►
Description
2020 CESM Workshop Day 3 - Wednesday, 17 June 2020 - 8:30-12:00 - Earth System Prediction Working Group
The 25th Annual CESM Workshop will be a virtual workshop with a modified schedule on its already scheduled date. Specifically, the virtual Workshop will begin with a full-day schedule on 15 June 2020 with presentations on the state of the CESM. On 16-17 June 2020, working groups and cross working groups have half-day sessions, some with presentations and some that are discussion only.
More information about the CESM Earth System Prediction Working Group:
http://www.cesm.ucar.edu/working_groups/earth-system-prediction/
More information about the workshop:
http://www.cesm.ucar.edu/events/workshops/2020/
Talks included:
A New CAM6+DART 80-Member Ensemble Reanalysis for 2011-2019: Initial Conditions for CAM, CLM and CICE Forecasts and Forcing for POP, CLM and CICE - Kevin Reader
S2S with CESM1 - Yaga Richter
Advantages of 25Km resolution ocean model in GEOS Seasonal Prediction System - Andrea Molod
Two-year Predictions of ENSO Event Duration during 1954-2015 - Xian Wu
On the prospect of developing seasonal to decadal soil moisture forecasting system - Sanjiv Kumar
Decadal Prediction with CESM: recent progress & outstanding challenges - Steve Yeager
The climpred diagnostic package - Riley Brady
Near-term Simulation Plans - ESPWG Co-chairs
General Discussion
D
B
A
C
E
C
And
then
it's
been
helpful
to
in
the
welcomes.
You
know
ground
rules
to
remind
people
about
the
code
of
conduct
and
I'll
put
that
link
over
in
the
sidebar
and
also
you're
just
deciding.
If
you
want
people
to
put
questions
just
in
the
chat
and
read
them
or
have
people
raise
their
hand
and
unmute
themselves,
it
seems
like
from
the
other
groups.
C
D
C
G
F
B
D
H
D
F
I
E
C
My
goodness
well,
it's
gorgeous
and
there
people
are
using
the
road
in
like
a
park.
There
are
so
many
bikers
and
joggers
just
down
the
middle
of
the
road,
but
it
is
so
beautiful
and
the
parking
lot
is
so
full
of
bikers
and
hikers
and
construction
workers,
but
other
than
our
our
team.
Here
running
the
meeting
I
haven't
seen
anyone
else,
one
one
cleaning,
woman
but
but
I
haven't
seen
anyone
well.
H
H
F
H
F
H
F
D
F
D
F
C
So
if
we
get
a
large
group-
and
there
happens
to
be
some
background
noise,
what
I
did
yesterday
was
just
mute
everyone
and
then
the
speaker
can
go
ahead
and
unmute
themselves.
So
we
had
some
people
having
like
an
animated
discussion
in
the
background
and
so
try
to
scroll
through
quickly
and
find
that
person
is
complicated.
So
it's
just
mute
everyone
and-
and
let
the
speaker
continue,
sounds.
D
F
D
Okay,
can
everybody
give
me?
Okay,
all
right,
good,
okay,
good
morning,
everybody,
this
is
the
inaugural
meeting
of
the
Earth
System
prediction
working
group
and
I'm
Kathy
page
own
from
George
Mason
University,
one
of
the
co-chairs
and
the
other
co-chairs
are
Yaga
Richter
and
Steve
Yeager
from
NCAR
and
of
course,
there
we
go
okay
first
thing:
first
I
want
to
just
remind
you
of
the
entire
code
of
conduct
and
respectful
dialogue.
D
D
We
have
a
mailing
list
and
if
you're
not
on
the
mailing
list,
please
please
join
it
right
now,
or
at
least
as
a
black
check.
We
had
149
members
so
really
exciting
to
see
so
many
people
excited
about
Earth,
System
prediction
and
we
also
have
created
a
zoo
look.
Channel
julep
is
like
swag,
it's
been
familiar
with
that
and
this
is
a
way
for
us
to
easily
interact
with
each
other
answer.
D
D
So,
as
a
new
working
group,
you
might
wonder
why
should
be
on
have
an
assistant
for
diction
working
group
and
so
I
just
want
to
go
over
some
of
the
reasons
why
this
working
group
exists
and
and
what
kinds
of
things
we
want
and
what
kind
of
things
we
want
to
do
in
this
group
and
the
goals
of
the
working
group.
So,
as
you
all
know,
s
to
D.
So
in
this
case,
I'm
referring
to
sub
seasonal
to
the
Cadle
prediction,
is,
is
of
increasing
interest
to
researchers
and
stakeholders.
D
And
that's
because
a
lot
of
these
kinds
of
experiments
are
done
using
operational
models
and
they're,
mostly
ensembles
of
opportunity,
and
it
makes
it
challenging
to
then
do
some
of
the
kinds
of
sensitivity,
experiments
and
things
that
we
would
like
to
be
able
to
do
for
understanding.
The
sources
of
predictability.
D
And
in
doing
these
kinds
of
experiments,
we
have
a
whole
list
of
outstanding
challenges
and
opportunities
and
things
that
we
can
do
many
of
these
challenges
and
opportunities
are
kind
of
universal
across
the
timescales
so
coming
together
from
so
seasonal
three
decade,
old
prediction
helped
us
to
bring
together
and
kind
of
leverage.
Many
of
the
efforts
in
doing
initialize
prediction
experiments
that
need
to
be
kind
of
brought
together
for
for
kind
of
a
concentrated
effort
around
solving
some
of
these
challenges.
D
D
So
this
is
a
CSN
assistant
prediction
working
group,
and
so
the
idea
is
that
this
we
will
use
the
CSM
model
for
doing
our
system.
Prediction
and
I
just
want
to
highlight
that
there's
really
some
unique
capabilities
of
the
in
karma
dolls
and
TSM.
Particular
it's
a
comprehensive
Earth,
System
prediction
system
and
one
of
the
things
I'm
very
excited
about
is
the
fact
that
we
can
look
at
multiple
time
scales
with
a
single
framework
and
that
this
framework
is
accessible
to
the
community
and
easy
to
use
the
research.
D
So
the
key
things
being
initialized
or
system
prediction
and
on
timescales,
ranging
from
functional
to
the
Cadle,
and
we
want
to
collect
community
input
on
priority
experiments,
infrastructure
needs
and
models
and
systems
development
needed
to
advance
persistent
prediction.
Research.
We
will
plan
and
execute
large
ensemble
initialized
pine
tasks
and
for
crash
simulations.
D
Those
who
computationally
burdens
them
for
individual
researchers
to
undertake
will
orchestrate
the
development
of
tools
and
frameworks
to
enable
a
broader
community
to
explore
predictability,
mechanisms
and
sensitivities,
prediction
systems,
design
and
will
promote
collaboration,
collaborative
multidisciplinary
research
and
serve
as
a
hub
for
the
Coordination
of
effort
and
will
provide
a
reliable
catalyzing
forum
for
sharing
the
latest
Earth
System
prediction.
Science
results.
D
So
there
are
already
some
existing
system
prediction
working
room
experiments:
these
are
the
forecasts
using
on
TSN
and
we
have
some
from
the
sub
seasonal,
so
seasonal
time
scale,
using
the
C
sn1
sub
seasonally
forecasts
that
have
been
contributed
to
this
sub
seasonal
experiment
database,
as
well
as
from
expertly
forecast
experiments
using
wacom
following
that
same
scene,
sub,
sub,
seasonal
experiment,
protocol-
and
these
were
these-
are
available
from
Yonder
Victor
on
the
seasonal
time
scale.
There
are
north
north
american
multi-model
ensemble.
These
forecast
experiments.
These
are
monthly
seasonal.
D
We
forecast
them
with
CES,
m1
and
they're
available
in
the
end
of
the
media
and
there's
the
decay
toll
prediction
large
ensemble
experiment,
which
is
available
on
home
and
car
10,
marked
in
gray
here,
I'm,
showing
from
the
experiments
that
are
planned
or
in
progress
that
will
be
part
part
of
the
artist
and
prediction
working
group
experiment.
On
the
sub
seasonal
time
scales.
There
are
CSS
2
sub
seasonally
forecasts
that
are
planned
on
the
inter
annual
time
scale.
There's
tsn2
multi-year,
we
forecast
their
plans
and
then,
on
the
decayed
all
time
scale.
D
So
then,
I
want
to
just
on
introduce
that
we
have
a
set
of
discussion.
Questions
plans
for
later
in
this
meeting
and
they're
also
listed
in
the
Google
Doc
I'm
not
going
to
read
through
them
now,
because
we'll
have
time
to
talk
about
them
later
in
discussion.
But
I
encourage
you
to
take
a
look
at
them
and
add
your
suggestions
and
ideas
in
the
Google
Doc.
D
In
terms
of
these
questions
or
other
questions
that
you'd
like
to
raise,
and
if
you
have
those
ideas
or
someone
else
listed,
something
you
agree
with,
or
support
or
I
want
to
say,
we
should
do.
Please
go
ahead
and
add
a
+1
to
that
in
the
Google
Doc,
and
that
will
help
facilitate
our
discussion
for
later
and
so
I'll
just
end
with
an
overview
of
today's
session
agenda.
D
D
D
F
K
A
A
So
this
new
reanalysis
is
similar
in
quality
to
the
inset
in
Korea
analysis,
but
with
eating
someone-
and
it's
also
similar
to
the
20th
century
right
now,
so,
hopefully
you're
familiar
with
those
products.
It's
an
upgraded
version
of
the
cam
for
to
to
reanalysis.
We
finished
in
2012,
but
this
uses
cam
six
at
one
degree,
which
is
the
more
the
workhorse
level
of
present
a
more
data
products
for
use
in
hein
casts
that
most
of
these
data
products
are
native
to
CES
M.
A
A
A
So
assimilations
have
three
components:
a
model,
observations
and
dark
to
combine.
The
first
two
here
are
some
of
the
details
of
the
model,
but
it's
important
to
keep
in
mind
that
the
reanalysis
these
look
mostly
like
the
actual
weather.
All
the
specific
cam
version
has
a
somewhat
marginal
influence.
A
A
So
this
assimilation
assimilated
several
million
observations
per
day
into
the
cam
model
state.
If
we've
defined
to
be
these
variables,
these
wind
and
temperature
observations
from
airplanes,
radio
signs
and
satellites
and
GPS
refractivity
observations,
which
is
basically
atmospheric
density,
and
what
this
yields
is.
Eighty
equally
likely
CES
M
states
consistent
with
the
actual
weather
and
cam
physics.
By
consistent,
we
mean
that
it
explicitly
uses
the
two
main
sources
of
uncertainty,
which
are
odd,
raishin
errors
and
model
ensemble
spread
to
balance
the
information
in
the
observations
and
in
the
model
hindcast.
A
So
the
benefit
of
having
lots
of
observations
is
that
a
one-degree
resolution
model
doesn't
generate
great
hurricanes
by
itself,
but
here's
hurricane
sandy
shortly
before
landfall
at
0z
on
October
29
2012.
This
is
in
the
analysis.
Ensemble
means
surface
pressure,
so
I've
truncated.
The
range
of
the
surface
pressure
at
950
just
to
highlight
the
hurricane,
so
the
central
pressure
from
the
National
Hurricane
Center
best
track
was
950
active
Baskervilles,
the
CAF's.
There
are
courtney
pressure,
wind
relationship,
they
calculate
to
be
950,
950,
Pascal's
and
our
ensemble
mean
came
in
at
956
point
six.
A
So
this
is
good
given
that
the
location
of
the
ensemble
central
pressure
could
be
wrong
by
up
to
sixty
kilometers.
Due
to
the
one
degree
resolution
and
the
N
South
Pole
standard
deviation
of
central
pressures
is
one
point:
seven
hectopascals.
So
the
ensamble
is
not
inconsistent
with
the
National
Hurricane
Center
numbers.
A
So
this
is
an
unprecedented
data
set
it's
a
full
ensemble
output
and
lens.
Then
previous
ensemble
down
assimilation,
experience
I've
shown
the
value
of
this,
and
the
output
is
four
times
per
day
for
some
variables
for
nine
years.
This
is
a
unique
combination
with
many
uses,
get
real-world
initial
conditions
for
cam
ensembles
with
justifiable
spread
also
for
CLM,
sea-ice
and
Mozart,
but
that
needs
a
little
investigation.
There's
realistic
atmospheric,
forcing
for
all
surface
components
in
simulations
and
assimilations
with
justifiable
variability.
A
It
consists
of
CSM
gridded
data
which
were
all
familiar
with,
but
also
observations,
space
data
which
are
on
some
mobile
model.
Estimates
of
the
observations
at
the
observation
locations
so
using
much
of
this
data
requires
no
knowledge
of
data
simulation,
but
people
are
interested
in
the
data
size.
Data
assimilation
will
find
a
trove
of
useful
and
interesting
data.
A
A
The
second
product
that
may
be
very
useful
for
the
Earth's
system
prediction
is
the
ensemble
of
restart
file.
Sets
that
we've
created
it's
available
weekly
for
hindcast
studies
and
intermediate
dates
can
be
generated
quickly.
If
you
need
a
specific
day,
here's
a
snapshot
from
each
of
the
four
active
components,
restart
files.
These
are
well
spun
up
model
states.
A
A
If
you're
not
comfortable,
looking
at
things
in
observation
space
like
we
just
did,
then
on
cams
grid,
you
can
see
the
spatial
and
temporal
distribution
of
the
corrections.
The
model
state
caused
by
the
observations.
Yawn
szabo
mean
of
these
quantities
is
available
four
times
per
day
for
the
whole
nine
years
and
members
are
available,
weekly.
A
So
returning
to
hurricane
sandy
we're
looking
at
ensemble
member
one
difference
from
the
mean
for
the
surface
pressure
turns
out
that
member
one
is
very
close
to
the
mean,
and
so
hurricane
sandy
is
really
not
visible.
