
of
them.
The
thermodynamic
contribution
dominates
over
this
intensity
contribution
so
for
cam,
5
and
observations.
Q
The
intensity
contribution
is
actually
around
zero,
and
this
is
mostly
due
to
this
last
term
shown
in
this
equation.
This
dv
dsst
term,
because
we
found
that
that
relationship
was
like
very
small,
even
slightly
negative
and
how
we
were
calculating
it,
so
that
kind
of
made
the
intensity
contribution
fairly
small,
but
it's
higher
in
the
rce
simulations
so
just
quickly
I'll
go
over
some
results
so
again
for
this
kind
of
apparent
short
duration
scaling,
it
seems
that
tc
precipitation
is
a
bit
less
sensitive
to
sst
in
the
models
than
in
the
observations.
Q
But
this
is
especially
apparent
for
the
rce
stimulations
for
all
of
the
data
sets.
We
found
that
the
majority
of
precip
change
comes
from
the
thermodynamic
contribution
and
again
be
in
cam
5
and
observations,
because
the
tc
intensity
change
with
sst
is
like
very
small.
That
leads
to
the
very
small
contributions
from
intensity
change
and
so
we're
still
working
on
trying
to
explore
how
we
could
possibly
use
these
rc
simulations
to
try
to
understand
the
climate
scaling
between
tc,
precipitation
and
sst
and
trying
related
to
observations.
Q
This
is
still
a
work
in
progress.
I
was
literally
like
messaging
with
kevin
this
morning
about
this
work
so
still
working
on
it,
but
ultimately,
I'm
trying
to
see
if
we
could
try
to
use
these
rc
simulations
to
like
project
future
tc
precipitation
increases
as
we
continue
to
see
sea
surface
temperatures,
warming
with
climate
change,
so
that's
all
I
have
for
today
again.
This
is
this
paper
that
came
out
about
the
rce
simulations.
Q
P
G
G
A
hundred
percent,
but
for
rc
and
offs
it's
a
little
lower.
What
are
some
of
the
thoughts
for
the
ops
part
of
like
where
that
principle
increase
is
coming
from,
like
aside
from
the
thermodynamic
contribution,
because
that
I'm
asking,
because
that
might
relate
to
like
how
people
want
to
interpret
the
idealized
simulations
so
just
wondering
what
your
thoughts
might
be.
Q
Yeah,
no,
that's
a
really
good
question.
I
think
I'm
still
yeah
trying
to
understand
this.
Exactly.
I
think
there
probably
is
like
at
least
on
the
climate
scale,
there's
probably
a
more
of
an
increase
from
intensity
that
is
not
showing
up
in
these
results
because
of
how
I
calculated
this
dv
dsst
term.
So
I
think,
probably
in
the
long
term
there
would
be
yeah
a
contribution
from
intensity.
Q
I
guess
there
probably
could
also
be
some
sort
of
maybe
some
micro
physical
changes
that
are
showing
up
in
the
observations
that
maybe
wouldn't
in
the
models
as
well.
So
that's
one
thought
yeah,
maybe
like
precipitation
efficiency
changes,
perhaps
just
a
guess
but
yeah.
I
think
I
need
to
think
about
this.
A
little
more.
It's
a
good
question.
P
P
P
P
N
N
N
I
would
like
to
show
the
recent
ipcc
report
about
tropical
cyclones
in
future
warmer
climate,
and
there
is
indication
that
there
are
more
major
tropical
cyclones
and
as
well
as
more
better
and
windier
tropical
storms,
and
also
they
are
moving
shifting
towards
north,
as
well
as
their
rapidly
intensifying
tropical
cyclones
due
to
increased
surface
temperatures
as
well
as
there
is
also
indication
of
the
slower
dc's
which
can
do
more
damage
and
coming
to
the
d
season
climate.
In
the
warmer
climate.
N
There
is
an
in
consensus
of
tropicals
increase
in
tropical
cyclone
intensity
between
models
and
as
well
as
the
theory
but
coming
to
the
frequency.
There
is
an
uncertainty
because
some
of
the
studies
project
an
increase,
and
there
are
some
studies-
project,
a
decrease
in
the
tropical
cyclone
frequency
in
the
warmer
climate
and
therefore
understanding
the
last
climate
controls
in
tropical
cyclone
frequency.
N
Coming
to
the
past
climate,
I
have
considered
two
simulations
which
are
due
to
changes
in
the
precision
index
like
one
is
5e
where
there
is
the
low
precision
index.
That
means
that
there
is
high
solar
insulation
in
the
northern
hemisphere
summer
and
the
nmis
5d,
which
has
high
precision
index.
That
means
there
is
high
solar
insulation
in
the
southern
hemisphere,
and
these
changes
in
the
solar
installation
may
lead
to
changes
in
the
last
climate
variables
and
and
the
future
climate.
N
N
One
is
the
traditional
tracking
schemes
like
which
uses
resolution
dependent
thresholds
to
detect
tropical
cyclone
like
vertices,
like
those
thresholds
and
the
surface
pressure,
absolute
verticity
and
metal
in
speeds
and
duration
thresholds,
and
the
second
one
is
the
phenomenon-based
tracking
scheme
which
usually
detects
the
location's
potential
for
cyclone
formations.
They
use
the
resolution.
Independent
thresholds
like
those
are
objects,
parameter
related
humidity,
vertical,
wind,
shear,
specific
humidity
and
the
duration
thresholds.
N
And
yeah
this
slide
shows
the
changes
in
the
large
scale.
Variables
like
the
left
side
shows.
The
the
first
is
the
precision
index
and
the
c
water
levels
and
the
tropical
ssds.
These
are
simulated
using
low
resolution
csm
model,
and
you
can
see
the
variations
of
the
co2
levels
match
with
the
tropical
assistance
and
like
the
past
warmer
climate,
like
the
past.
Conditions
like
mis5e
can
be
like
can
be
an
analog
for
the
future
warmer
climate.
N
So
we
see
an
asymmetry
in
northern
hemisphere
summer
and
southern
has
been
summer
seasons
of
like
the
temperature,
specific
material
related
humidity,
and
these
changes
may
lead
to
the
changes
in
the
tropical
cyclone
and
genesis
our
intensity.
Variations
like
to
look
at
that.
