►
From YouTube: SimPEG Meeting August 6
Description
- Discussion on a trajectory for generating examples and answering user-questions on SimPEG
- John presents on using SimPEG to invert DCIP data / Mag Gradiometry data and progress on implementing new receiver types in the natural source electromagnetic code
A
Perfect,
so
we're
no
lie
cool.
Well,
thanks
for
coming
in
after
your
long
weekend,
everyone
so
there's
a
few.
We
can
go
around
and
do
sort
of
quick
updates
click
reports.
If
anyone
has
been
notes
open,
please
feel
free
to
add
yourself
to
the
list
of
folks
who
are
attending
one
of
the
things
actually,
and
this
would
be
I
think
a
good
habit
for
us
to
try
and
start
doing
is
actually
like
everyone
just
list
your
name.
A
If
you
are
attending
versus
not
one
of
the
things
that
it's
been
kind
of
interesting,
just
keeping
in
touch
with
the
pen
geo
project
they've
had
to
work
with
the
NSF
just
to
show
like
that
they
are
making
progress
and
stuff
like
that,
and
simply
just
having
meeting
notes
with
like
who
showed
up
was
actually
a
huge
step
there.
So
it's
kind
of
a
low
bar,
but
a
nice
thing
for
us
to
keep
track
of
so.
A
I
have
a
couple
quick
updates,
but
maybe
we
can
go
around
first
and
if
there's
anyone
else
who
wants
to
start
us
off,
go
for
it.
So
maybe
I'll
go
by
the
order
on
on
my
screen.
John
do
you
have
any
quick
updates
you'd
like
to
share
before
what
will
go
into
empty
longer,
but
feel
free
to
give
a
quick
summary
if
you
like?
Oh
you,
you
were
muted,
oh.
D
C
Volume
averaging
operation
and
I
implemented
something
pretty
quickly
in
Python
to
test
it
out.
So
I
think
I've
got
a
nice
way
to
work
that
implemented
and
that's
extensible
for
1d
2d
and
3d,
at
least
for
the
tensor
meshes
for
now,
but
I
think
I
have
a
good
understanding
of
what's
happening
there.
So
should
be
a
quick,
a
dish
addition
if
we
want
that.
C
E
C
It
might
have
been
I
was
trying
to
look
at
how
it
was
implemented
there.
I
can
I
got
maybe
of
what
was
happening
really
quickly
and
implemented.
My
own
way
of
doing
it
without
really
looking
what
was
happening
there,
but
essentially
I
was
just
looking
for
overlapping
cells
in
each
dimension.
So
along
one
dimension,
I,
look
for
the
overlapping
cells
and
then
inter
and
then
go
from
one
to
the
other.
Okay.
E
F
F
When
that
came
up
is
that
I
guess?
What's
on
Geo
side
of
XYZ
car
computations
we're
we
haven't
updated
the
notebooks,
so
I
I
did
the
I
did
the
update
for
the
mag,
but
we
need
I
created
an
issue
on
the
on
the
computation
repo,
but
we
we
need
to
review
all
those
all
those
examples
because
they're
they're
breaking
you
know:
they're
not
matching
master.
G
A
A
F
A
F
F
F
A
Well,
I
think
you
actually
go
to
computation
done
to
your
side
on
X,
Y
Z.
Another
thing
it's
on
then
yeah.
Sorry,
it's
not
easy
to
find.
But
if
you
actually
go
to
computation
to
your
side
orexi,
it
does
exist
somehow
googled
the
right
thing
they
could
have
landed
here.
We
haven't
really
advertised
this
page
much.
G
F
A
A
Well,
cuz
cuz
Devon
did
a
nice
job,
organizing
both
the
examples
and
the
tutorials
indiscreet
eyes,
and
so
John
and
Devon.
If
he
does
a
game
to
work
together,
that
might
be
yeah
just
a
productive
way
of
moving
forward,
and
maybe
you
can
sort
of
divide
and
conquer
some
of
the
examples
so
that
you
know
not.
One
person
has
to
carry
all
of
them.
A
F
Yeah,
once
I
can
bring
him
back
so
see.
We
have
a
bunch
of
notebooks
that
exist
under
computation
that
Joe
statics
way,
they're
not
exposed,
but
some
people,
someone
has
managed
to
get
to
that
side
and
all
those
books
are
breaking
cuz
they're
enough
to
master
so
we're
thinkin,
Mets
camera
these
guys
back
in
bringing
it
back.
A
A
A
H
A
Yeah,
this
actually
I
might
jump
in
with
my
update.
If
that's
okay
and
then
we
can
search
on
this-
is
totally
relevant
as
well.
Cuz
I've
noticed,
especially
Dom
you've
been
fielding
a
ton
of
questions
which
is
awesome
and
extremely
helpful
and
I'm
wondering
so
I've
seen
like
some
of
them.
It
looks
like
we
actually
are
in
place
where
you
can
just
were
in
a
place
where
you
can
send
them
a
link,
and
it's
like
here's
what
you
need
to
know,
or
here's
an
example
that
you
can
use
to
answer
your
question.