In
this
difference
plot
the
central
pressure
of
the
mean
was
nine
hundred
fifty
six
point.
Six
and
member
one
was
nine
fifty-seven
point
one.
A
A
A
Here's
the
status
of
the
project,
the
first
tree
analysis
using
the
two-degree
camphor,
was
completed
about
eight
years
ago.
The
current
project
for
one
degree
kam
six
reanalysis
as
years
old
2011
to
2018,
completed
with
one
more
to
go
that
should
be
finished.
In
the
next
few
weeks
we
may
run
the
reanalysis
from
1999
to
2010.
If
there's
enough,
Community,
Interest
and
computer
time.
A
A
So
our
top
priority
is
to
find
collaborators
to
analyze
the
data,
use
it
for
research
rate
papers,
and
then
we
may
do
the
reanalysis
for
1999
to
2010
and
first
division
interest.
An
alternative
is
to
rerun
completed
years
with
additional
radiance
observations
and
final
notes.
Our
students
we'd
love
to
work
with
you
and
if
you
identify
with
an
upper
upper
underrepresented
group,
we
are
eager
to
hear
your
ideas
and
questions
and
that's
all
I
have.
Thank
you
very
much.
D
D
A
We
don't
have
the
the
increments
explicitly
be
aware,
you
have
the
files
that
can
be
used
to
derive
those,
and,
yes,
those
are
available.
A
six
hourly
frequency
for
the
whole
period,
for
the
ensemble
mean
we
have
it
more
frequently
for
the
ensemble
less
frequently
for
the
ensemble
members.
Those
are
available
about
once
per
week.
A
Yes,
each
time
there's
an
assimilation,
the
statistics
that
we
use
reading
the
observations
and
they
push
the
model
state
towards
the
observations.
They
don't
exactly
match
the
observations,
because
the
observations
have
some
error
in
them
and
we
don't
want
to
match
those
exactly
and
so
the
amount
that
those
model
variables
are
pushed
towards.
The
observations
are
the
observation
increments
and
seeing
a
map
of
that
or
a
time
series
of
that
can
give
you
some
indication
of
how
the
model
is
being
persistently
pushed
towards
observations.
D
D
E
C
F
F
I
have
done
over
the
years
with
the
people
listed
below
Julie
korone,
Sasha
hyemi,
who
came
to
the
/
with
Lantau
son,
Kathy,
Poisson,
Ahmed
and
Steve
Yeager,
and
the
project
started
actually
with
a
NOAA
funded
proposal
that
I
had
with
Judith
pearl
it's
from
NOAA
to
an
issue
in
Vesta
Gate,
the
role
of
the
stratosphere
on
NAO
predictability,
and
this
was
the
first
time
ever
that
we
ran
such
seasonal
hindcast,
the
end
car.
So
in
this
process
we
learned
several
lessons
and
I'll
briefly
describe
to
you.
F
The
data
set
that
we
have
created
and
then
lessons
that
have
come
out
of
it.
So
we
can
shape
the
discussion
of
what
we
will
do
with
CSM
to
later
so
for
the
STS
hindcast
with
CSM
one,
we
primarily
followed
the
sub
seasonal
protocol
or
the
subjects
protocol.
So
we
have
45
day
long
simulations,
starting
with
start
dates
from
1999
to
2015
every
Wednesday.
We
have
10
ensembles
with
the
default
30
level
model
and
10
ensembles
with
the
46
level
model.
So
you
have
little
cartoons.
F
Here's
the
46
level
model
just
has
a
better
representation
of
the
stratosphere
in
a
particular
better
representation
of
sudden
warming's
and
the
QBO.
We
initialize
the
atmosphere.
From
era
interim
we
initialize
the
land
from
an
offline
simulation,
it's
fun
up
with
crew
and
sub
data
as
forcing,
and
then
we
used
the
GRA
55
Forest
Ocean
hindcast
as
ocean
initial
conditions
and
as
Kathy
mentioned
data
is
available
in
the
IRI
library,
with
all
the
other
sub
X
models.
F
So
the
lesson
first
lesson
we
learned
it
was
about
infrastructure.
The
computational
infrastructure
for
us
to
us
hindcast
is
simply
complicated.
So
just
for
the
Honka
set
that
we
created
there
was
8840
model
runs,
and
you
can
imagine
it's
just
not
an
easy
setup
on
the
we
leverage
scripts
created
by
Julie
Curran
for
the
enemy
contributions
with
CSM
one
because
of
the
more
frequent
initialization.
It's
still
quite
a
challenge
to
put
these
through
Cheyenne
or
any
other
system.
So
even
automatic
running
requires
a
lot
of
human
time
and
especially
when
things
don't
go
right.
F
F
F
The
orange
is
the
combination
of
oean
others
and
the
blue
bar
is
showing
you
the
30
level
CSM
one,
and
you
see
that
it
is
quite
on
par
with
NOAA's
operational
model,
and
when
you
add
that,
together
with
the
other
models,
you
can
further
increase,
the
multi
model
mean
we
also
have
really
good
mjo
prediction:
skill
in
CS
m1.
It
is
not
close
to
what
ECM
fw,
what
ECMWF
has,
which
has
much
longer
predictability.
F
All
the
way
up
to
32
days,
CSM
one
is
at
about
27
28,
but
it
still
leads
the
pack
among
the
other
models
that
we
have
in
the
United
States
and
I
just
mentioned
that
in
the
freerunning
model,
the
njo
is
not
the
degrade
and
see
us
someone.
So
we're
excited
to
see
what
will
happen
in
CS
m2
when
the
representation
of
the
MgO
is
much
better.
F
The
third
thing
we
learned
and
is
perhaps
not
a
new
lesson,
but
it
was
nice
to
visualize
it
with
CSM.
Is
that
ensemble
size
matters
a
lot
so
here's
again
a
CC
of
global
two
meter
temperature
as
a
function
of
ensemble,
size
and
CSM,
is
here
in
gray
and
we're
comparing
against
ECMWF
and
n
set
for
the
same
number
of
ensembles.
F
The
other
thing
we
learned
is
that
better
stratospheric
model
does
give
you
better
stratospheric
predictability.
So
what
I'm
showing
on
the
left
here
is
the
representation
of
the
stratospheric
winds
or
particular
tropical
winds
to
show
the
lack
of
the
qbo
in
the
30
level,
freerunning
model
and
46
level
km
v,
which
has
a
QB.
Oh,
it's
not
best
QB,
oh,
but
there
is
an
internally
generated
qbl.
F
At
least
this
is
our
experience
so
I'm,
comparing
here
2
meter,
temperature
ACC
for
the
30
level
and
46
level.
C
is
someone
for
different
seasons
and
there
are
very
small
differences
between
the
bars.
So
both
models
have
very
similar
skill
which
later
on
lets
us
combine
them
into
a
20
member
ensemble,
and
you
see
those
also
for
the
nao
on
the
right.
F
We
have
pretty
good
nao
skills,
so
we
have
correlation
coefficient
of
0.5
all
the
way
out
to
week,
3,
which
is
quite
good
as
far
as
sub
seasonal
models
are
concerned,
but
there
is
no
discernible
difference
between
the
high
top
and
the
low
top
model.
So
if
you
remember,
we
called
up
the
plot
on
the
ensemble
size.
One
could
argue
that
you're
better
off
putting
in
your
computational
power
into
running
more
ensembles,
rather
than
increasing
the
model
lid.
F
Okay
and
the
last
lessons
we
learned
that
this
20
member
ensemble
says
you
can
combine
the
models
together
to
have
a
20
member
ensemble
they're
very
useful
for
basic
research
on
predictability.
So
this
is
a
figure
from
a
paper
long
tail
Sun
is
putting
together
on
the
attribution
of
nao
predictability
and
on
the
left.
Here
you
have
a
CC
for
nao
for
strong
and
pull
and
weak
polar
vortex
events
in
the
middle.
You
have
the
same
for
just
neutral
events
and
on
the
right
for
ENSO
events.
F
So
in
short,
we've
demonstrated
the
utility
of
CSM
one
as
a
tool
for
us
to
is
research
through
the
high
logistical
burden
of
hindcast.
The
working
group
is
a
perfect
venue
for
these
coordinated
experiments.
Csm
one
has
over
a
really
good
sub
seasonal
skill.
However,
the
better
representation
of
the
stratosphere
does
not
increase
surface
predictability.
F
D
F
So
I
think
that's
the
conclusion
we're
coming
to
that
in
that
short
window
is
fish.
Do
you
focused
on
week
three
and
four,
if
you
initialize
the
stratosphere,
it
really
holds
its
initial
state
for
this
time
window.
So,
therefore,
if
you're
stratosphere
is
better
in
an
offline
model,
that
does
mean
you
may
be
doing
different
things,
but
in
its
short
time
window
the
benefit
is
small.
F
E
D
A
D
F
Know
because
I
don't
know
of
groups
that
are
able
to
run
with
two
different
versions
of
a
model.
I
ain't
know
Andrea's
on
the
line
and
they
know
they
have
a
stratosphere.
I,
don't
know,
they've
tried
to
degrade
the
stratosphere
or
change
it
to
see
if
it
makes
a
difference
Andrea.
Can
you
comment
on
that.
I
Yeah,
so
we
haven't
done
the
side-by-side,
you
know
lower
and
better,
but
Larry
coy
at
Gao
has
done
a
lot
of
work
looking
at
forecasts
of
opportunity,
so
we
we're
predicting
a
stratospheric
warming
event,
let's
say
and
then
doing
a
better
job
at
predicting
the
surface
meteorology
in
the
wake
one
of
those.
So
again
that
doesn't
say
that
if
we
lower
delayed
or
something
that
it
we
wouldn't
buy,
we
are
getting
some
benefit
from
resolving
that
stratosphere,
predicting
the
warming
and
and
the
aftermath
yeah.
F
Just
FYI
in
our
between
the
two
versions
of
the
model,
the
sudden
warming
predictability
is
about
the
same,
so
I
didn't
show
it,
but
for
some
dates
the
30
level
model
actually
does
better
and
for
some
days
the
46
level
model
does
better.
So
out
of
this
hindcast
period,
there
were
11,
sudden,
warming's
major
sudden
warming's,
which
is
not
that
many,
but
from
the
ones
we
looked
at.
I
I
Our
version
2
was
released,
November
2017,
the
paper
just
came
out
and
our
version
3
that
I'm
going
to
talk
to
you
about
here,
and
the
testing
is
due
for
release.
We're
gonna
freeze
it
later
this
year,
we're
basically
there
and
then
we're
gonna
run
the
retrospective
reanalysis
and
the
end.
The
hindcast
suite
just
a
little
reminder
of
what
this
system
looks
like.
We
basically
grabbed
what
at
the
time
was
the
atmosphere
that
was
being
used
in
our
nwp
systems,
and
that
includes
the
aerosol
model
and
we
added
to
it
than
to
moment
cloud
microphysics.
I
I
We
also
reworked
how
the
runoff
from
the
ice
sheets
gets
to
the
ocean,
and
I'm
gonna
just
show
you
a
little
bit
about
that:
a
new
atmosphere,
ocean
interface
layer
that
that's
a
nice
way
to
get
the
diurnal
cycle
without
having
ten.
You
know
a
really
high
vertical
resolution
near
the
surface
of
the
ocean.
I
We
do
run
a
so-called
weakly
coupled
data
simulation
that
initializes
all
of
this,
and
basically
the
way
that
that
goes
and
this
actually,
as
I,
was
listening
to
Kevin
speak.
It
occurs
to
me
that
that
this
is
something
that
could
be
done
with
that
reanalysis
as
well.
Is
that
you're
running
the
coupled
system,
you're,
replaying
or
nudging,
the
atmosphere
to
that.
I
So
here's
the
ice
sheet
runoff
and
the
old
system
for
a
place
like
Greenland,
we
were
essentially
looking
at
the
closest
woops.
We
were
essentially
looking
at
the
closest
ocean
point,
and
so
if
we
had
melt
and
runoff
from
the
continent,
it
was
running
to
the
closest
ocean
point.
If
you
look
at
the
asymmetry
in
the
topography
over
Greenland,
this
essentially
means
that
at
some
places,
if
I
touch
it
again,
it's
going
to
move
at
some
places.
I
It's
made
a
big
change
on
how
much
of
that
meltwater
gets
where
into
the
ocean
we've
seen
some
nice
impacts
on
the
salinity
and
then
the
old
system
ran
on
the
left.
What
I
would
call
it's?
It's
left
over
from
old
paper,
buy
price
at
all
in
the
mid-80s.
86
I
think
that
basically,
we
put
sort
of
a
junk
skin
layer
in
the
middle.
I
The
atmosphere
talks
to
that
and
then
that
talks
to
the
ocean
and
but
basically
you're
building
in
or
setting
by
the
timescale
of
that
communication
you're
setting
the
the
timescale
of
the
journal
cycle
and
we
went
to
something
that
Santa
Cal
and
Max
Juarez
developed
with
the
cool
skill
in
diurnal
warming.
Basically,
within
the
top
ocean
layer,
we
set
a
profile
of
the
temperature
and
that
skin
temperature
depends
on
that
and
that's
now
a
function
of
the
stability
and
it's
a
function.
It's
just
a
much
nicer
way
to
diagnose
the
diurnal
cycle.
I
Here's
the
difference
between
the
this
is
a
it's
a
change
in
vertical
and
horizontal,
but
we
played
around
with
this,
and
most
of
this
is
because
of
the
change
in
the
horizontal
here's,
the
bathymetry,
so
not
only
on
the
left,
the
old
one
on
the
right,
the
new
one.
Not
only
are
we
seeing
deeper
ocean
basins,
but
if
you
take
a
look,
I
could
have
blown
this
up
and
shown
it
a
little
bit
better.