I
used
genesis
potential
index
like
to
look
at
the
tropical
cyclone
genesis
and
the
left
panel
shows
the
compares
between
two
orbital
forcing
simulations
and
the
the
other
panel
shows
the
two
times
co2
and
the
present
compared
with
the
present
present-day
climate.
N
We
see
a
overall
decrease
in
the
frequency
of
storms
in
across
the
globe,
and
these
differences
in
the
genesis
potential
are
can
be
explained
by
use
of
like
using
the
component
analysis.
The
gpa
component
analysis
to
see
which
variable
is
able
to
explain
the
changes
in
the
genesis
potential
index
and
of
all
the
variables.
I
can
see
that
the
moisture
entropy
deficient
variable
is
explaining
the
changes
of
genesis
in
both
the
orbital
force
incinerations
as
well
and
in
the
future
warmer
two-time
co2
simulations.
N
I
have
used
another
genesis
potential
index
as
well,
which
uses
different
variables
like
here
instead
of
most
entropy
deficient.
The
additional
variables
are
the
related
humidity
and
the
vertical
velocity,
and
these
genesis
potential
changes
also
indicate
that
here
also
on
the
700
hectare.
Pascal
related
humidity
explains
the
changes
in
the
genesis.
N
Tropical
cyclone
genesis
changes,
and
this
slide
shows
the
the
tropical
cyclone
genesis
and
tracks
which
are
explicitly
simulated
using
the
tropical
cyclone
tracking
scheme,
which
I
have
explained
earlier,
and
here
we
can
see
that
in
the
mis5e
there
is
higher
annual
frequency
compared
to
the
present
day
and
on
two
times
co2.
We
observe
a
decrease
in
the
frequency
of
formations
compared
to
the
present
day.
N
We
can
see
that
there
is
a
symmetry
in
the
orbital,
forcing
simulations
like
increase
in
the
southern
hemisphere
basis
or
and
a
symmetry
in
the
northern
hemisphere,
like
a
decrease
in
the
western
pacific
and
in
a
north
atlantic
basins,
an
increase
in
the
north,
indian
and
eastern
north
pacific
and
in
210
co2.
We
see
an
overall
decrease
in
the
across
all
the
ocean
basins.
N
We
see
an
increase
in
the
frequency
of
intense
storms
and
two
times
co2
simulations
when
coming
to
the
orbital
forcing
simulations
we
see
a
symmetry
in
the
northern
hemisphere
and
southern
hemisphere
in
northern
hemisphere.
We
see
a
decrease
in
the
frequency
of
intense
storms
in
amis
5v,
whereas
in
southern
hemisphere
we
see
an
increase
in
the
frequency
of
intense
storms
and
misfires
conditions
compared
to
them
as
5d,
and
these
changes
like
we
can
see
the
changes
in
the
tropospheric
temperature
and
moisture
conditions
like
here.
N
And
here
we
can
see
that
relative
humidity
changes,
explain
the
changes
in
the
tropical
cyclone
frequency
like
in
the
northern
southern
hemisphere
summer
season.
There
is
higher
relative
humidity,
which
is
leading
to
favorable
for
tropical
cyclone
formations
and
in
the
northern
hemisphere
summer
season.
There
is
a
decrease
in
the
relative
humidity
in
the
orbital,
forcing
simulations
which
is
leading
to
the
decrease
in
the
tropical
cyclone
frequency
and
looking
at
the
two
times,
co2
conditions.
N
We
see
that
globally,
there
is
a
decrease
in
the
relative
humidity,
which
is
leading
to
a
decrease
in
the
frequency
of
formations
and
also
I
have
looked
at
the
static
stability
of
the
atmosphere,
and
we
also
see
this
in
mis5e
the
so
northern
hemisphere
has
highest
stable
conditions
compared
to
the
mis5d,
which
is
also
leading
to
the
reduced
frequency
in
northern
hemisphere,
seasons
and
in
southern
hemisphere.
We
see
mis
5d
has
a
higher
stability
in
stable
conditions
than
minus
5v.
N
P
P
R
R
R
Our
grid
has
a
base
base
resolution
of
one
degree
all
over
the
globe,
but
our
on
our
domain
of
interest.
It
is
refined
to
a
quarter
degree,
high
resolution
and
sorry
between
the
high
resolution
region
and
the
base
resolution.
There
is
a
there
is
a
region
of
transition
which,
which
has
a
resolution
of
quite
a
half
degree.
R
When
we
want
to
compare
the
cost
of
variable
resolution
grids,
we
can
assume
linear
scalability
between
the
number
of
elements
within
the
grid
grid
and
a
job
o'clock
time.
So
we
can
say
that
our
mod,
our
variable
dissolution
grids,
has
as
less
than
two
times
the
cost
of
uniform
one
degree
spectral
element
grid,
but
it's
it's
likely
to
give
us
comparable
results
of
a
uniform
quarter.
Degree.
High
resolution
grids
over
our
domain
of
interest.
R
For
our
study.
We
made
two
we
completed.
Two
sets
of
simulations.
One
set
consists
of
control
simulations,
which
has
a
uniform
one
degree
resolution,
and
the
other
step
consists
of
a
variable
resolution
simulations
with
a
highest
resolution
of
a
quarter
degree.
We
made
our
simulations
using
a
fs
concept
for
for
between
1998
and
2015,
and
both
control
and
variable
resolution
simulations
consist
of
seven
ensemble
members.
R
R
These
are
for
a
total
press,
any
total
precipitation.
Again
we
obtain
a
globally.
We
obtain
comparable
results,
results
with
a
uniform
one
degree
control
simulation.
But
when
we
look
at
our
region,
we
can
see
better.
We
can
see
that
our
model
performed
better
than
the
control
simulation
again,
especially
over
a
complex
terrain.
R
R
R
P
P
F
R
Very
preliminary
results.
We
are
planning
to
make
comparisons
with
observation,
that
data
sets
to
study
on
many
variables
to
evaluate
its
added
value
because
of
time
restriction
and
because
our
results
are
quite
fresh,
and
these
are
just
for
two
variables,
but
we
are
planning
to
study
many
variables
and
make
many
observe
many
comparisons.
With
other
observation
data
sets.
P
And,
and
where
do
you
envision
this
research
in
the
future?
What's
really
your
science
pursuit
here?
So
I
couldn't
understand,
oh,
where,
where
do
you
want
to
go
in
the
future
with
that
approach
of
the
variable
resolution?
What's
kind
of
really
the
science
approach
that
that
you
want
to
follow
in
the
future?
P
R
B
B
R
B
G
P
Thanks
adam-
and
I
actually
also
would
like
to
point
out
that
who
we
won
from
doe
the
pnn
lab.
She
also
looks
a
lot
at
the
time
step
sensitivity
in
e3sm
in
this
case,
but
it
turns
out
yeah
that
a
lot
of
the
physics
parametrizations
are
very
sensitive
to
time,
step
settings
and
it's
often
really
grounded
in
the
numerics
of
these
physical
prioritizations
and
how
they
are
coupled
to
the
dicor.
So
it's
a
very
important
topic
adam.
I
agree
on
that
which
we
should
tackle
for
csm
as
well.
H
G
P
P
H
H
H
As
a
result
on
this
title
slide,
you'll
see
that
earthworks,
which
is
being
led
by
dave
randall
at
colorado
state,
has
a
very
strong
list
of
in-car
and
cesm
co-pis
and
collaborators,
and
also
we've
really
benefited
from
contributions
from
the
esmf
and
the
c-seg
teams
as
well.
So
in
terms
of
how
julio
started
in
terms
of
new
car
code
of
conduct,
this
is
very
much
a
team
effort
and
I'd
like
to
acknowledge
all
of
these
team
members.
H
In
this
very
short
talk.
I
don't
have
time
to
really
articulate
the
reasons
why
we
might
want
to
consider
high
resolution
global
climate
modeling.
I
think
many
people
in
this
working
group
are
very
familiar
with
that,
as
well
as
the
challenges
around
it.
This
is
just
one
paper
that
I
was
fortunate
enough
to
contribute
to.
This
was
a
briefing
paper
prepared
in
advance
of
cop26.
H
In
addition
to
all
of
the
interesting
scientific
opportunities
high
resolution
modeling
presents,
we
make
the
case
in
this
paper
that
high
resolution
earth
system
models
can
potentially
revolutionize
the
quality
of
information
available
from
mitigation
and
adaptation
decisions
in
the
face
of
climate
change,
on
both
global
and
regional
scales.
Since
the
extreme
weather
events
that
cause
the
greatest
damage
are
generated
by
processes
operating
at
small
scales,
while
crucially
they're
shaped
by
the
much
larger
global
circulation
of
the
ocean
and
the
atmosphere
and,
of
course,
there's
feedback
between
the
two.
So
it's
really
a
continuum
of
scales.
H
So,
just
briefly,
earthworks
is
a
five-year
project
funded
by
the
nsf
directorate
for
computer
and
information,
science
and
engineering.
The
objective
is
to
develop
a
global
storm.
Resolving
coupling
coupled
model
capability
that
uses
a
single,
uniform
global
grid
earthworks
consists
of
cam6
with
the
impasse
non-hydrostatic
dynamical
core
the
impasse
ocean
model
that
has
been
developed
at
los,
alamos
national
labs,
the
impasse
sea
ice
model,
as
well
as
the
community
land
model
and
cmeps
the
goal.
H
Ultimately,
where
we're
headed
we're
not
there
yet
we're
not
even
two
years
into
the
project,
but
indeed
is
a
you
know
a
kilometer
scale,
fully
coupled
modeling
capability,
a
lot
of
the
effort
is,
is
really
moving
toward
utilizing.
Gpu
enabled
components
with
end-to-end
workflow
portability
across
u.s
leadership,
computing
systems
and
I'll
say
a
bit
more
of
that.
In
this
talk
a
few
initial
objectives,
I
just
want
to
highlight
and
call
out
in
today's
presentation.
H
One
of
the
first
tasks
was
porting.
The
impasse
ocean
model
and
the
sea
ice
models
into
the
csm
framework,
and
we
have
done
that
successfully
initially,
while
we're
moving
toward
gpu
enabled
components,
we
wanted
to
assemble
a
cpu
version
of
this
earthworks
configuration
and
test
it
out
at
relatively
coarse
grid
resolutions
and
and
we've
achieved
both
of
these
goals.
We're
also
beginning
to
test
cam
6
physics
at
convection,
permitting
scales-
and
I
think
importantly,
and
of
interest
to
to
this
group
perhaps
is
that
we've
been
working
very
hard
to
port
critical
path.
H
Portions
of
earthworks
to
gpus,
and
we've
made
good
progress
on
that,
and
let
me
just
quickly
highlight
some
of
these.
So,
as
I
mentioned,
the
first
task
assemble
this.
This
version
of
the
model
get
it
to
run,
do
it
based
on
a
cpu
version
and
and
see
what
the
simulations
look
like,
and
can
we
produce
stable
climate
simulations
at
relatively
coarse
resolutions?
H
We
have
diagnostics
for
a
number
of
simulations,
including
a
10-year
simulation
that
was
recently
completed
at
60
kilometers,
this
available
from
the
earthworks
website,
which
is
indicated
here,
and
this
is
not
a
talk
to
talk
about
the
fidelity
of
the
simulation
at
this
point,
but
I
just
wanted
to
put
up
a
couple
of
plots
to
show
you
that
we
are
simulating
planet
earth.
This
is
surface
air
temperature.
An
estimate
of
the
observations
on
the
left
and
from
the
10
10
year
annual
mean
surface
air
temperature
from
earthworks
on
the
right.
H
There
are
biases,
I
don't
mean
to
say,
there's
not
biases,
some
relatively
large
biases.
At
this
stage.
You
can
see
with
your
naked
eye,
for
instance,
the
cold
bias
over
the
arctic
region.
If
you
look
off
the
west
coast
of
the
continents,
as
is
typical
for
a
couple
models,
earthworks
is
also
simulating
near
surface
air
temperatures
and
sea
surface
temperatures
that
are
too
warm
to
climatology.
H
We're
not
only
focused
on
mean
state
when
we're
looking
at
the
diagnostics,
but
also
the
variability
ten
years
is
kind
of
short
to
say
much
about
enso
variability.
But
here
we
just
made
a
plot
of
the
north
atlantic
oscillation,
based
on
monthly
anomalies
of
sea
level,
pressure
over
the
north
atlantic,
pretty
good
reproduction
of
this
dominant
mode
of
climate
variability
and
precipitation
people
always
like
to
see
plots
of
precipitation
an
estimate
from
observations
on
the
left
from
the
10-year
earthworks
simulation
again
annual
mean
on
the
right.