I
If
you
take
a
look
in
some
of
the
pinch
regions
in
the
Indonesian
throughflow
and
the
Florida
Strait
you're,
seeing
some
nicer,
deeper
sort
of
digging
out
of
things
that
allows
I'll
show
you
in
a
minute
it's
allowing
the
flow
to
move
much
much
nicer,
much
more
much.
You
know
much
more
realistic
flow
in
there.
I
So
here's
an
example
with
the
surface
currents
again
in
the
upper
plot,
we're
looking
at
the
Brazil
current
the
Florida
The
Loop
Current,
and
the
Florida
straight
flow
on
the
right
hand,
side,
the
the
lower
set
is
obviously
Indonesian
through
flow
and
Kuroshio
on
the
right.
Are
these
Oscar
sort
of
observation
based
estimates
on
the
left?
Is
the
old
with
the
coarser
resolution
and
in
the
middle
is,
is
the
is
the
25
kilometer
and
you
can
see
much
more
realistic?
You
know
little
Eddy
motions
in
the
in
the
Gulf.
I
More
robust
flow
through
that
Florida
Strait
same
is
true
for
Indonesia
Kuroshio,
so
this
this
just
also
it
may
be
a
little
hard
to
tell,
but
it's
also
moved
the
location
of
the
current
off
the
coast
a
little
bit,
which
is
again
it's
sort
of
a
more
realistic
set
of
flows.
So
we're
very
happy
with
this.
What
this
is
done
for
us
instance.
If
you
look
at
the
top
on
the
left,
it
looks.
It
may
look
subtle
to
you
but
I'm.
Looking
at
the
salinity
bias
relative
to
the
world
ocean
Atlas
version
13.
I
If
we
look
at
the
newer
one,
it's
also
it's
a
similar
issue,
not
forget
the
high
latitudes,
but
what
we're
seeing
is
the
in
the
version
two
with
the
with
the
coarser
model,
we're
too
fresh
and
too
salty.
Here
we're
not
getting
this
transport
across
here
properly
done
properly
and
those
issues
are
basically
gone
with
the
new
system.
So
we're
very
happy
with
how
we're
transporting
the
the
salinity
in
it
with
the
new
resolution
has
bought
us.
I
A
lot
of
nice
characteristics
also
relative
to
the
analysis,
and
if,
even
if
we
look
at
something
like
the
a
mock
on
the
bottom,
we
see
the
a
mock
sort
of
you
know
left
to
its
own
devices.
These
aren't
seasonal
forecasts,
but
if
we
were
thinking
of
initializing
and
and
running
it
to
its
own
devices,
it
was
dropping
and
we're
doing
a
nice
job
of
retaining
the
a
mock
and
again
this
is
probably
a
mock
that
the
salinity
is
probably
the
horizontal
resolution
that
bought
us
almost
all
of
it.
I
I
Basically,
the
system
used
to
be
under
dispersive
early
in
the
forecast.
These
are
Mason
only
mainly
seasonal
timescales,
and
we
also
had
too
small
of
an
ensemble
size
for
the
month
out
and
AOP
and
AE
kind
of
things,
but
based
on
what
we
look,
what
we
saw
in
the
literature,
we
see
little
evidence
of
additional
skill
from
ensemble
size
beyond
a
few
months.
Let's
say
three
months
out,
and
so
we
ended
up
doing
a
sort
of
a
sub
sampling
strategy
to
do
this
yeah.
I
This
is
just
the
justification
or
justification
that
people
use,
but
what
we
ended
up
doing
was
we
run
in
each
month.
We're
gonna
have
a
full
set
of
40
forecasts
that
run
for
three
months
and
after
three
months,
we're
gonna
sub
sample,
there's
a
k-means
clustering
technique
that
was
developed,
and
so
this
is.
This
is
just
a
schematic
of
what
the
idea
is.
We've
got
the
40
that
we
run
out.
I
For
the
last
six
seven
months
of
the
forecast
on
the
right,
the
top
and
the
bottom
show
the
original
clusters.
If
we
let
all
of
them,
run
out
all
40
run
out
for
the
whole
period
of
time.
That's
the
evolution
that
you're,
seeing
at
the
top.
The
clusters
are
identified
by
the
different
colors
and
then
on
the
bottom.
I
This
is
the
30
days
where
we
do
the
clustering.
So
basically
the
system
is
going
to
be
running.
We're
still
gonna
run
every
5
days.
I
know
that
Kathy
is
not
thrilled
with
that
choice
for
sub
X,
but
we're
gonna
end
up
doing
this
for
T
member
for
the
first
two
months,
three
months
actually
and
then
ten
for
the
rest,
we're
still
gonna
stay
with
this
lag
burst.
Ensemble
strategy,
we've
tested,
many
different,
we've
tested,
lag,
we've
tested,
bursts,
we've
tested
the
come
and
the
combination
still
gets
us
the
best.
I
I
D
I
I'm
trying
to
figure
out
how
to
stop
sharing
there
we
go
so
yes,
the
answer
is
yes.
In
particular,
we've
started
to
smell
it
out
with
MJ
Oh.
What's
hard
for
us
to
do,
is
sort
out
because
there's
a
bunch
of
other
improvements
with
this
system.
So
it's
it's
going
to
be.
We're
not
gonna,
run
a
separate
ensemble
with
the
old
model
with
the
new
resolution
and
things
like
that,
we
don't
have
a
clean
side
by
side.
D
All
right
history
reminder:
if
you
maybe
you
think
of
a
question
later
or
you'd
like
to
ask
the
question
you
I've
reached
added
the
Google
Doc
link.
You
can
put
the
question
in
the
Google
Doc
under
their
presentation
and
then
that'll
help
make
sure
that
your
question
can
get
answered
and
facilitate
discussion
with
the
seeker.
E
L
L
Individual
El,
Nino
and
La
Nina
events
show
a
large
diversity
in
their
temporal
evolution,
as
shown
by
the
time
series
of
monthly
nino
3.4
index.
For
example,
some
strong
El
Nino
terminates
after
peaking
in
winter,
followed
by
multi-year
La
Nina
events
only
other
hand.
Some
El
Nino
lasted
two
years
and
there
are
also
la
nina
in
bands
lasted
only
one
year.
The
duration
of
ENSO
events
is
important
to
the
duration
of
their
associated
climate
impacts.
L
Our
recent
study
have
been
looking
to
the
mechanisms
and
potential
predictability
of
invent
duration
based
on
diagnostic
analysis
and
perfect
model
experiments
conducted
with
the
model
CES
m1,
the
current
operation.
Also,
forecasts
are
limited
to
12
months.
Suppose
you
explore
the
real
world
predictability
of
invent
duration.
We
analyzed,
as
sues
of
sweets
of
two
year
forecasts
conducted
with
the
CES
m1
initialized
on
March
June
and
November.
L
Also,
those
forecasts
are
conducted
under
the
Simek
5
external,
forcing
for
each
invent.
We
compare
the
forecasts
initialized
at
6
differently
times
relative
to
the
second
December,
so
the
lead
time
ring
is:
has
the
range
from
25
months
to
6
months,
first
I'm,
showing
you
one
forecast
baked
them
how
of
a
strong
El
Nino
event
that
lasts
one
year
in
1972.
L
The
onset
and
termination
of
this
El
Nino
is
well
predicted
when
the
forecasts
are
initialized
in
March.
Before
the
onset
of
this
El
Nino,
while
the
subsequent
La
Nina
state
is
not
predicted
until
the
forecast,
our
usual
eyes
in
November,
near
the
peak
of
El
Nino
and
in
the
left
panel,
I'm
overlaying,
the
ensemble
Ming
in
your
3.4
index,
at
all
the
six
different
forecasts
to
assess
the
overall
predictability
of
so
invent
duration,
I'm,
showing
you
the
composite
forecasts
for
all
one
year
and
two
year.
L
Events
during
this
period,
so
for
one
year
any
new
on
average
is
development
and
termination
are
first
predicted
in
forecasts
initialized
in
June.
However,
the
subsequent
Lenina
state
is
not
predicted
until
the
forecasts
is
initialized
in
June
in
the
second
year
and
the
failure
of
predicting
La
Nina
conditions
in
those
earlier
forecasts
is
associated
with
the
model
bias
in
simulating
to
slow
and
so
termination.
L
Here
is
first
year
onset
and
the
two
year,
La
Nina
can
be
predicted
with
the
longest
the
lead
time
of
25
months
when
forecasts
are
initialized
from
the
peak
of
preceding
El
Nino.
So
overall
by
the
peak
first
peak
of
also
invent,
we
can
predict
their
duration
of
both
El
Nino
and
La
Nina,
with
30
monthly
time
I'm,
not
showing
it
here.
The
correlation
skill
is
over
0.6
and
this
high
pre
high
correlation
scale
can
persist
into
the
spring
after
the
second
peak.
L
That
provides
a
high
predictability
up
to
one
year
and
a
half,
and
in
the
next
few
slides
I'll,
be
focusing
on
explaining
the
dynamical
processes
behind
the
predictability
of
invent
duration.
So,
first
I'm
focusing
on
the
equatorial
pacific,
the
oceanic
processes,
by
showing
you
the
longitude
time,
sections
of
the
sea
surface
temperature
in
color,
shading,
thermocline
depth
in
black
controls
and
surface
winner
nominees
in
a
quote:
hurry
up
cific
for
one
year
and
two
year:
el
nino
in
observation
and
november
forecasts.
L
In
one
year,
el
nino,
the
sea,
surface,
temperature
warming
and
surface
wind
anomalies
developed
a
few
months
earlier
over
the
western
pacific.
Compared
to
that
of
two
year,
el
nino,
the
early
onset
lists
you
the
earlier
arrival
of
the
delayed
oceanic
feedback.
So
there
is
a
thermocline
showing
develops
in
the
late
summer
in
the
first
a
year
denoted
by
the
negative
values
in
contours.
Here.
L
Why
are
you?
You
only
knew
due
to
its
late
onset,
the
thermocline
shoaling
occurs
a
few
months
later,
when
the
forecasts
initialized,
with
the
opposite,
thermocline
depth
and
nominees.
They
show
distinct
evolutions
in
the
subsequent
seasons.
The
initial
thermocline
shoaling
propagate
eastward
and
lists
you
the
termination
of
any
new
in
springtime,
while
into
year
any
new.
The
later
thermocline
surely
do
not
arrive
at
the
eastern
coral
pacific
in
spray
and
the
el
nino
start
to
redevelop
in
boreal
summer
when
there
is
some
seasonal
intensification
of
pyrenees
feedback.
L
The
timing
of
oceanic
feedback
is
also
important
to
the
predictions
of
La
Nina
duration.
We
see
that
the
ensemble
forecasts
are
also
initialized
with
the
obvious
signs
of
thermal
contact
anomalies.
Besides
the
timing
effect,
the
magnitude
of
initial
thermocline
surely
is
also
important.
This
initial,
surely,
is
stronger
in
the
peak
of
two
year,
El
Nino,
which
can
help
the
La
Nina
last
longer,
and
this
initial
family
only
is
also
related
to
the
preceding
stronger
El
Nino.
L
Besides
the
difference
in
the
oceanic
feedback,
we
also
observe
differences
in
the
surface
wind
variability
over
the
western
Pacific
between
one
year
to
year,
events
for
both
El
Nino
and
La
Nina
case.
We
observe
the
westerly
wind
decay
in
one
year
in
advance,
but
persist
into
the
following
year
into
year.
L
Events
and
the
ensemble
forecasts
reproduce
this
difference
to
understand
what
causes
the
different
wind
responses:
I'm,
showing
you,
the
spatial
patterns
of
sea,
surface
temperature
anomalies
and
surfaced
wind
anomalies
near
the
peak
of
one
year
into
year,
El
Nino
in
observations
the
two
year,
El
Nino,
showed
less
warming
over
the
Indian
and
Atlantic
Ocean,
and
also
there
is
a
sea
surface
temperature
warming
over
the
Northeast,
the
tropical
Pacific
into
year,
El
Nino.
But
this
warming
is
nearly
absent
in
one
year
El
Nino.
L
In
a
difference
map
we
see
that
this
Indian
and
the
authentic
difference
is
and
the
subtropical
warming
difference
persist
into
the
following
seasons
and
the
subtropical
warming
even
extends
to
the
western
equatorial
Pacific.
That
indicates
the
lower
of
North,
Pacific,
Meridian
or
mode
so
the
sea
surface
temperature
anomalies
in
the
Indian
and
Plenty
and
North
Pacific
Ocean
can
may
walk
together
to
influence
the
wind
variability
over
the
western
Pacific
and
they
help
to
maintain
the
westerly
wind
anomalies
in
two
year:
El
Nino
and
prolonged
the
El
Nino
events,
the
November
initialized.
L
The
forecasts
well
reproduces
this
sea
surface
temperature
anomalies
in
both
the
initial
States
in
the
subsequent
seasons.
So
oceanic
precursors,
outside
of
the
equatorial
Pacific,
also
contribute
to
the
high
predictability
of
ENSO
in
ventilation.
Similar
results
were
also
found
for
the
La
Nina
case.
D
E
L
Thank
you
for
the
question.
I
actually
have
a
slide
showing
that
we
found
that
this
El
Nino
decays
very
slow
in
the
western
Pacific
as
I
shown
earlier,
and
this
feature
is
even
pronounced
in
forecasts
initialized
even
earlier
that
is
associated
with
the
West's
disciplines
meant
of
this
sea
surface
temperature
and
wing
anomalies.
When
those
anomalies
are
shifted
to
the
west.
They
are
less
affected
by
the
delayed
feedback
from
the
Indian
Ocean,
so
that
slowed
down
the
termination
of
any
new
and
will
be
further.
L
D
M
M
So
what
yeah
I
think
I
forgot
to
tell
you
that
clicks
I'll
follow
the
damn
advice
that
I'll
say
some
of
the
joke
and
I
liam
like
you
are
laughing,
or
at
least
it
smiling
on
my
job.
So
I'll
go
through
the
three
main
questions.