H
You
can
see
that
there
are
aspects
of
the
earthwork
simulation
that
match
up
with
observations
pretty
well,
but
again,
there's
biases,
pretty
diffuse
itcz
or
the
atlantic
too
much
rainfall
over
the
western
equatorial
indian
ocean
region
and
something
in
the
central
pacific.
That's
quite
reminiscent
of
the
long-standing
double
itcc
bias.
H
Very
recently,
in
fact.
Just
yesterday
we
got
some
plots
from
a
new
30
kilometer
coupled
simulation.
That
is
running
as
we
speak,
so
this
is
just
the
january
from
year,
one
nothing
more
than
that
at
30
kilometer
resolution,
the
total
precipitation
rate.
So
I
can't
really
compare
or
say
anything
scientifically
valid
at
this
point,
but
I
do
think
it's
interesting.
For
instance,
if
you
look
over
the
atlantic,
you
see
a
much
better
defined
itcz
or
atlantic.
H
Again,
I'm
not
saying
that's
a
robust
feature,
but
this
simulation
is
just
now
in
the
machine
when
we're
making
these
diagnostic
plots
that
are
available
on
the
web
page.
We
are
used
using
the
number
of
existing
diagnostic
packages,
such
as
the
climate,
variability
diagnostics
package
and
does
impass
analysis
package,
but
ultimately
using
impasse.
H
H
What
you
see
in
the
observations
is
a
classic
squall
line
structure
within
a
larger
scale
structure,
a
comma
shaped,
larger
scale
structure.
You
see
a
very
leading
clear,
leading
line
of
very
strong
convection
and
a
trailing
stratoform
structure
with
a
weak
echo
region
between
the
two.
Now,
when
you
look
at
the
impasse
simulation
based
on
warf
physics,
overall,
the
forecast
is
very
good.
H
Many
aspects
of
the
observations
are
reproduced,
although
you
will
note
that
at
three
kilometers
the
model
is
overdoing
the
reflectivity.
This
is
typical
of
all
models
of
three
kilometer
resolution
and
there's
no
significant
weak
echo
region.
When
we
look
at
the
simulations
with
impasse
with
cam
6
physics,
again,
we
see
the
overall
structure,
but
there
are
larger
differences
than
we
see
in
observations
compared
to
observations
with
convection
being
forced
primarily
by
the
strong
cold
front.
H
While
for,
for
example,
in
the
war
physics
simulation,
the
convection
is
triggered
in
a
warm
sector
ahead
of
the
front
more.
As
was
observed,
so
there's
some
takeaways
from
this.
The
key
thing
I
want
to
emphasize
today
is
we
are
doing
this
testing.
We
have
a
lot
more
to
do
in
this
area,
but
cam
six
physics
is
stable
at
three
kilometers,
at
least
for
this
test
case.
H
The
first
is
the
portability
of
the
environment
where,
in
earthworks
we
are
experimenting
with
software
containers
like
docker
and
singularity,
and
also
performance
portability,
where
we've
taken
a
gpu
offload
via
directives,
approach,
using
primarily
openacc
and
openmp,
and
also
beginning
to
experiment
with
mpi
if
you're
familiar
with
with
those
acronyms,
it's
also
important
to
demonstrate
scalable
end-to-end
workflows,
which
includes,
of
course,
scalable.
I
o,
which
is
really
critically
important
for
high
resolution,
modeling
efficient
analysis
because
we're
talking
about
hundreds
of
terabytes
of
data
here
on
unstructured
grids.
H
I
already
mentioned
how
we're
working
with
project
region
high
speed
point-to-point
data
transfers,
for
example,
via
globus
and
data
portals,
which
means
data
sharing
vis
via
easily
deployable
containerized
science
gateways
related
to
this.
Our
goal
is
to
build
and
test
across
multiple
platforms
and
centers
to
ensure
that
earthworks
is
easy
to
build
and
run
we're
initially
focused
on
a
subset
of
large
computing
platforms
on
nsf
and
boe
systems
at
nsf
of
with
nsf
systems.
This,
of
course
includes
in-car.
H
We
have
a
large
allocation
on
cheyenne
that
we're
utilizing
and
then
also
at
tac,
where
impasse
atmosphere
will
serve
as
a
characteristic
science,
application
for
nsf's
next
generation
leadership,
class
computing
facility
system,
which
is
tentatively
being
called
horizon
on
doe
systems.
We've
received
allocations
to
do
work
at
argonne,
which
we're
using
primarily
for
gpu,
porting
and
testing
and
at
nurse
where
we're
doing
both
cpu
and
gpu
testing.
H
Finally,
in
terms
of
the
testing
status
first
on
cpus,
the
initial
goal
of
earthworks
has
been
to
assemble
a
working
cpu
version,
as
I've
highlighted
earlier
in
the
talk,
this
plot
shows
our
progress
in
achieving
this
goal.
The
x-axis
is
the
level
of
test
integration
from
unit
tests
and
gray
to
increasing
levels
of
component
integration
within
cesm,
which
are
the
shades
of
blue
and
the
y-axis
is
the
resolution,
and
you
can
see
that
our
approach
here
is
to
increase
the
resolution
incrementally
by
a
factor
of
two.
H
The
key
point
is
that
we're
trying
to
work
up
toward
the
upper
right
corner
of
this
diagram
and,
overall
you
can
see
that
we're
making
good
progress
and
we've
completed
now.
A
couple
of
fully
coupled
runs
on
cpus
at
120
and
60
kilometers
and
again,
a
30
kilometer
simulation
is
in
the
machine
right
now
in
terms
of
gpu
testing.
This
chart
breaks
down
earthworks
by
components
that
need
to
be
refactored,
namely
the
atmosphere
of
the
ocean
and
the
sea
ice
components.
H
Note
at
this
point
we're
not
touching
the
land
surface
and
the
cmeps
mediator,
which
we
envisioned
saying
on
cpus
in
this
table.
The
green
shows
those
parts
that
have
been
ported
and
optimized
to
gpus,
noting
that
we've
only
looked
at
nvidia
gpus
today,
but
the
main
point
is
that
some
of
the
most
critical
path,
portions
of
the
model,
have
been
ported
to
gpus
we're
working
on
club
right
now,
and
when
this
supported,
we
will
also
be
able
to
have
the
cam6
physics
package
operating
on
gpus
wrapping
up
here.