So
can
we
develop
signal
to
the
idols
or
master
forecasting
system?
The
second
question
is:
why
can
we
develop
you,
that's
where
I
will
spend
most
of
my
time
and
then
how
can
you
movie
develop
me?
M
M
Okay,
so
going
to
the
first
question:
can
we
develop
the
signal
to
Decatur?
So
what
you
see
here
is
this
town
side
map
its
sowed
year,
one
to
year,
10,
swarm,
archer
predictability
in
the
dpla
forecasting
system
and
on
in
each
of
the
column
are
the
four
different
season.
The
thing
to
focus
here
is
the
color
bar
and
they
stipling.
If
there
is
a
stippling,
that
means
it's
statically,
significant
predictability
and
access
we
go
even
to
the
year
ten.
M
So
that
that
brings
to
the
answer
of
the
question
number
one:
yes,
we
can,
at
least
there
is
suggestive
evidence
in
the
DPLA
experiment
that
we
can
develop
seasonal
indicators.
Almost
a
forecasting
system
go
to
the
next
so
now
coming
to
the
Y
part,
so
there
should
be
a
baseline.
So
what
you
see
in
this
slide
on
the
Left
panel
is
the
same
slide.
That
I
showed
you
earlier
and
in
the
in
the
right
panel.
M
You
you
see
the
same
analysis
but
done
for
the
precipitation
forecast
and,
as
you
start
going
here
for
and
on,
but
in
the
precipitation
forecast
you
see
most
of
the
area
that
white.
That
means
there
is
no
stippling.
That
means
there
is
no
significant
predictability
in
the
precipitation
beyond
year,
four
or
so.
So
why
does
the
there
is
no
predictability
in
precipitation?
But
what
can
are
we
able
to
predict
the
soil
moisture?
So
this
is
exactly
the
same
system,
but
two
different
variable
I'm
I'm,
comparing.
M
So
we
will
look
into
a
few
other
experiment.
This
is
result
from
the
MF
type
experiment.
There
are
various
different
lines
but
main
thing
to
focus
on
the
black
line
and
the
blue
line
black
line
is
the
observation
and
blue
line
is
the
experiment,
MF
type
experiment
at
ten-member
and
sambal
in
the
top
panel,
you
see
the
precipitation
anomaly
in
the
Great
Plain
in
the
bottom
panel.
M
You
see
this
almost
an
anomaly
in
the
great
length
and
then
it's
comparison
of
the
MA
experiment
with
the
observation
and
what,
as
we
already
know,
that
Great
Plain
is
a
noisy
region.
So
that
means
we
have
very
less
predictability,
particularly
for
the
fresh
stage,
and
then
you
see
a
correlation
of
point
three
five.
M
That
means
roughly
around
ten
percent
of
the
precipitation
variance
can
be
explained
by
the
SST,
but
as
we
go
to
the
soil,
moisture
and
compared
with
similar
kind
of
in
detail,
in
this
case
e
d
di
and
then
the
correlation
in
created
two
point.
Seven
two.
That
means
almost
around
fifty
percent
of
the
variance
in
the
soil
moisture
can
be
explained.
So
the
finding
that
so
Alma's
there
is
a
more
predictable,
then
precipitation.
M
It
is
a
robust
across
different
experiment,
at
least
in
the
CSM
family
of
the
model,
so
coming
back
to
scientific
effect,
why
why
we
are
able
to
predict?
That's
almost
it.
Basically,
there
are
four
main
driver.
One
is
a
small
muscle
memory.
One
is
one
master
and
precipitation
coupling
or
that
mostly
looked
into
kind
of
first
component
of
this
almost
variability,
then
soil,
moisture
and
it
interacts
and
inform
master
and
run
of
interaction.
M
What,
if
you
focus
only
on
the
first
turn,
what
is
there
each
one
is
the
memory
component
there
Rho
s
T
and
s
t
minus
one,
but
second
one
is
the
variability
component
that
has
a
time
integral
minus
1.
That
means,
if
I
know
previous
month,
swordmaster
anomaly
we
can.
We
can
have
a
estimate
of
its
contribution
to
the
next
month
and
that's
a
very
important
factor
in
terms
of
predicting
this
all
myself.
M
So
what
we
can
do
it
we
can
integrate
this
equation
by
replacing
the
standard
deviation,
turn
with
the
anomalous
term
and
then
keeping
the
correlation
wrote,
ran
from
the
climatology
or
from
the
30
years
in
correlation
term,
and
we
integrated
this
term
for
the
2012
drought
and
what
we
found
that
the
first
term,
that
is
a
memory
term,
contributed
to
almost
50
57
percent
of
the
objects
for
master
anomaly.
Dealing
the
2012
loud
and
8
percent
was
from
the
precipitation
anomaly.
M
It's
not
to
say
like
precipitation
is
not
important,
but
here
main
point
is
precipitation
is
not
predictable.
What
what
you
see
in
the
bottom
panel?
There
is
a
soil,
moisture
prediction
in
the
middle
panel
and
then
the
precipitation
forecast
is
skill
in
the
again
third
column-
and
you
see
we
have
this
skill
in
this
hall
master
forecast
for
the
2012
grout,
but
we
don't
have
this
skill
for
the
precipitation.
M
There
are
other
processes
that
continues
to
the
to
the
psalmist
predictability.
One
of
this
plot
I've
talked
several
times
so
I'll
not
spend
too
much
time.
What
you
see
here
but
collects
is
the
soil
depth,
and
then
this
is
the
anomaly
propagation
from
the
1988
drought.
This
is
observation,
and
you
see
anomaly
persist
at
1
meter
like
heck
below
the
root
zone
for
a
year
and
longer
so,
and
that
can
contribute
to
the
psalmist
predictability
so
because
that
was
one
single
event.
M
So
we
looked
into
the
climatology
and
you
see
the
lag
autocorrelation
in
the
left
panel.
Top
is
4
for
moisture
and
bottom
is
for
the
precipitation,
as
we
expected,
like
autocorrelation
decays
with
the
time,
but
after
a
year
it
against
D
bounce
back
and
at
pure
selected
month
it
becomes
aesthetically
significant
that
a
top
layer
panel,
and
then
we
can
do
the
signal
analysis
for
all
the
season.
What
we
see
that,
yes,
the
correlation
decayed
and
then
again
it
comes
up.
That's
what
we
would
describe
as
a
potential
for
mastery
margin.
M
So
that's
another
processes
can
contribute
to
the
higher
predictability
of
this
all
muscle.
The
last
one
is
the
land
atmosphere,
coupling
and
the
vegetation
interaction.
So
what
you
see
here
is
the
difference
in
the
small
Marcel
residence
time
between
two
experiment.
One
is
couple
land
atmosphere
couple
and
another
is
land
atmosphere,
uncouple
experiment
and
on
the
Left
panel
you
see
the
BDC
on
the
right
panel.
M
So
to
answer
the
second
question:
yes,
so
al
mater
has
a
predictability
due
to
memory
due
to
anomaly:
reemergence
and
land
atmosphere,
coupling
and
vegetation
in
tracks-
and
these
are
three
plausible
physical
hypotheses
in
terms
of
what's
warm
water-
has
a
higher
predictability
going
on
to
the
most
difficult
part
of
my
presentation
is
how
can
we
develop
it?
So
we
know
some
physical
basis
why
we
expect
land
pro
contain
contribute
to
the
predictability,
so
we
tested
in
that
case
yes,
M,
D,
ple
prediction
and
what
particularly
I
want
you
to
pay
attention
to
this
line.
M
There
is
a
magic
number
before
they're,
so
that
34,
that
isn't
omble
member
which
was
used
to
do
the
land
initialization.
So
there
is
anomaly
in
before,
and
we
correlated
with
the
forecast
anomaly
and
what
you
see
when
you
use
the
34
member
and
tumble
anomaly,
it
says
it
gives
you
significantly
higher
predictability
than
the
remaining
39
large
ensemble
initial
condition
that
so,
like
a
land,
contributes
to
the
predictability.
M
The
total
predictability
are
shown
by
by
this
solid
line.
So
what
we
find
is
land
contributes
almost
30%
of
the
total
predictability,
so
land
in
initialization
is
required.
You
know
addition
to
the
ocean,
then
the
second
one
needs
so
most
of
the
or
DPL
II
was
initialized
is
never
first.
So
when
we
did
seasonal
probability
analytics,
we
find
there
is
a
strong
seasonal
cycle,
and
so
we
go
ahead
and
run
the
frequency
analysis
and
we
find
that
is
exactly
like
a
tenth
cycle.
M
In
in
that
10
years
forecast,
so
that's
where
it
picks
up,
particularly
for
this
Walmart
area,
this
poverty
aesthetically
significant.
So
what
at
least
to
my
session
is
at
least
run
at
least
one
more
initialization,
that
is
in
the
May,
a
collage
than
some
of
the
experiment
back
in
2013,
so
they
we
can
look
into
their
experimental,
see
like
what
is
the
exact
that
we
want
to
run
it.
And
finally,
this
is
Yahoo
difficult
part
of
the
presentation.
M
This
is
what
we
are
envisioning
for,
the
next
generations,
our
master
forecasting
system,
I,
don't
know
how
much
time
I
have
it.
This
is
complicated
figure,
but
what
a
key
key
component
there
is,
let
climate
model
do
their
best
and
let
Big
Data
to
the
list.
So
we
are
trying
to
combine
both
the
recent
at
once
in
our
system,
modeling,
with
more
like
what
artificial
intelligence
or
the
big
data
techniques,
including
deep
learning,
and
although
I
made
these
figures
were
kind
of
a
conceptual
figure
very
easily.
M
D
Thank
you.
We
have
about
two
minutes
for
questions.
So
if
you
have
a
question
on
IC
from
Iowa
for
the
DPLA
skill
in
soil
moisture,
what
observations
are
you
comparing
again
and
since
the
Deepthi
le
land
is
initialized
from
a
freerunning
simulation?
Is
the
predictability
you
described
coming
from
external
forces,
unknown.
M
D
D
M
That
watch
for
the
JJ
and
we
use
the
monthly
data
to
the
way
I
wrote
the
equation.
What
I
do
I
take
the
correlation
and
then
take
the
previous
month
of
this
all
monster
anomaly
to
integrate
equation,
but
I
thought
it's
a
based
on
yeah
monthly
in
my
type
experiment,
as
well
as
the
previous
month
as
Walmart
said
anomaly
in
that
my
EMF
type
experiment,
yeah
I,
don't
know
if
I
exactly
answered
the
question
that
was
did
a
good
and
difficult
question.
Alright,.
D
D
B
B
Right
I
thought
it
would
be
useful
during
this
first
meeting
of
the
ESP
working
group
to
give
I
guess
an
overview
of
the
lay
of
the
land
as
far
as
takeo
prediction
at
CES,
M
and
I'll
start
off.
With
the
caveat,
this
is
my
own
personal
perspective
and
I
apologize
in
advance
for
any
omissions
or
oversimplifications
of
other
people's
work
in
his
presentation.
So
at
this
point
our
main
detail
prediction
data
set,
as
we've
heard,
is
the
CSM
dko
prediction
large
ensemble.
This
was
presented
to
the
community
in
a
Bands
article
in
September
2018.
B
It
represents
the
latest
Gale
prediction:
data
resource
for
the
CSM
community.
It's
about
a
26,000
simulation
year
experiment.
So
a
very
expensive
undertaking
took
about
upwards
of
40
million
core
hours
to
complete
this
experiment.
So
this
is
not
something
that
we
will
be
doing
often
and
it's
very
hard,
obviously
to
do
sensitivity,
experiments
off
of
an
experiment
like
this.
It
represents
our
contribution
to
the
dCPP
of
cnips
6.
B
So
as
we
extend
it
further
and
further
in
the
future,
we're
less
and
less
confident
in
realism
of
that
ocean
and
sea
ice
simulation,
the
data
are
available.
Here's
the
link,
you
can
also
get
it
from
the
earth
system
grid,
Federation
and
I'd
like
to
point
out.
This
was
made
possible
by
multi
agents
and
support.
Noah
Dooley
and
Anna
staff
and
I've
presented
many
times
results
from
this
overview
paper.
B
Do
you
just
thought
I
would
cursory
takeaway
points
from
that
from
that
paper,
which
is
that
we
see
high
global
de
kado
predictions
for
SST
and
service
air
temperature
pretty
much
ubiquitous
over
the
globe?
But
much
of
that,
of
course,
is
attributable
to
external,
forcing
and
the
skill
improvement
over
the
uninitialized
ensemble,
which
is
also
a
large
40
member
ensemble,
is
most
evident,
4d
trended
variability.
B
There's
clear
evidence
of
surface
air
temperature,
precipitation
skill
enhancement
from
initial
ovation,
for
example,
summer
precipitation
over
the
Sahel,
but
there's
a
complex
dependence
on
recent
region
season
and
lead
time.
We
showed
that
the
large
ensemble
size
enhances
the
skill
and
we
showed
evidence
of
the
signal-to-noise
paradox
in
the
new
system,
precipitation
predictions,
and
so
this
talk
will
be
about
asking
the
question:
what
more
have
we
learned
about?
Csm
D
ple?
B
And
so
all
this
work
I
think
highlights
the
value
of
thoughts,
the
initialization
for
exploring
potential
predictability
of
poorly
observed
ocean
fields,
and
these
are
fields
like
ocean
budget
chemistry
for
which
data
assimilation
may
not
be
a
viable
option
for
initializing
the
system.
This
terrestrial
work,
also
in
particular,
highlights
the
need
to
upgrade
to
CLM
5
and
then
train
the
land
modeling
community,
with
a
model
that
they're
interested
in
looking
at
and
exploring
realistic
land
initialization
in
the
future.
Csm
initialized
prediction
methods,
including
the
decade
or
efforts.
B
Another
important
paper
that
came
out
was
led
by
lo
Simpson.
She
focused
on
low
frequency
variability
of
March
precipitation
over
Europe
in
particular
Portugal,
so
she
identified
a
strong
observed
link
between
multiple
yield
fluctuations
in
North,
Atlantic
jet
and
manic
assets
T,
and
she
used
that
to
show
that
it
combined
any
animal
statistical
prediction
of
late
winter.
European
precipitation
can
be
skillful
at
the
cable
lead
times,
even
though
the
dynamical
DP
OE
system
itself
is
not
that
precipitation.
B
So
you
can
see
my
cursor
I'm,
pointing
at
the
raw
precipitation
skill
for
low
frequency,
precipitation
of
the
UK
and
blue
here,
it's
quite
low
foodie
Kelley.
If
you
use
the
the
skilful
prediction
of
the
ple
SST
combined
with
this
statistical
relationship,
then
we're
able
to
achieve
really
impressive
skill
scores
for
both
the
UK
and
Portugal
and
so
they've.
This
work
demonstrates
the
potential
of
the
hands
prediction
system,
skill
and
circumvent
model
air
through
novel
statistical
approaches,
and
it
hints
at
an
anemic
atmospheric
response
to
skillfully
predicted
SST.
B
B
A
paper
that
has
just
come
out
in
climate
atmospheric
science
demonstrates
that
the
dpla
system
has
high
skill
at
predicting
winter,
blocking
frequency
and
nao
in
the
North
Atlantic
sector.
So
the
top
left
plot
here
showed
the
time
series
blocking
a
blue
region
on
the
upper
right.
This
is
diagnosed
from
Mali
500
and
then
smooth
isolate
variability.
The
blue
curve
is
from
observations
and
sat.
Bank
are
the
analysis
and
then
the
the
dashed
red
curve
is
the
meteor
one
through
eight
prediction
from
the
dpla
system.
B
Look
at
that
generates
a
skill
Explorer
of
0.65,
which
can
be
enhanced.
You
close
to
0.9
by
additional
temporal
smoothing,
are
averaging
over
a
window
of
different
honk
at
start
gates,
and
so
this
high
school
for
Atlantic
blocking
is
mirrored
in
high
school
Floridian
a
harrowing
on
the
bottom
left.
We
get
a
skill
score
for
winter
time
in
a
year
o
of
0.85.
After
doing
this,
additional
smoothing,
so
this
is
I
would
say
unprecedented
scale
for
an
individual
prediction
system.
B
This
paper
showed
the
the
scale
for
both
high
level
blocking
and
nao
is
strongly
sensitive
to
the
ensemble
size.
So
these
different
lines
here
show
different
lead
year.
Combinations
and
the
top
most
skillful
curve
is
highlighted
by
the
X
on
the
left
hand
side.
So
the
highest
skill
curve
years
for
the
leader
1
through
8
for
any
other
3
year,
2
through
8
way,
she's
that
the
skill
increases
with
ensemble
size
that,
before
it
hasn't
saturated
in
40,
since
potential
for
even
higher
still
throw
even
one
ensemble
members
at
this.
B
It
also
points
out
the
signal-to-noise
issues
in
the
system.
So
we
look
at
the
predictability
of
the
model
itself
is
shown
by
blue
it's
much
much
lower
than
when
we
verify
against
observations.
So
there's
a
much
lower,
predictable
variance
component
and
dpo
lead
in
an
observation.
That's
consistent
with
the
signal-to-noise
paradox
literature.
So
this
work
highlights
the
potential
for
higher
for
each
in
skill
by
enhancing
a
signal,
variance
through
model
improvement
or
just
by
increasing
the
ensemble
size,
a
paper
that
I've
written,
it's
that's
going
to
be
coming
out
in
climate
dynamics.
B
I
think
will
show
the
origins
of
our
detailed
predictions
Silliman
scale
in
the
Atlantic
sector.
What
I'm
showing
on
the
right
here
is
our
skill
for
a
persian
heat
content
so
for
North
Atlantic
Black
is
the
Fossey,
and
this
is
a
potential
predictability
study
asking.
How
are
we
able
to
achieve
this
high
skill
when
we
compare
the
red
curve
to
the
black
list,
even
out
to
leave
your
5
through
9?
B
So
the
upper
right
hand
blot
here
in
color,
shoot
the
Labrador
sea
water
thickness
in
late
fall
of
1999
and
the
contours
show
the
sea
surface
height
anomalies
that
are
basically
driven
from
below
by
this
Labrador
sea
water.
Anomaly,
giving
rise
to
this
very
strong
surface
flow
into
the
eastern
central,
so
polar
dire.
And
so,
if
we
look
at
a
time
series
averaged
over
this
mid-atlantic
red
fox
of
laven
or
sea
water
thickness,
you
can
see
that
the
DP
le
does
an
amazing
job
of
predicting
this.
B
Here's
from
satellite
records
with
the
variability
in
surface
geostrophic
flow
over
this
mid-atlantic
ridge
region
was,
and
you
can
see
that
the
model
isn't
perfect
clearly,
but
it
seems
to
be
getting
the
right
sort
of
behavioral
fluctuation
in
surface
effective
heat
convergence,
so
this
work
suggests
that
unobserved,
deep
ocean
variations
are
key.
This
successful
be
able
prediction
and
highlights
the
value
of
fossa
visualization
for
understanding
predictability
mechanisms,
so
I'm
gonna
just
blaze
through
this
there's
a
new
paper
coming
out
by
Hannah
Christensen
in
general
climate.
B
That
demonstrates
state-dependent
predictability
for
our
ocean
heat
content
on
our
planet,
using
the
relative
entropy
metric,
and
it
showed
that
for
some
started
years,
information
content
actually
increases
with
lead
time.
So,
looking
out
for
that
paper,
which
will
be
coming
out
soon,
another
paper
I
want
to
highlight
is
a
multi
model
paper
that
Nick
Dunne
stone
from
UK
Met
Office
has
coming
out
in
Environmental
Research
Letters
and
in
this
he's
combined
the
depresses
3
and
the
DPLA
system,
which
are
both
40
member
systems
and
looked
at
inter
annual
climate
prediction
spill
over
week
months.
B
One
16
and
the
takeaways
are
that
the
direct
comparison
with
a
higher-resolution
depressive
system
shows
the
DPLA
has
quite
comparable
skill,
but
somewhat
lower
than
depresses,
and
it
also
highlights
the
initialization
shock
that
we
have
in
DPLA.
If
we
look
at
the
lead
months,
7
through
11
prediction,
monsoon
prediction
precipitation
over
North
Africa,
let's
see
the
DPL
II
has
negative
skill
or,
as
depressants
got
positive
skill,
so
in
suggesting
that
early
lead
times,
we've
got
some
shock
behavior
that
is
diminishing
our
skill
in
our
prediction
system.
B
B
She
showed
that
actually
got
negative
skill
in
the
tropical
Pacific
that
early
lead
times
and
then
high
engine
followed
that
up
in
2017,
with
an
analysis
that
same
system
showing
that
for
early
start
years
we
tend
to
get
predominance
with
ENSO
events
and
later
years,
La
Nina
events
that
is
giving
us
this
very
wonky
year.
In
so
evolution,
that's
need
to
be
corrupting
spill
more
widespread.
B
We
look
at
SS
to
scale
and
an
Eastern
tropical
Pacific,
it's
low
or
negative,
and
as
we
go
forward
and
lead
time,
it
gets
better
and
better
and
we
can
see
that
more
clearly.
If
we
look
at
the
the
skill
difference
for
different
Penta
relative
to
media,
one
through
five
I
can
see
that
skill
actually
increases
as
we
both
Lord
and
lead
time.
B
And
so
this
is
an
issue,
that's
still
present
in
our
system,
and
it
might
explain
these
sorts
of
phenomena
like
the
mast
of
Sahel
precipitation
scale
of
meteor
three
through
seven
rather
than
earlier
on
eviction
system.
If
we
could
correct
the
shock
we
might
be
able
to,
you
know,
get
a
more
well
behaved
system
that
should
hire
steel,
that's
relieve
times
so
very
quickly.
B
Closing
thoughts,
the
DP
le
I
think,
at
least
from
my
perspective,
continues
to
exceed
our
expectations,
and
the
preliminary
work
suggests
that
at
the
top
to
your
contribution
to
dCPP,
recent
analyses
contributed
greatly
to
our
understanding
of
DP
le,
but
more
scrutiny
by
the
community
is
needed,
so
please
download
it
and
explore
this
data
set.
Another
point
I
want
to
make
is
that
successful
prediction:
our
high
school
yields
high
profile
papers,
but
to
advance.
B
We
need
to
develop
deeper
understanding
of
prediction
system
behavior
in
general,
both
moving
bad
and
we
need
efficient
ways
to
test
new
methods.
How
do
we
share
negative
or
disappointing
results
that
exposed
prediction
system
failure,
and
so
here
I've
listed
some
key
challenges
for
advancing
esmd
Kaila
prediction
and
its
amount
of
time
I'll
stop
there.
D
Since
we
are
over
time
and
I
know
that's
sort
of
virtual
meeting
sitting
inside
of
a
computer
for
a
while,
everybody
probably
needs
a
break.
This
point
so
I'm.
Let's
take
our
break
we'll
come
back
at
ten
thirty.
If
you
have
questions
for
Steve,
you
can
put
them
into
the
Google
document
or
the
chat
window
and
we'll
make
sure
that
we
follow
up
with
those
questions
at
a
later
point.
So
take
a
break
and
returning
at
ten
thirty.
D
Everyone
welcome
back
to
the
earth
system
prediction
working
group
meeting.
It's
10:30
and
I
want
to
keep
this
all
on
schedule,
so
we're
going
to
start
with
our
next
presentation,
but
before
that,
I
just
want
to
say
that
I'm
really
excited
and
impressed
that
our
very
first
inaugural
meeting
and
hi
early
virtual
I've
seen
a
high
number
of
participants
on
Zoom
at
97
participants
go
whoo-hoo.
D
D
All
right
and
also
I'll,
thank
you
for
I
noticed
during
the
break
that
people
are
putting
questions
it's
the
Google
Docs,
so
please
continue
doing
that
and,
if
you're
one
of
the
speakers,
please
take
a
look
at
those
questions
and
try
to
answer
respond
to
the
questions
that
were
put
in
there.
Our
presentation,
this
presentation
is
Riley,
Brady
and
he'll,
be
talking
about
the
climb
for
a
diagnostic
packet.
J
All
right
so
today,
I'll
be
talking
about
climp
red,
which
is
a
diagnostic
package
that
I
have
co-developed
with
Aaron
Spring
who's.
A
graduate
student
at
MPI
I
mean
the
package
is
briefly
described
as
an
x-ray
wrapper
for
analyzing
and
sample
forecast
models
for
climate
prediction,
I'm
going
to
go
over
the
motivation
here
and
share
with
you
guys
some
of
the
API
and
then
hopefully
encourage
some
some
new
users
to
join
us.
J
Where
we've
got
these
massive
databases
like
DP,
le
the
dCPP
from
cf6,
sub,
X
and
then
and
so
on.
So
we
have
this
unprecedented
volume
of
data
to
analyze,
and
the
vision
here
is
that
if
we
can
unite
around
this
open
source
framework
for
analyzing
predictions,
we
can
spend
more
time
actually
answering
these
scientific
questions
about
predictability
and
less
time,
potentially
writing
and
rewriting
each
other's
code.
J
So
I
want
to
briefly
introduce
the
two
main
packages
that
were
leveraging
here
with
Clint
bread
if
you're
unfamiliar.
This
is
sort
of
in
the
framework
of
this
pan
geo
team.
So
there's
a
website
can
GI.
Oh
it's,
this
community
of
geoscientists
that
helped
to
develop
and
support
tools
that
are
that
are
helpful
for
big
data
Geoscience
analysis.
The
main
package
we
wrap
is
both
x-ray
and
it's
this
wonderful
on
package
that
allows
you
to
do
easy
analysis
on
labelled
multi-dimensional
arrays
of
data.
So
it's
it's!
J
A
label
of
where
operation,
so
you
can,
you
can
slice
and
call
the
actual
dimensions
instead
of
worrying
about
the
order
of
them
and
having
to
know
the
index
and
worrying
about
whether
or
not
you're
in
a
1
or
0
based
index
language
and
the
other
main
package.
We
leverage
is
called
desk,
which
is
this
pure
Python
package.
It
allows
for
out
of
memory
and
parallel
computation,
so
I
actually
use
this
to
run.
J
You
know:
5
5,
nodes
and
workers
on
Kasper
I
can
have
a
terabyte
of
memory
and
hundreds
of
workers
loading
in
chunks
of
data
into
memory.
So
you
can
work
even
on
a
hundred
big
node.
You
can
work
with
a
terabyte
of
data,
so
this
you
can
see
will
help
out
with
things
like
bootstrapping
in
these
high
resolution
prediction
and
samples
that
are
highly
dimensional.
J
So
I
want
to
give
you
some
intuition
into
how
we
structured
this.
This
architecture
for
clim
bread,
so
the
main
thing
you'll
see
as
a
user
is
we
offer
these
two
high
level
objects
called
a
hike,
a
stencil
and
perfect
model
ensemble
and
Aaron's
leading
the
perfect
model
side
of
this.
But
we've
really
divvied
up
the
initialize
prediction
world
into
these
two
bins
where
the
I
guess
example
is
something
that's
initialized
from
an
observation
like
product
and
the
perfect
model.
J
Inaudible
briefly
is
initialized
from
control
around,
but
it's
sort
of
this
contained
environment
with
one
model,
the
former.
You
want
to
actually
verify
against
observations
in
that
ladder,
you're
really
just
doing
a
theoretical
framework
and
verifying
against
itself,
but
really
the
way
we
divvy
this
up
is
that
the
my
constants
on
both
you're,
comparing
differently
you're
taking
in
sample
means,
are
looking
on
a
member
to
member
basis.