H
This
talk
is
related
to
further
testing
of
cam6
physics
at
three
kilometer
grids,
using
variable
mess
resolutions
and
then
a
couple
of
talks
of
interest
in
the
software
engineering
group
tomorrow.
So
thank
you
very
much
for
your
attention.
I
think
this
is
a
summary
of
the
key
points
that
I've
mentioned.
So
I
won't
read
them
again.
H
P
S
S
So
in
this
short
presentation,
I'm
going
to
discuss
some
of
recent
work
as
jim
also
briefly
mentioned.
Analyzing
the
moisture
flux
transport
pattern
over
white,
serious
regions
using
the
csm
covered
with
the
non-hydrostatic
dynamical
that
is
impulse,
and
this
is
a
steel
primary
result.
So
the
key
of
this
work
is
built
up
built
upon
this
non-hydrostatic
csm
simulations
with
the
major
support
from
the
ancient
nsa,
funded
earthworks
project.
S
First,
I'd
like
to
briefly
introduce
the
background
and
on
the
main
questions
we'd
like
to
answer
for
this
study.
So
first,
this
work
is
connected
to
one
of
our
priests
work
that
is
focusing
on
evaluating
the
campus
performance
and
preservation
features,
and
we
did
the
cat.
We
did
a
case
study
over
the
avatar
over
the
west
u.s
u.s
regions.
S
Here
it
is
here,
are
the
great
match:
domains.
That's
during
the
transient
skills
from
the
60
kilometer
to
30
meter.
As
you
can
see
here,
the
sticker
meter
covers
the
the
hawaii
u.s
specimen
and
also
adjacent
ocean
region,
so
for
connect
to
foreign
by
this
evaluation.
Work
we'd
like
to
we'd
like
to
understand
a
large
scale.
S
Research,
I'd
like
to
bring
up
point
out
that,
due
to
this
distinct
different
dynamics
over
the
northwest
and
the
southwest
coastal
regions
for
this
ai,
driven
for
ai
driven
price
visions,
heavy
presentations,
I'd
like
to
focus
on
separating
the
analysis
for
these
two
on
two
different
of
the
regions
and
the
wind
dynamics.
Shield
here
is
for
the
near
surface
average
wind
wind
pattern.
S
So,
first,
first,
let's
look
at
the
thermodynamic
component
for
the
moisture
concentration
for
this
heavy
person
days.
When
I
say
the
heavy
president
is
I'm
using
the
75
percent
higher,
as
the
stretch
heard,
to
to
select
the
heavy,
the
pres
and
precision
days
above
this
value
so
for
northwest
division.
I
did.
I
averaged
the
original
mean
persuasion
for
for
for
the
land
on
domain
and
similar
way
for
california
region
and
for
the
heavy
present
days
it.
S
Here
is
the
here's
just
this
integrated
water,
vapor
transport
again
for
this
heavy
pressure
index
of
75
ounce,
entire
stress
heard,
and
this
because,
as
briefly
as
I
mentioned,
like
the
noisemaster
during
a
careful
range
on
there
on
this
first
happy
person,
this
associated
with
different
different
tracks
of
different
tracks
of
different
types
of
tracks,
ai
tracks
and
dynamics,
as
you
can
see
here
like
the
on
the
model,
actually
did
a
pretty
good
job.
When
we
averaged
the
order,
we
composite
the
result
of
all
these
on
high
refreshment.
S
This
it
had
this.
It
has
this
reasonably
special
distribution
for
the
moisture
flux,
transport
pattern
and
also
the
magnitude
is
pretty
close
to
marriage
too.
However,
we
do
see
some
underestimation
for
the
chem
impasse
compared
to
the
marriage
you
decide
and
also
because
we
have
much
final
resolution
for
chemin
pass,
which
is
nuclear
meter
versus
the
0.5
degree
in
marital.
So
we
do
see
like
we
do
see.
Some
final
features
in
campus
for
the
most
transported
pattern.
S
Analysis
average
over
this,
the
longitude
from
130
to
140,
watts,
degree
and
again.
This
first
law
is
a
result
for
the
northwest
region
and
it's
capturing
the
the
result
for
california
region,
and
we,
if
you
can,
if
you
can
take
first
look
at,
we
have
the
vertical
dash
lamp
dinners
the
latitude
bombs
roughly
for
each
of
this
study
domain.
S
First
here
is
the
dominant
donor
range
profile,
it's
similarly
as
this
specific
humidity,
but
we,
but
we
actually
increase
the
heat
levels
here,
just
to
want
to
show
the
overall
picture
of
the
wind
on
the
vertical
when
the
profile,
how
it
looks
like
in
a
moderator
and
how
to
compare
it
to
the
realistic
side,
and
we
did
recreate
it.
The
result
for
cameron
passed
to
the
same
resolution
as
the
marriage
tutor
said,
for
so
for
the
northwest
region
has
the
top
roasters.
S
The
the
profile
is
generally
matched
or
imagined
between
the
campus
and
the
realness.
Is
this
that?
But
there
is
a
slightly
shift
of
this
gesturing
location
to
northward,
to
northward
in
the
model
in
the
model
and
when
we
look
at
and
then
we
look
at
the
more
redundant
wind
armor
one
is
relatively
secondary
compared
to
the
dominant
thunder
wind
pattern,
especially
over
the
northwest
region.
S
As
you
can
see
from
the
from
the
magnitude
here
that
over
the
california
region,
it
is
your
measure
significantly
and
we
we
do
select
the
to
see.
The
cadmium
has
tends
to
have
this
software
bearing
software.
The
balance
of
the
source
of
the
balance
for
this
kind
of
question
this,
comparing
to
the
married
to
analysis.
S
That
is
average
over
this
heavy
pressure
days,
and
this
is
the
anomalies
between
these
having
present
days
to
and
the
mean
stage
that
is
average
of
all
the
orders
of
these
racisms,
and
the
brief
message
here
is
that
cameras
has
has
consistent
the
sea
level.
Quite
short,
normally
pattern
combat,
also
lows
of
surface
light
pressure.
S
Norway
pattern
when
on
comparing
to
the
married
too-
and
this
signal
is
consistent
between
northwestern
california,
but
of
a
different
center
of
different
central
regions
here
and
for
the
land
for
impression
here
is.