Pursues
your
observations.
Where's
the
perfect
model,
Istanbul,
there's
a
lot
of
ways
to
construct
these
super
vectors
of
different
ensemble
member
pairs
to
compare
them
to
each
other.
J
J
The
the
other
thing
that
we
think
about
with
with
Clint
bread
is
because
we're
using
x-ray.
We
can
anticipate
certain
dimension
labeling,
so
any
any
prediction
system
that
comes
in
through
x-ray.
We
we
check
and
guide
you
towards
labeling
into
this
framework,
where
you
have
this
init
dimension
here,
which
labels
you
know
the
date
it
was
initialized.
This
can
be
integers
of
the
year
or
it
could
be
actual
date
strings
down
to
the
to
the
minute.
J
We
also
look
for
this
lead
time
to
mention
we're
not
working
in
that
frame
where
you
sort
of
have
this
time
axis
with
target
dates.
We
convert
that
over
to
a
lead
axis,
for
you
to
say
we're
this
many
integer
time
steps
away
from
your
initialization
and
then
we
can
read
this
netcdf
units
attribute
that
you
depend
on
to
know
what
how
to
translate
those
lead
times.
J
Does
it
years
seasons
months
weeks
spent
as
days,
and
so
that's
how
we
take
your
initialization
date
and
we
use
that
information
to
shift
it
and
align
it
with
with
your
verification
product,
and
then
we
can
also
ingest
a
member
to
mention
if
you're,
using
probabilistic,
metrics
or
if
you
want
us
to
take
an
sample,
mean
for
you
and
then
any
any
additional
dimensions
like
spatial
dimensions
that
are
tacked
on
get
carry
along
for
the
ride
and
it's
fully
vectorized.
It
gets
broadcast
on
the
desk.
J
The
last
sort
of
high-level
thing
I
want
to
show
you
before
we
jump
into
what
this
looks
like
is
that
I
spent
a
lot
of
time.
Thinking
about
forecast
alignment,
it
seems
like
different
communities,
use
different
strategies
to
actually
select
the
initializations,
so
we
offer
this
keyword
now
or
you
can
select
how
to
align
your
initialization
in
day
time.
You
can
go
with
this
same
various
keywords
so
that
you're
you're,
finding
a
common
verification
window
at
all
leads
citizen
obviously
depended
on
the
observation
you
give.
J
Our
group
works
a
lot
with
ocean
biochemical
data,
so
we
tend
to
have
sparse
data
sets.
So
these
these
decisions
really
influence
our
skill.
Another
alignment
strategy
is
to
use
a
common
set
of
initializations
that
can
verify
across
all
leads.
And,
lastly,
this
maximize
keyboard
just
says
that
each
lead,
we're
gonna,
select
all
available
initializations
that
can
verify
what
the
observations
provided.
Maybe
this
seems
like
over
the
top,
but
when
you
actually
use
these
different
strategies,
you
get
wildly
different
skill,
so
this
is
just
verifying
the
DPL
ii
SST
local
time.
J
Series
on
the
left
is
trended
and
on
the
right,
that's
d,
trended
and
we're
verifying,
except
that
forced
ocean
sea
ice
product.
The
different
colors
here
show
the
different
alignment
strategies
where
the
y-axis
is
the
anomaly
correlation
coefficient.
Unsurprisingly,
the
maximum
degrees
of
freedom
gives
you
the
best
skill,
but
surprisingly,
depending
on
whether
you're
trended
or
d
trended
deciding
between
common
and
hits
and
common
verifications
changes
the
scale
and
the
comment:
it's
is
something
that
by
design
is
used
in
the
perfect
mountain
world,
for
instance.
J
So
now,
we've
got
all
that
background.
We're
in
the
Clint
pred
space
we've
opened
a
juber
notebook.
We
loaded
in
x-ray
and
I've
loaded
in
the
CSM
DPL
the
output.
This
is
from
the
dCPP
directory.
So
this
is
a
version.
That's
regraded
onto
one
by
one
degree,
grid
I
haven't
done
any
bias,
correction
or
anything
and
I've.
J
Also
loaded
in
the
fossee
product
for
validation
also
won
a
one
by
one
degree,
global
grid,
but
simply
we
just
loaded
that
high-level
hindcast
in
Samba
object
and
Flint
I
simply
call
my
guests
and
sample
on
that
DP
le
data
set
and
it's
loaded
in
we
do
a
lot
of
checking
under
the
hood,
make
sure
you
have
those
dimension.
Labeling
make
sure
the
coordinates.
Look
good
check
the
time
and
we
have
a
lot
of
intelligent
error
messages
to
help
you
out
there
and
a
lot
and
a
lot
of
examples.
J
The
documentation
online,
which
I'll
show
the
link
at
the
end,
and
then
we
can
just
append
on
these
observations
by
using
this
observations,
method,
I
passed
in
Fosse
and
I,
give
it
this
aesthetic
name
and
no,
you
can
add
any
arbitrary
number
of
observations.
So
the
nice
thing
here
is
we
can
carry
along
three
verification
products
and
when
we
run
the
verifies,
the
verify
method
is
good
verify
against
all
three
in
parallel.
J
So
you
can
get
a
sense
of
sort
of
uncertainty
due
to
your
verification
product
and
the
nice
thing
is
because
we're
working
with
x-ray,
we
get
this
nice
HTML
representation.
So
you
know-
and
what's
going
on
under
the
hood,
so
I
print
out
my
hindcast
in
sambal
and
on
the
top
row
here
I
can
see.
Okay,
this
is
my
initialized
EP
le
here's,
my
dimension
names
I
can
see
all
the
coordinates,
and
this
is
HTML,
so
you
can
actually
drop
down
and
look
at
inspect
some
of
the
coordinates.
J
The
attributes
will
be
carried
out
long
if
there
are
any
and
then
I've
got.
This
excuse
me
bossy
data
set
at
the
bottom,
with
the
coordinates
as
well.
So
you
get
a
sense
of
what
time
period
we're
covering,
and
you
also
see
the
variables
if
there's
multiple
variables
here,
it'll
it'll
verify
multiple
variables
and
in
synchrony.
J
Now
that
we've
got
a
sense
of
the
title
of
you,
what
does
it
look
like
to
actually
verify
forecast?
Well,
this
is
where
you
spend
all
of
your
time.
Writing
these
loops
hand.
Writing
your
metrics
trying
to
line
the
daytime
and
checking
and
rechecking
the
indices,
making
sure
they
look
right
here.
We
just
run
up
verifying
climp.
It
will
handle
the
rest
under
the
hood
and
here's
all
those
design
decisions.
J
Erin
recently
said
nobody
should
get
their
skill
for
free
because
we
used
to
have
a
few
of
these,
but
you
know
set
as
default
and
we're
switching
to
a
workflow
where
you
have
to
you
have
to
you
have
to
pass
in
each
of
these
keywords
so
that
you
know
what's
going
into
still
computation.
So
here
I
pick
an
anomaly
correlation
coefficient
for
my
metric
maximize
the
degrees
of
freedom
for
my
alignment,
my
comparisons
gonna,
be
the
in
sambal,
mean
two
observations
and
I
so
on
a
persistence.
J
Reference
forecast
and
I'm
running
this
on
five
nodes
on
Casper
and
you'll.
Note
I
verify
1010
late
years
for
the
entire
global
grid,
180
by
360
and
less
than
three
seconds.
So
this
is
the
performance
we're
really
aiming
for
that.
We
spent
a
lot
of
time
working
on
so
the
verify
stuff
really
scales.
Well,
so
you
can
you
can
flip
out
metrics
and
not
wait
minutes
two
hours
to
get
a
global
grid
of
this
and
again
you
get
this
nice
HTML
representation
of
that
output
and
we've
got
this
appended
skill
dimension.
J
So
you
can
flip
between
your
initialize
skill
and
that
persistence.
The
skill
from
your
persistence
forecast
we're
working
on
adding
more
references.
You
can
also
have
an
uninitialized
reference
like
the
CSM
le,
but
we're
planning
on
adding
Dan
persistence
and
climatology,
and
things
like
that,
and
then
you
can
leverage
the
rest
of
the
Python
Panji
ecosystem
to
make
nice
publication.
Writing
figures.
I
will
note
before
before
Steve's
scared
about
what
I'm
showing
here.
This
is
just
straight
from
dCPP
I
didn't
buy
his
correct
drift,
correct
pushed
anomalies.
J
I
just
wanted
to
really
show
this
workflow,
so
I'm
just
plotting
Rio
output.
So
this
just
this
is
kind
of
arbitrary,
but
you
can
see
since
we're
working
in
an
x-ray
framework.
We
can
very
quickly
and
easily
plot
these
sort
of
things,
so
here
we're
looking
at
scale
on
the
top
row
and
persistance
on
the
bottom
row
that
that
came
out
of
this
one
print
verify
command
and
then,
because
you
know,
we've
loaded
it
with
all
of
these
different
customizable
features.
You
just
switch
out
the
metrics
string,
which
they're
all
available
in
the
docs.
J
So
now
we
switch
it
to
normalize,
mean
absolute
error
and
rerun
it.
It's
done
in
less
than
three
seconds,
and
we've
now
got
anim
you've
globally
again.
The
colors
here
are
really
meaningful,
because
I
didn't
post
process
the
data,
but
you
can
see
how
so
easy
is
this
version
of
the
metric,
and
we
have
28
metrics
available
right
now,
both
deterministic
and
probabilistic.
J
The
last
thing
I
want
to
touch
on
is
an
important
topic
in
the
field
which
is
bootstrapping
and
significance
testing
right
now
me
we
have,
you
know
the
easiest
way
to
do.
It
is
through
that
metric
argument.
I
just
showed
so
you
can
pass
in
Spearman
and
Pearson
p-values
and
actually
get
the
P
significance
at
each
grid
cell
and
then
use
other
x-ray
tools
to
get
field
significance
after
and
you
can
stipple
that
way.
So
this
is
a
purely
sort
of
temporal.
J
J
J
This
deep
ele
SST
global
grid
and
this
this
basi
global
grid
they're
chunked
per
desk,
and
this
is
on
a
5
by
5
degree,
global
grid,
which
is
what
Steve
uses
in
his
DPLA
paper,
and
this
500
means
I'm
doing
500,
bootstrapping
iterations,
so
we're
essentially
resampling
with
replacement
the
in-sample
members
in
the
hindcast
and
then
computing.
The
scale
500
times
and
we
can
do
that
on
five
nodes,
I'm
Casper
in
about
12
seconds-
and
this
is
something
that
takes
a
horrendously
long
amount
of
time.
Normally,
when
you
put
your
hand,
writing
it.
J
So
this
is
where
ramming
4
is
this.
You
know
minute
minute,
you
know
order
minute
timescale
for
bootstrapping,
and
here
I
returned
the
five
and
95%
quantiles,
but
you
can
imagine
also
returning
significance
based
on
you
know
and
how
many
iterations,
the
boss
skill,
beats
out
the
hindcast
skill,
for
instance.
J
So
we
can
really
do
hundreds
of
thousands
of
iterations
of
bootstrapping
in
a
timely
manner,
I'm
working
on
an
S
us
project
this
summer,
so
we're
gonna
really
hope
to
have
a
more
robust,
s2s
interface
here,
where
you
can
work
with
arbitrary
timescales
on
the
stuff
daily
level,
even
we're
also
working
on
refactoring
the
code
base.
So
if
you're
interested
in
contributing
to
code,
this
will
make
it
a
lot
easier
to
help
us
develop
it
and,
lastly,
we're
gonna
push
for
a
Jaa's
publication.
J
This
fall
so
look
out
for
all
of
this
and
I'm
gonna
end
with
this
slide.
This
shows
you
the
documentation,
which
has
a
ton
of
information,
the
codes
open
source
on
github.
We
encourage
you
to
try
to
try
and
post
issues
and
reach
out
to
us
with
email
for
ideas,
questions
support,
but
if
you
want
to
help
we're
really
eager
to
have
a
community
involved.
So
thanks.
So
much
for
having
me
in
this
session
is
exciting
to
talk
about
software
Thanks.
F
Okay,
so
I
will
start
off
the
discussion
on
this
and
I'll
hand
it
over
to
Steve.
So
these
are
simulations
that
were
already
planning
a
10
card
that
we
have
acquired
computing
allocations,
and
after
that
we
can
start
the
discussion
on
what
else
would
you
like
to
see
if
we
got
an
allocation
for
the
working
group
after
November,
so
we're
already
planning
following
on
the
CSM
1s
2s
5s
20s
m2,
so
these
will
be
very
similar
to
what
I've
done
with
CSM
one.
F
However,
we
would
like
to
see
this
as
a
community
project,
so
we
would
like
to
have
contributions
and
analysis
from
the
broader
community
and
we'll
plan
to
do
is
just
provide
the
baseline
simulations
and
an
additional
simulations
could
be
carried
out
later,
let's
say
with
perturbed
physics
or
changes
to
initial
conditions,
etc.
So
the
set
up
we're
setting
up
is
just
default.
F
Csm
to
the
32
level
standard
atmosphere
model
we
will
do
weekly
Monday
starts
from
1999
to
2019
right
now,
we're
planning
for
11,
ensemble
members
and
again
that's
something
we
can
discuss
in
a
discussion.
If
there
was
a
desire
to
have
more
members,
that's
something
we
could
write
into
the
ESP
WG
computer
allocation
for
the
atmosphere,
we're
going
to
switch
from
era
interim
to
CFS,
be
to
just
because
the
data
is
available
up
to
date.
F
We
already
have
some
community
interest
in
looking
at
MgO
and
tropical
cyclones,
where
we
would
there's
so
many
other
things
to
Lokhande.
We
would
really
like
to
help
so
just
send
me
an
email
or
we
can
start
putting
names
in
a
Google
Doc.
While
we
talk
if
you're,
interested
and
output
is
quite
comprehensive
and
we
have
a
list
of
variables
that
were
outputting
on
3d
levels
and
we're
adding
land
ocean
output
as
well
as
atmospheric
output
and
so
Longley
and
I
think
it's
extremely
comprehensive.
F
But
if
there
is
something
specific
you
would
like
to
look
at
and
send
me
an
email
and
I
can
connect
you
with
the
current
document
and
the
issues
come
to
me.
We
want
to
kick
off
these
hindcast
as
soon
as
possible,
hopefully
in
the
next
couple
of
weeks.
So
please
send
me
output
suggestions
by
the
end
of
the
day
on
next
Monday.
That
would
be
great
I
mean
again.
I
can
share
with
you
what
we
already
have
planned,
okay
and
then
I'll
cane
it
over
to
Steve
are.
B
F
B
Right
so
the
Earth
System
prediction
group
within
cgb
submitted
and
car
strategic
capability
for
compute
proposal
in
the
spring
time
entitled
multi-year
earth
system
predictions
and
some
ESM
to
the
idea
here
is
that
we
would
use
the
new
model
at
nominal
one-degree
resolution.
That's
cam,
six
pop
to
CLM
5
and
size
5
with
Craig
prognostic
ocean
biogeochemistry
and
run
20
member
honk
asts
initialized
four
times
per
year
over
the
time
period
1970
through
2019
and
the
simulation
length,
whose
hindcast
would
be
three
years
in
the
proposal.
B
We
thought
that
we
would
basically
do
an
all
J
or
a
55
initialization,
so
we
can
use
Jerry
55
rate
analysis
for
the
atmosphere,
initial
condition
use
our
JRA
do
force
pop
pop
and
size
5
simulation
for
ocean
and
sea
ice,
and
this
is
our
contribution
to
the
omote
to
protocols
and
mick
6,
and
then
the
land
initial
condition
remains
a
bit
of
a
question
and
I'm
not
an
expert
on
it.
But
I'm
curious.
If
there
are
any
experts
who
would
know
whether
we
could
do
a
offline,
JRA,
55
forced
spin
up
with
CLM
5.
B
To
begin
these
simulations
again,
we
need
to
finalize
the
land
and
initial
conditions,
and
you
know:
I
saw
Cecilia
bits
as
comment
on
the
for
simulation
and
sea
ice
and
our
DP
le
initialization
and
so
I
guess.
There's
a
question
of
you
know:
what
exactly
is
the
fidelity
of
the
CI
simulations
to
simulation?
That
we'll
be
doing?
We
want
to
take
a
closer
look
at
that
and
we'll
revisit
the
Neptune
simulation
and
an
option.
That's
up
for
discussion.
B
B
D
E
B
F
B
F
B
M
I
think
one
need
regarding
they
Steve
that
you
asked
for
around
forty
four
million
and
they
gave
you
18
million,
so
I
guess
what
I
wanted
and
I
might
have
it.
Can
you
hear
me
yeah
yeah,
so
instead
of
having
for
initialization
month
like
those
are
targeted
around
for
season,
we,
you
may
think
about
having
the
two
initialization
month
like
summer
versus
winter
or
a
spring,
but
yet
lake
fault
and
as
we
progress,
if
we
get
some
more
computer
time
like,
then
we
can
do
two
more
season.
M
The
comment
that
I
have
it
from
the
my
experience
of
the
CLM
CLM:
it
is
the
kind
of
very
complex
model.
There
are
many
pills
to
be
initialized,
so
I'm,
not
sure
if
we
can
directly
take
the
GRE
initialization
and
in
see
like
the
land
again,
this
needs
to
be
talked
with.
Will
this
needs
to
be
talked
with
the
day?
But
what
I
can
see
what
we
can
think
about?
We
can
take
the
GR
f-14.
We
can
run
a
continuous
CL
m5
grant
and
save
the
CL
m5
history.
M
E
E
It's
you're
essentially
forcing
the
offline
CLM
with
the
Gerry
55
until
the
system
is
spun
up,
and
then
you
start
from
that
spun
up
point
and
you
give
it
the
actual
years
and
you
do
it
again
essentially,
but
starting
from
that
spun
up
point.
So
sometimes
you
run
it
for
you
know
eight
hundred
years
offline
or
seven
hundred
years
or
something
to
get
it
spun
up.
And
then
then
you
run
it
and
put
your
output
out
through
the
time
period
like
let's
say
it's
1992
current
or
something
who's.
B
E
We
this
is
man,
see
if
we've
done
that
as
well.
For
some
of
the
the
kept
prediction
runs
that
we
were
running
with
Brian
Medeiros
and
with
Julie
and
Joe.
So
that's
something
that
I
can
help
with
as
well.
One
thing
we
did
mention
this
they're
working
with
the
new
fates
model,
so
you
just
might
want
to
ask
if
that's
gonna
become
part
of
their
standard
set
up
and
how
that
might
affect
us
going
forward.
F
E
Yeah,
just
it's
based
on
the
questions
it
came
up
later
unless
we
feel
like
it's
already
been
answered
by
your
l46
versus
l32
comparison.
But
if,
since
calm
is
gonna,
be
moving
toward
a
higher
top,
maybe
it
would
be
good
to
see
what
the
benefits
of
that
are
by
having
the
new
grid
whenever
we
decide
it
and
then
the
old
vertical
resolution
and
see
if
there
are
any
clear
benefits,
yeah.
F
I
think
it
would
be
great
to
do
that
on
a
little
longer
time
scales.
Do
it
on
like
a
seasonal
to
inter-annual
timescale,
yeah
I,
don't
know
after
then,
I
run
controls
for
it
as
well,
but
it
could
be
one
of
the
projects
we
proposed
because
last
time
I
talked
to
Julie,
Carolina,
I,
think
she's.
On
the
line
too,
we
were
not
planning
a
running
CSM
for
anima
me,
so
we
could
consider
using
this
a
higher
top
model
for
a
similar
purpose
set
of
friends.
F
You
came
office
has
a
whole
line
of
research
that
there
for
longer
seasonal
and
slightly
longer
time
scales
attribute
a
lot
to
the
stratosphere
in
the
commune,
and
we
have
not
had
a
model
with
which
we
could
do
these
experiments
with.
So
it
is
mid
top
cam
with
what
about
91
levels
would
be
an
ideal
model
to
do.
Those
experiment
was,
for
example,
I,
guess.
E
The
trouble
is
to
get
their
predictability.
We
need
a
large
number
of
members
and
yeah.
We
need
a
lot
of
computing
time
to
be
able
to
do
that
with
the
high
top
yes
right
and
just
to
clarify
it's
just
that
the
funding
didn't
seem
to
come
through
on
doing
the
official
enema
me
in
collaboration
with
University
of
Miami
for
CSM,
but
I'm
not
opposed
to
running
some
something
similar
to
it
for
some
overlapping
period.
E
E
G
Good,
yes,
so
I
had
suggested
a
couple
of
experience.
One
was
doing
the
fully
coupled
CES
and
two
sub
seasonal
forecasts,
which
I
think
Yaga
you're
already
working
on
that
and
then
another
set
of
experiments
is
more
started
from
a
conversation
at
the
international
s2
s
group,
which
is
to
do
a
ocean
of
observing
system
sensitivity,
experiment.
So
OS
sees
where
he
seemed
W.
G
What's
the
impact
of
different
ocean
observations
on
sub
seasonal
focus,
with
understanding
that
the
ocean
is
one
of
the
main
sources
of
predictability
on
those
timescales
on
timescales
of
weeks
to
months
so
ECMWF
and
the
japanese
group
jamei
have
done
experiments,
subsystem
forecasts
experiments
initializing
the
ocean
from
these
different
ocean
reanalysis
products.
So
as
wondering
if
we
could
do
something
similar
with
CSM.
So
the
initial
conditions
are
ready.
G
F
G
Can
do
a
subset,
also
yeah
and
we
I
think
we
should
have
enough
number
of
cases
to
do
at
least
have
a
reliable
forecast
skill,
but
we
can
definitely
do
a
subset
compared
to
what
ECMWF
and
J
I
may
have
done.
They
have
done
a
25
years
period,
but
we
can
do
a
shorter
time
even
a
5
year
period,
but
you
are
looking
at
specific
cases
of
like
MJ,
one
answer
which
is
of
my
interest,
but
we
could
do
other
case
studies
at
least
10
different
cases,
I
guess
to
get
some
reliable
forecast.
Skill
like.
F
Each
sub
seasonal
hindcast
said
I
think
from
my
calculates
about
3
million
core
hours
so
doing
additional
four
of
those
that
would
be
additional
12
million,
so
it
just
depends
how
much
allocation
we
would
have
and
I
don't
know
if
gokon
is
still
on.
If
he
can
comment
on
as
a
working
group,
what
would
be
a
total
reasonable
number
of
hours
to
put
in
in
a
proposal?
I.
H
I'm
actually
going
to
go
over
those
numbers
over
the
next
few
weeks
and
get
back
to
each
working
group,
but
it
is
probably
not
unusual
to
expect
order
15
million
ish
per
year,
but
I
need
to
double
check
those
numbers,
and
especially
in
this
second
year
last
six
months
is
going
to
be
on
the
new
machine
that
is
essentially
going
to
be
tripling
the
amount
of
allocation
that
each
working
group
can
get
during
that
segment.
So
I
need
to
get
back
with
the
numbers.
I
mean
expect
order
15
million
per
year,
if
not
more.
B
F
Think
that's
really
important.
There
is
also
some
promising
changes
in
physics,
for
example,
addition
of
mesoscale
convective
transport
changes
how
the
MgO
propagates,
and
that
would
be
a
great
one
to
test
and
again,
if
we
just
had
a
framework
in
which
we
could
run
a
year
or
a
subset
of
these
hindcast,
it
would
be
much
easier
and
far
more
fruitful
than
re
running
now.
I
guess
that.
H
This
is
your
thing
and
I
wanted
to
make
I
think
Steve
already
suggested,
but
I
wanted
to
reassess
it.
This
essentially
sensitivity
to
sort
of
model
parameters,
ation
changes
or
model
physics
changes
and
how
they
impact
Paquito,
well
predictions
or
system
predictions
they've
been
on
the
radar
for
the
last
I
mean
since
the
inception
of
this
whole
thing,
I
remember
attending
meetings
in
2010
or
even
earlier.
This
comes
up,
but
I.
D
Just
to
add
to
the
concerns
about
prices
to
full
speed,
applied
task
force,
trying
to
select
some
case
studies,
there's
also
the
issue
any
would
raise,
and
also
we
oughta
mentioned,
and
her
presence
basis
of
that
ensemble
size
seems
to
be
really
important,
so
to
be
experiment
and
test
larger
ensemble
sizes
and
think
about
the
ways
larger
ensemble
suite.
Maybe
four,
maybe
four
case
studies.
D
K
Yeah,
sir
thanks
for
bringing
it
up,
it
just
occurs
to
me
when
you
see
it
in
Steve's
presentation
and
others
that
this
really
is
a
key
dimension
leading
to
more
skill.
And
obviously
it's
not
you
know,
especially
given
the
limitations
on
core
hours.
It's
not
really
possible
to
you
know,
use
large
ensembles
for
the
entire
hindcast,
but
I
do
wonder
if
a
case
study
approach
in
looking
at
ensemble
size,
as
well
as
some
of
the
other
methodological
choices
and
perhaps
like
parameter
stations,
could
be
useful.
You
know.
K
F
D
And
I
also
see
from
James
Dunne
mentioned
six
hourly
output
for
work
down
ceiling,
so
it's
probably
difficult
to
do.
Six
hourly
output
for
a
whole
might
have
suite,
but,
along
with
the
case
studies,
there
could
be
maybe
certain
cases
where
more
preset
output
is
provided
such
that
some
work.
Downscaling
studies
should
be
done
as
well.
B
F
Since
rich
mention
is
a
regional
refinement,
kind
of
brings
up
the
topic
of
higher
horizontal
resolution
and
go
in
particular
has
asked
us
to
asked
a
group
word
whether
we
feel
going
to
high
resolution
and
CSM
to
so
quarter
degree
in
the
atmosphere
and
ocean
of
about
0.1
degree.
How
important
is
that
to
this
research
into
this
group?
M
Yeah
so
I
wanted
to
make
a
comment
regarding
like
yeah
what
these
peeps
hid
and
what
you
are
also
saying
that
yeah
there
is
already
a
constant
in
terms
of
the
compute
time
or
in
terms
of
the
resources,
and,
as
you
pointed
in
your
presentation,
it's
not
only
the
compute
time.
It's
also
human
resources
like
who
is
going
to
monitor
these
done
and
something
fell.
So
so
it
did
something
like
a
mechanism
can
be
built.
M
Like
you
as
a
group,
we
run
the
default
run
or
default
set
of
experiment,
and
that
can
be
supplemented
by
the
individual
researcher
with
the
in
versity
grant
like
it
and
then
invert
CT
research
said
we
can
also
get
the
compute
time,
for
example,
if
I
am
interested
in
kind
of
running
the.
What
is
the
contribution
of
such
subsurface
or
mastered
memory?
So
I
can
use
your
default
experiment,
but
I
can
only
run
the
subsurface
part
and
I
can
decrease
the
compute
time
only
for
the
subsurface
part.
M
So
if
we
we
both
were
kind
called
a
very
terribly
or
in
terms
of
per
like,
like,
for
example,
I
have
run,
it
seemed
that
experiment,
but
in
that
case,
I
needed
to
run
both
the
control
and
tumble,
as
well
as
the
perturbance
number.
Here.
What
I
am
saying
as
a
part
of
the
group,
let's
run
the
control
ellipsometry
specialized
and
come
through
the
university
elevator
I.
F
Think
that's
a
great
suggestion,
so
yeah
we
are
trying
to
store
all
the
initial
conditions
and
as
much
as
possible
across
timescales
and
again
at
least
now,
we've
all
moved
on
to
the
same
version
of
CSM.