The
here
is
just
a
general
picture
of
the
mean
of
the
mean
preservation
average
over
its
heavy
price
preservation
days
in
campus
and
how
it
looks
like
in
america.
S
S
P
Yeah
this
also,
you
know
we
we're
entering
the
discussion
period,
so
let's
spend
maybe
the
next
15
minutes
on
questions
that
we
might
have
for
all
the
speakers
here
of
the
second
half
and
at
12
we
transition
over
into
this
informal
discussion
and
that
discussion
should
focus
on
these
aspects.
Can
can
we
remove
some
features
from
the
current
configuration?
The
development
configuration
of
the
community
atmosphere
model.
B
G
G
B
We'll
have
four
moderated
breakout
rooms
with
christiana
cecile,
haney,
brian
medeiros,
and
I
as
the
moderators
plus
two
unmoderated
breakout
rooms,
they'll
run
for
about
10
minutes,
and
there
are
some
possible
topics
but
you're
welcome
to
let
the
discussion
go
anyway,
any
way
it
does
for
the
unmoderated
rooms.
I
would
you
know
it
would
be
nice
if
you
might,
if
you
named
or
nominated
a
a
rapper
tour
who
could
take
at
least
you
know
rudimentary
notes
and
send
them
back
to
the
co-chairs,
but
other
than
that?
B
That's
that's
all
and
then
after
the
breakout
rooms
will
come
back.
As
christiana
said,
for
kind
of
a
group
discussion
about
the
survey
that
was
sent
out
last
month
on
removing
features
from
cam.
So
I
guess
the
first.
Are
there
any
questions,
any
any
questions
that
people
would
like
to
ask
speakers
from
this
session
or
the
last.
P
I
actually
had
a
question
for
the
earthworks
team.
I
hope
the
bridge
or
others
are
still
available
or
the
summer
I
saw
that
the
transition
to
gpu
technology
is
partly
supported
by
companies,
for
example
amd,
and
I
was
wondering
whether
we
are
still
now
dealing
with
open
source
code.
Or
is
this
not
proprietary?.
P
A
Well,
that's
our
intent
and
we're
still
we're
working
on
that,
for
example,
we're
working
with
the
impasse
core
team
to
get
the
impasse,
7
dicor
with
gpu
directive
markups
in
into
cam,
and
we're
working
with
pumas
to
ensure
that
pumas
group
to
ensure
the
microphysics
markups
get
back
in
and
we're
also
doing
that
with
with
club
which
the
doe
has
given
them
funding
in
order
to
move
to
the
directives.
So
our
strategy
is
to
work
with
the
code,
development
teams
and
the
science.
E
K
A
Okay,
if
I
could
just
add
a
little
side
comment
about
that,
you
know
if
you,
if
you
don't
do
that,
it
becomes
very,
very
difficult
to
maintain
synchronization
with
the
the
evolving
code,
especially
in
the
physics
which
is,
as
as
everybody
probably
well
knows,
it's
constantly
mutating
under
scientific
activity.
So
that's
a
key
effort
of
doing
any
any
of
these
kind
of
refactoring
efforts.
P
A
Four
yeah
a
full
atmospheric
and
full
full
ocean
implementation
again
with
the
impasse
ocean,
but
yeah
the
that's.
The
the
vision
is
to
do
the
full
offload,
and
you
know
that,
if
you
don't
do
that,
then
you
get
you
know
amdahl
bottlenecks
and
stuff
like
that
is
with
the
data
transfers
back
and
forth
to
gpus.
A
P
I
A
H
P
G
B
I
I
would
say
then,
if
it
looks
like
there
are
no
more
questions.
Why
don't
we
go
into
the
breakout
rooms
for
10
minutes
and
then
return
to
talk
about
the
survey
and
the
sort
of
software
management
issues
that
the
survey
was
the
result
of
so
so
elizabeth?
Could
you
could
you
launch
the
the
breakout
rooms?
D
Q
P
P
P
P
A
B
B
B
B
Another
another
less,
I
would
say
less
salient,
but
still
significant
thread
was
that
you
know
people
are
people
seem
to
be
willing
to
let
the
oil
air
and
die
core
go,
but
that
this
is
going
to
require
require
development
of
scam,
because
scam
right
now
relies
on
the
existence
of
the
valerian
dicor
for
coupling
and
other
things.
So
I
someone
added
this
to
the
slides
that
this
this
development
will
will
take
place
later
in
the
year.
B
Okay.
But
so
now,
let's,
let's
let's
go!
Let's
go
and
talk
about
sdcam,
and
at
this
point
you
know
if
steve
or
cheryl
want
to
jump
in
and
correct
me
or
or
anything
else,
please
do.
But
this
is.
This
is
a
list
that
was,
you
know,
a
list
that
we
we
compiled
over
the
last
couple
of
days
listing
reasons
why
why
sb
cam
is
hard
to
maintain?
There
are
two
versions
of
sp
cam.
B
There
are,
you
know
there
are
copies
of
physics
in
each
version
of
sp
cam,
including
a
copy
of
club,
which
apparently
is
is
pretty
outdated
there
there.
You
know
there
are
frequent
instances
where
lots
of
files
during
pull
requests
that
have
to
be
modified
are
our
many
of
the
files
are
in
sbcam
and
obviously
sp
cam
is
is
expensive
to
run
so
that
climate,
validations
of
sbcam
are
are
are
are,
are
not
frequent.
B
So
anyway,
the
software
engineers
have
been
modifying
sp
cam
to
keep
it
up.
You
know,
keep
it
passing
tests
with
the
current
cam
configurations,
but
we're
not
sure
it
it's.
It's
still.
You
know
producing
sensible
results
with
with
the
modifications
that
have
been
introduced
to
the
model
in
the
last
few
years,
so
I'm
going
to
turn
it
over
now
and
let
steve
or
or
cheryl
and
actually
kevin
kevin
has
his
hand
up.
C
Slower
yeah
right
so
I,
but
I
think
that
it's
worth
mentioning
and
now,
I'm
being
a
little
bit
provocative
right
that
each
year
that
goes
by
that
we
that
we
make
cam
more
comprehensive
and
complex
the
more
in
line
or
the
more
potential
that
sp
cam
is
actually
more
affordable.
Relatively
it's
just
a
side
thought
now.
That's
not
completely
true,
because,
of
course,
some
of
the
updates
that
go
into
the
full
cam
are
needed
for
for
parameterizations,
outside
of
or
within
sp
cam
on
the
side.