There
are
some
software
engineering
issues
depending
on
weather,
which
timescale
you
run
but
I
think
that's
a
good
idea.
We
provide
the
baseline
pieces
and
if
you
have
students
to
run
specialized
cases,
I
think
that
would
definitely
help
to
distribute
the
workload.
F
M
K
Yeah
I
had
a
comment
on
the
resolution
question.
It
strikes
me
that
you
know
ECMWF
and
going
from
there
system
for
a
system.
5
you
know,
went
from
I
think
you
know
down
to
like
a
25,
kilometer
resolution
and
I
think
there
your
the
general
sense
was
to
some
extent
those
skill
improvements
were
a
bit
underwhelming.
Give
them
that
effort
and
improving
some
things
like
that
and
I'm.
D
I
think
that
there
was
Cola
and
he
thinks
of
us
together
it
did
a
bunch
of
it
was
Metis
and
Athena
and
I
think
and
car
may
have
been
involved
in
some
of
that
there
was.
These
high
resolution
runs
at
different
resolutions
and
a
lot
of
exploring
what
the
impact
was
and
I
again
I'm,
not
as
up
on
the
literature
as
everybody
who
was
involved
in
the
projects.
But
if
I
recall
that
still
improvements
are
they
even
in
representation
and
variability
by
resolution?
Was
it
as
impressive,
as
maybe
we
would
have
hoped
so.
D
F
B
B
They've
got
a
60
kilometer
out
and
spirit
quarter
degree,
ocean,
high
top
atmosphere,
and
what
that
paper
shows
is
that
the
skill
improvement
is
really
very
marginal
over
our
1
degree
system,
and
probably
you
know,
the
primary
phenomena
that
they
beat
us
at
would
be
the
nao
skill
in
your
one
and
two.
You
know.
So,
if
that's
your
interest,
but
maybe
maybe
all
that
resolution
is
worth
it,
but
you
know
there
just
isn't
a
lot
of
other
improvement
team
globally.
F
A
Sure
it's
partly
a
kind
of
biased
viewpoint
from
the
ensembl
data
simulation
vantage
point,
whereas
doubling
the
horizontal
resolution
costs
us
four
times
as
much
in
resources.
So
we
can
tend
to
do
fewer
experiments
that
way,
and
the
vertical
seems
to
from
from
what
I've
heard
over
the
last
couple
years
bring
in
more
important
physics,
improvements
like
the
qbo
and
so
forth
that
we'd
really
like
to
have
in
the
models
and
I'm,
not
as
aware
of
what
the
horizontal
increase
in
resolution
would
bring
us
does.
A
B
E
E
Responsive
impacts
to
that
coupling
then
high
resolution
in
terms
of
ography
and
in
terms
of
you
know,
extreme
events,
you
really
need
you
know,
high
resolution
to
give
you
a
better
chance
of
capturing
those
statistics
you
know,
is
that
fast-moving
statistics
for
us
as
the
teleconnections
of
your
line?
Okay,.
B
Is
Jeff
I'm
gonna
jump
in
and
say
that,
of
course,
the
gradients
between
resolved
grid
points
in
the
vertical
are
much
more
significant
than
in
the
horizontal.
And
so
when
you're
trying
to
fit
observations,
say
at
the
top
of
a
boundary
layer.
Increased
horizontal
vertical
resolutions
going
to
be
a
much
more
significant
impact
on
improving
analyses
and
pointing
out
differences.
H
This
is
your
cutting
and
I.
Guess
I'll
a
commented
on
this
thing
just
a
few
minutes
ago
we
will
be
doing
some
limited
prediction.
Simulations
as
part
of
our
eye
has
efforts
and
that's
using
the
older
version
of
the
high
resolution
model.
So
perhaps
we
can
wait
for
those
simulations
to
come
online
and
then
assess
the
impact
at
that
point,
so
that
might
be
a
path
to
just
go
that
way.
Yeah.
B
G
Just
wanted
to
make
a
quick
comment
about
the
resolution
issue
and
I.
Think
some
of
the
resolution
issue
could
be
a
fearless
solving
certain
phenomena
or
not
like,
for
instance,
in
the
ocean.
If
you
go
from
one
big
knee
to
a
quarter
degree,
you
still
are
not
resolving
Eddie's
and
most
of
the
ocean
you're
like
at
max
Eddie
permitting
in
the
tropics.
So
if
you
really
do
resolve
certain
phenomena
that
are
sources
of
predictability,
then
I
think
it
does
make
a
difference
for
prediction.
B
F
D
I'd
like
to
circle
back
to
the
state
studies,
I
thought
that
was
a
really
good
idea
and
it
c2
I'm
sort
of
coming
in
different
ways
and
I
was
wondering
if
people
would
be
willing
to
put
in
Google
Doc
specific
cases
that
they
would
really
like
to
see
and,
as
we
start
the
process
of
what
trying
to
come,
trying
to
find
a
way
to
come
together
on
a
set
of
set
of
cases.
We
could
work
with.
K
I'm
sure
Kathy,
listen
Andy
occurs
to
me
several
years
ago
when
I
was
helping
lead
the
drought
task
force
for
NOAA.
We
call
that
about.
Maybe
it
was
four
or
five
different
drought
case
studies,
I'm,
not
suggesting,
then
all
the
case
they
should.
You
know,
focus
on
drought
necessarily,
but
you
can
certainly
I
can
least
list
those
in
your
document
and
maybe
one
or
two
could
be
part
of
the
cases.
E
D
F
F
Discussed
a
lot
about
this
to
us
hindcast,
but
if
you
would
like
to
work
with
the
data,
then
if
you
could
just
send
me
an
email
or
put
your
name
in
the
Google
Doc,
that
would
be
great
and
the
second
question
I
think
it's
also.
A
big
question
is
what
other
Diagnostics
metric
people
would
like
to
see
and
if
they
would
be
willing
to
work,
for
example,
with
the
Clint
Fred
package
and
contribute
metrics
I.
D
F
Nobody
so
far,
volunteering,
so
just
speak
up
if
you
have
an
opinion
so
right
now
there
is,
you
know
many
other
working
groups
like
a
name
WG.
There
is
the
nwg
diagnostic
package.
Climate
variability
has
its
diagnostic
package.
If
you
ever
try
to
analyze
any
of
these
simulations,
you'll
realize
there
isn't
an
easy
way
and
everybody's
been
raining
their
own
code,
but
we
have
an
opportunity
to
leverage
what
Riley's
been
doing.
I
mean
Riley
has
a
raised
hand.
So
maybe
he
can
just
talk
about
it.
A
little
more
yeah.
J
Think
I
think
it's
a
little
intimidating
if
you
have
been
contributed
open-source
to
contribute
those
kind
of
things,
but
what
we've
done
on
the
documentation
is.
We
have
a
whole
outline
on
on
exactly
the
steps
to
to
opening
a
pull
request
in
contributing
in
Aaron
and
I
are
passionate
about
this.
So
we
definitely
guide
people
through
their
contributions.
D
J
Kathy
helps
a
ton
kicking
off
yes
to
us
stuff
that
we
were
dragging
our
feet
on
getting
to
because
it's
a
lot
of
daytime
stuff.
So
we
you
know
as
maintainer
zuv
this
thing
we
are
more
eager
to,
but
things
that
the
community
wants.
Then
we
were
to
go
down
our
own
development
roadmap.
Essentially
any
push
from
the
community
for
your
given
thing
is
going
to
go
to
the
top
of
the
priority
list
and
adding
metrics
is
really
simple.
Just
you
know,
it's
really
copying
a
code
template.
That's
in
the
document.
J
J
It's
at
some
point,
whether
it's
a
student
or
a
community
effort
to
take
Clem
cred
and
create
a
spin
off
Diagnostics
form
of
it
that
you
can
run
on
your
route,
but
I
think
maybe
that's
what
people
are
getting
at
is,
instead
of
every
time
having
open
it
up
and
use
it
interactively
it'd
be
nice
to
have
some
well-written
script,
that
just
ingests
atmospheric
predictably
output
and
then
leverages
Clint
read
internally
but
the
end
product.
Is
you
get
a
document
or
a
set
of
figures
or
something
hosted
online?
G
So
it's
wondering
in
terms
of
the
question
of
what
other
metrics
should
we
include
it.
Would
it
be
useful
to
talk
to
operational
centers
about
this
as
well,
because
I
know
some
of
the
operational
centers?
They
have
this
thing
called
as
a
scorecard
or
a
scoreboard
where
they
have
like
a
standard
set
of
metrics,
and
they
use
that
to
evaluate
any
improvements
in
the
forecasting
system.
Does
it
improve
or
degrade
forecast
skill
in
any
of
those
standard
metrics,
so
that
could
be
something
useful
for
our
sensitivity.
G
D
Think
that's
a
good
idea
and
I'd
also
like
to
think
about
it
kind
of
more
broadly
in
that,
maybe
as
a
working
group
with
this
and
more
specifically
scientific,
rather
than
forecast
skill
approach
or
focus
that
maybe
we
think
about
some
of
our
own
metrics
that
are
relevant
to
the
particular
scientific
forces
of
predictability
that
we're.
After
because
my
my
personal
opinion
is
that
I
get
really
bored
of
sitting
at
skill
scores
over
and
over
from
this
system
and
that
system
and
that
system.
And
it's.
D
Feel
like
come
Fred
gives
me
the
ability
to
crank
through
those
Kilgore's
see
what
the
skill
is
like
and
then
I
can
jump
in
the
scientific,
looking
at
predictability,
component
kind
of
more
quickly
and
so
having
having
metrics
that
are
or
Diagnostics
that
are
around
us.
The
predictability
sources
of
predictability
components
are
something
that
I
think
it
would
be
great
to
see
in
this
group
beyond
fuel
towards
themselves.
F
K
Yeah
and
in
fact,
I
alive,
probably
is
a
better
person
to
talk
about
this
than
I
am,
but
since
she
has
a
new
project,
but
it
does
strike
me
that
you
know
we
tend
to
think
of
a
need
for
land
evolution,
kind
of
dynamics
being
lower
in
shorter
range
predictions
and
I.
Wonder,
given
that
there's
kind
of
a
you
know,
limitations
on
the
computing
resource.
K
You
know
it
might
be
interesting
to
know
how
far
you
have
to
push
complexity
and
the
land
surface,
to
really
get
feedbacks
that
maybe
matter
on
in
a
climate
prediction,
sense
and
I'm
sure
they
vary
as
you
go
out
to
Catalina
three
four
weeks
ahead.
All
right,
it
just
seems
like
maybe
there's
some
kind
of
experiment,
particularly
in
some
of
the
drought
cases
that
can
be
done
to
look
at
how
look
influence.
You
know
that
land
complexity
has
on
skill.
E
Yeah
definitely
I
think
yeah,
it's
in
the
sense
of
the
kind
of
perturbed
physics
ideas
that
so
our
project
was
I
think
mostly
in
like
a
perfect
model
world,
but
it
would
be
interesting
to
do
it
for
observed
drives,
but
to
look
at
how
various
new
complexities
in
CLM
5
affect
the
evolution
of
drive.
So
things
like
the
plan
hydraulics
the
time
we'll
talk
about
Monday
or
just
simply
soil,
moisture,
preconditioning,
irrigation
and
fires
and
dynamic
vegetation,
and
things
like
that.
E
So
it
would
be
nice
to
do
that
for
some
observed
events
to
see
whether
well,
maybe
we
at
first
we
could
look
at
it
in
the
perfect
model
framework
like
we
had
proposed.
But
if
you
can
see
the
adding
in
particular
features
in
the
land
model,
get
you
closer
to
observations
to
the
extent
that
we
have
observed
these
things,
then
that
would
be
good
as
a
kind
of
a
check
on
how
relevant
is
for
reality.
F
D
I
think
when
some
of
these
kinds
of
things
were
discussed
previously
in
some
other
meetings
that
there
is
a
strong
fisheries
application
and
component
to
this,
so
I
think
looking
into
some
of
those
types
of
applications
would
be
and
I
think
a
Water
Resources
component.
Maybe
Andy
can
speak
to
that
I
I.
Think
of
there
being
some
really
exciting
applications
related
research
that
could
be
done
with
these
multi
I
think.
D
B
So
it
looks
like
Nicki
Levin
dusky
is
no
longer
on
the
call,
but
you
know
I
would
have
thought
based
on
her
work,
looking
at
terrestrial
predictability
and
deep
ele
potential
predictability
that
extends
out
to
two
or
three
years
that
this
this
is,
that
might
be
the
real
way
to
entrain
land
modelers
into
the
ESP
working
group
will
be
doing
it
with
CLM
five
and
it's
on
a
sort
of
timescale
that
is
very
relevant
to
land
processes,
and
so
you
know,
would
be
good
to
have
anyone
from
that
community.
Who's
on
chime
in
I.
C
F
Thank
you
all
for
coming
and
hopefully
we'll
see
you
at
the
next
working
group
meeting.
Oh
yes
and
I
think
somebody
has
posted.
There
is
also
discussion
lunches
and,
in
particular
the
polar
working
group
overlapped
with
us,
so
they
were
going
to
try
to
discuss
if
they
wanted
to
contribute
or
look
at
the
output
in
some
of
these
simulations
over
their
lunch
time,
discussion
which
will
be
in
about
15
minutes.
D
K
C
C
E
C
I
know
that
it
was,
it
was
kind
of
like
Tetris
when
gokon
and
I
were
trying
to
figure
out
which,
which
ones
to
put
at
the
same
time.
But
if
people
do
come
back
to
its
recorded
and
by
this
evening,
or
maybe
even
later
this
afternoon,
people
can
go
to
the
CTD
YouTube
channel
and
watch
it.
So
it's
uploading
pretty
soon
so
great
and
then
Ryan
and
I
are
gonna
work
on
loading
the
presentations.