E
B
B
K
For
instance,
mike
in
his
paper,
mentions
that
the
csm
load,
balancing
infrastructure
is
not
capable
of
really
handling.
The
kind
of
focus
he
wants
to
put
on
having
different
grid
cells
have
different
number
of
super
parameterized
cells.
I
think
I'm
I'm
getting
that
correct.
Let
me
correct
if
I'm
wrong,
but
we
have
the
ability
to
do
that.
We
just
don't
have
the
time
to
implement
it,
because
we're
spending
all
of
our
time
bandaging
up
the
current
infrastructure
and
that's
really
what
motivated
this
discussion.
G
G
O
K
D
O
Yeah,
so
there's
been
a
lot
of
developments
right,
marat
has
sort
of
moved
on
to
g
sam
and
that
lives
externally
of
sp
cam.
The
biggest
mmf
development
is
happening
inside
the
doe,
the
e3sm
mmf
project,
so
as
part
of
the
xscale
computing
project,
a
whole
new
pot
of
money
emerged
that
was
the
only
gpu
portable
target
and
there's
been
a
ton
of
development
in
that
code
base,
which
is
open
source,
but
I
think
it
would
be
like
oil
and
water.
O
That's
going
to
have
all
sorts
of
wonderful
features
like
we
know
and
things
that
you
would
like
in
an
less,
but
it's
going
to
be
pretty
difficult
to
read
and
pretty
difficult
for
a
grad
student
to
do
any
experiments
with
and
the
part
of
me
that
thinks
would
be
really
important
to
keep
a
version
of
sam
alive
inside
csm
is
the
part
of
me
that
thinks.
Grad
students
should
be
able
to
read
cloud
resolving
code
and
muck
with
it
and
interfere
with
it
and
do
experiments
with
it
yeah.
K
Yeah
that
that
sounds
like
a
one
of
those
areas
where
you
know
we
struggle
with
this
trying
to
make
code
performance,
but
also
keep
it
readable,
because
you
know
that's
part
of
what
scientists
need
to
be
able
to
do
is
to
is
to
understand
what
they're
running.
So
this
is
why
and
there's
going
to
be
a
talk
on
this
tomorrow
morning.
K
We
would
like
to
move
to
the
new
openmp
which,
unlike
openhcc,
it
supports
more
architectures,
different
types
of
gpus,
but
also
other
types
of
accelerators
like
fpgas
and
stuff
that
are
coming
up,
that
the
openmp
standard
is
now
capable
of
handling,
so
that
is
sort
of
the
opposite
approach
that
doe
is
taking
with.
I'm
spacing
the
name
of
their
tool.
Now
it
looks
like
c
plus
plus,
but
it's
not
c,
plus
plus,
and
it
turns
into
either
c
plus
plus
or
some
gpu
code
yackle.
No,
it's
something
else,
cocos
cocos!
K
O
The
laws
to
respond
to
there
I
mean
yeah,
so
the
that
the
mmf
team
in
doi
has
has
moved
to
this
yackle
thing.
That's
a
back
end
that
can
be
c
plus,
plus
and
run
on
on
gpus
or
on
amd
machines,
and
you
know
front
things
are
changing
under
foot
with
architecture.
So
there's
a
lot
of
appeal
to
trying
to
bring
that
into
csm.
Perhaps
if
you
want
to
do
something
architecture,
agnostic
without
reinventing
the
wheel
and
that's
a
strategy
that
could
be
considered,
but
it's
a
big
resource.
O
K
K
B
Could
I
ask
another
question,
please
so
I've
I've
heard
you
know
kind
of
the
self-contained
aspect
of
sb
cam
being
I
I've
heard
it
described
both
as
an
advantage
and
a
disadvantage.
I
I
you
know,
michael
this
morning
said
yeah.
Well,
you
know
sp
cam's
a
is
a
rapper
and
it
calls
its
own.
You
know
whatever
parameterizations
it
needs
it
calls
it
has
them
internally
and
and
uses
them,
and
then
I've
also
heard
that
you
know
it's
a
it's
a
it's
a
burden
that
we're
maintaining
several
versions
of
club.
B
K
K
I
think
you
know
the
ultimately
have
you
know
an
engineer,
stare
at
it
and
say:
okay,
here's
here's
a
way
so
that
it
doesn't
have
to
be
one
thing
right
now:
it's
two
things
inside
of
of
cam
right,
it's
it's!
Basically,
you
run
one
of
the
two
microphysics
versions
and
you
have
two
versions
of
svcam,
but
that
in
itself
is
limiting
and
sp
cam
is
not
really
amenable
to
growing
out
of
that
box.
B
K
I
think
we
need
a
connection
to
the
community.
That's
doing
this
so
that
we
can
start
talking
about
I'm
a
big
fan
of
co-design,
where
we
actually
meet
with
people
before
they
go.
Do
something
and
figure
out
a
good
way
to
do
it
and,
in
this
case,
leap
is
seems
to
be
ready
to
fire
up
a
bunch
of
work
and
if
we
can
get
involved
and
help
guide
some
of
that
work,
we
could.
G
K
I
K
K
And
you
know
just
glancing
at
mike's
paper,
I
say:
well,
okay,
we
should
go
ahead
and
implement
you
know
we
should.
We
should
have
a
target
running
the
load
balancing
that
would
basically
balance
the
physics
regardless
you
know
so,
essentially
you
would.
You
would
specify
how
much
work
is
going
on
in
a
grid
cell
and
that
would
essentially
be
proportional
to
the
number
of
of
mmf
inside
of
each
grid
cell,
which
is
variable
and
automatically
load
balance
that
why
not?
O
Yeah,
I
think
it's
I
I
do
think
it's
the
beginning
of
a
longer
conversation
right.
I
feel
a
little
self-conscious
that
I've
like
projected
my
myopic
interest
in
sdcam
on
really
busy
csm
developers.
You
know-
and
you
know
leap
does-
is
very
interested.
There
is
real
important
work
that
can
be
done
for
prototyping
machine
learning
models,
but
it
you
know
that's
different
than
maintaining
it
as
a
feature
that
that
you
know
is
there
with
the
model.
I
think
some
market
research
is
needed
to
understand
in
the
wider
csm
community.
G
O
G
O
K
E
K
Feels
like
like,
like
there's
a
lot,
which
is,
which
is
why
you
know
we're
I'm
trying
to
you,
know
pivot
from
we
need
to
get
rid
of
it
too.
We
need
to
figure
out
how
to
do
this
right.
You
know
in
order
to
support
the
community,
which
is
clearly
out
there
and
yeah.
So
this
isn't
sorry
mike
not
about
you.
B
F
Which,
I
think
is
it
basically
brings
up
the
point
that
I
was
about
to
make,
which
is
that
if
they,
if
these
people
don't
have
cesm's
version
of
sp
cam
to
go
to,
they
have
nowhere
else
to
go
to.
I
don't
think
that
there's
anywhere
I
mean,
maybe
the
doe
version
is
somewhat
accessible,
but
it's,
I
think,
more
expensive
and
harder
to
use
than
than
our
version
of
sp
cam,
and
so
I
think
that
we
have
a
we
have
a
tool
here.
F
K
G
K
P
Maybe
I
want
to
just
add
a
couple
of
thoughts,
so
I
think
there
is
a
lot
to
do.
Actually
if
we
decide
to
keep
the
sp
version.
For
example,
we
move
to
the
new
grid
configuration
in
the
vertical.
This
must
be
looked
at.
You
know
also
talking
of
scientifically
validated
versions.
We
have.
These
other
die
cores
coming
on
board,
which
are
no
longer
on
the
lat
long
grid.
I
guess
mikey
indicated
that
spectral
element
has
been
used
in
your
group,
but
we
really
don't
know
whether
there's
a
scientific
basis.
P
P
Also
any
attempt
to
use
sp
can
with
cam
six
physical
prioritizations,
because
maybe
that's
this
concept
that
they
are
all
embedded
in
the
old
club
version.
So
all
these
aspects
should
be
really
should
be
addressed
and
of
course,
leap
would
be.
A
great
con
you
know
would
could
make
a
great
contribution,
but
again,
as
mike
mentioned-
maybe
not
they
probably
don't
want
to
be
the
owner
of
sp
cam.
Also,
I'm
concerned
that
sam
the
underlying
clock,
resolving
model
is
basically
no
longer
maintained.
P
O
Yeah
just
to
respond
quickly
to
that
yeah.
I
think
I
think
opening
a
line
of
comm
with
the
doe
group
would
be
really
helpful
towards
this
visioning,
because
there
has
been
a
lot
of
vertical
resolution,
sensitivity,
testing
inside
that
project,
as
well
as
dicor
testing,
and
there
have
been
some
decisions
made
to
upgrade
the
physics
to
p3
and
shock
inside
pam
and-
and
it
is
possible
to
collins
comment
that
there
might
be
some
opportunity
that
the
engineers
would
recognize
to
leverage
that
to
make
it
an
external
line.
O
I
know
at
the
beginning
of
that
project.
Everyone
was
admiring
cheryl's
implementation
of
sp
and
csm
1.1
as
a
basis
for
their
starting
point,
and
maybe
there's
a
reciprocation
there.
That
could
happen,
but
you
know
and
that
the
architecture
agnostic
aspects
of
pim
might
be
quite
strategic
for
a
future.
I
don't
know
the
the
format
of
directio,
but
that
conversation
will
be
interesting
to
see.
O
I
think
that
the
doe
project
seems
to
prioritize
these
massive
simulations
that
just
have
to
do
around
a
thousand
nodes
of
insight,
no
matter
what
the
science
context
is
there,
so
they're,
probably
less
interested
in
load,
balancing
niche
visualized
resolution
for
the
grassroots
community
that
and
may
have
overlooked
some
of
the
load,
balancing
and
regionalization
opportunities
that
yeah
that
csm's,
more
diverse
user
base
could
benefit
from.
So
maybe
there's
like
a
middle
ground
strategy
there
between
the
two
code
bases
that
could
really
fill
a
niche.
O
G
O
G
G
G
B
B
G
It
does
but
can
I
can?
I
ask
there,
are
some
questions
in
the
chat
about
the
possibility
of
using
of
doe
and
us
using
the
same
externals
and
leveraging
all
the
work
that
walter's
done
to
their
cloud.
Resolving
model
is,
if
that's
not
feasible,
to
marriage.
R2
works
together,
then,
okay,
but
if
it
is,
then
maybe
we
should
include
walter
in
these
conversations.
K
G
B
K
Well,
and
also
possibly
with
galen,
because
you
know,
there's
we've
had
a
problem
in
the
past
and
we
even
have
a
problem
with
cs
impunity
of
this.
You
know
the
number
of
scientists
wanting
to
contribute
stuff
relative,
the
number
of
software
engineers
who
are
available
to
to
put
it
in
there's.
You
know,
there's
a
very
large
discrepancy
in
numbers
and
what
I
saw
yesterday
was
a
whole
lot
of
postdocs
and
one
software
engineer,
and
I
just
don't
know
how
to
make
that's
gonna
work.
K
K
B
J
At
least
yeah
hi
everyone-
I
was
just
going
to
follow
up
real
briefly
on
sheryl
craig's
comment
and
really
take
this
back
to
cmap
julio.
I
wonder
whether
it
might
not
be
worth
a
discussion
with
eric
weaver
at
nsf.
There's
a
general
issue
that
the
science
and
technology
centers
are
doing
really
important
things
and
they
have
legacies
like
sp.
J
Cam
leap
is
likely
to
have
some
earthworks,
possibly
going
forward
and
look
into
the
possibility
of
some
dedicated
funding
to
keep
these
very
important
legacy,
projects
that
involve
the
cesm
going
that
really
are
stretching
beyond.
You
know
the
capabilities
of
ncar
with
current
resources
to
do
I
don't
know
what
will
happen,
but
you
know
the
msf
has
made
an
enormous
investment.
J
Mike's
made
a
really
strong
case,
how
important
the
sp
cam
investment
was
to
let
those
falter
because
of
a
lack
of
resource
for
something.
That's
you
know
in
the
overall
scheme
of
things.
Kind
of
minor
would
seem
sort
of
tragic,
so
it
I
mean
I
don't
know.
If
anything
will
happen.
If
I
were
back
on,
you
know
in
my
days
as
cab,
chair,
I'd,
be
putting
this
on
the
agenda
for
the
cab
as
something
prep
to
talk
to
that's
up
about,
but
I
would
encourage
that
discussion.