A
But
that's
not
like
always
the
case
and
soggy
I've
noticed
that
too,
even
with
like
Craig
there's,
there's
not
always
examples
for
like
what
what
they're
looking
for
and
so
I'm
wondering.
If
there's
a
trajectory
where,
because
Devin
I
know
you're
invested
in
working
on
some
examples
for
the
docks,
so
I'm
wondering
if
there's
a
way
where.
A
Basically,
if
there's
a
question
that
comes
in
that
we
don't
have
an
answer
for,
if
folks
can
can
work
together
and
get
connected,
so
that
devon
has
the
tools
that
he
needs
to
actually
just
like
create
an
example
that
you
could
then
send
to
that
person
in
the
future.
So,
even
if
you
need
to
immediately
answer
the
question
like
that's
fine,
but
hopefully
the
outcome
of
like
any
new
question,
basically
should
be
that
if
that
question
we're
asked
again
in
two
weeks
time,
we
just
have
a
URL.
That
has
an
example.
H
You
were
gonna,
try
and
sorry
trying
to
answer
all
the
basic
questions.
How
do
I
create
Meister
they
and
how
do
I
go
and
do
this
sort
of
basics
for
a
simulation,
or
something
like
that,
so
they
find
a
lot
of
the
time
when
I'm
answering
questions.
If
you
had
a
very
basic
tutorial
simple
example
with
enough
text
that
it
would
answer
90%
of
the
questions
we
asked.
G
H
A
Yeah
totally
agree
and
I
think
that
so
all
that
I'm
suggesting
is
that,
rather
than
it's
just
a
question
of
which
way
to
approach
this
problem
is
there's
one
way
where
you
can
identify
like
here's,
the
entire
suite
of
tutorials,
that
we
need
to
solve
the
problems
of
any
user.
So
that's
sort
of
step,
one!
That's
like
a
build
up,
build
it
from
all
of
the
building
blocks.
A
So
as
it's,
it's
just
a
question
of
prioritizing
but
I
think
because
we're
getting
so
many
questions
coming
in
they're,
focusing
in
on,
like
let's
just
make
sure
that
we
have
examples
to
answer
these
exact
things
that
that
might
help
us
more
so,
especially
in
the
short
term,
while
we're
doing
some
refactoring
and
such
rather
than
sort
of
sketching
everything
out
and
knowing
that
you
know
some
things
are
changing
some
things.
Some
things
are,
will
continue
working
working
as
they
are.
Does
that?
Does
that
make
sense.
H
A
A
If
folks,
you
like
helped
prompt
the
creation
of
a
tutorial
are
willing
to
jump
in
and
help
out
with
a
review
that
can
streamline
things
and
if
you
give
a
thumbs
up
but
need
help
merging
and
stuff
like
that,
like
I'm
happy
to
do
that
and
I'm
happy
to
sort
of
answer
targeted
questions,
but
if
we
can
spread
out
like
we're
where
and
when
reviews
are
happening,
I
think
that'll
help
smooth
over.
You
know
any
individuals
busy
schedule
so.
A
E
A
E
A
In
some
ways
this
is
a
question
for
Devin,
so,
like
I
mean,
if
soggy
sort
of
had
the
start
of
an
example,
how
do
you
two
want
to
work
together
because
I
mean
soggy?
You
can
start
a
rough
floater
question.
You
could
work
from
there
or
an
issue
or
what?
What
do
you
think
is
the
best
best
road
forward.
H
E
I
think
Mayon
is
the
easiest
way
like
I
could
make
an
issue
and
provide
a
notebook
that
I
provided
to,
let's
say
crack
and
then
put
a
description.
What
what
the
question
was
and
how
I
answered
that,
like
that's,
probably
the
easiest
way,
man,
if
Debbie,
can
turn
that
into
either
example
or
whatever
documentation
that
that
sounds
like
a
good
workflow.
My.
H
G
A
That's
that's
totally
something
we
can.
We
can
create
that
tag.
One
quick
thing
with
creating
examples.
I
think
this
is
an
implementation
detail,
but
it
will
be
easier
if,
for
like
each
issue,
there
is
a
new
pull
request.
So
if
there's
a,
for
example,
there's
a
single
pull
request
that
goes
to
that,
so
not
trying
to
map
a
branch
with
a
whole
bunch
of
examples.
So
too
many
issues
if
we
can
keep
it
sort
of
one
to
one.
A
G
F
Of
like
he's
building
his
own
operator
so
I
think
yeah,
that's
that's
one
for
Joe,
actually,
yeah
I
think
you
would
like
to
be
able
to
do
backward
difference
or
even
Center
their
friends.
It's
probably
not
hard
on
your
side
too,
to
add
this
right
for
treatment.
Should
you
be
able
to
do
backward
differences?
No.
C
No,
no,
it's
just
make
sure
I
mean
I
was
looking
through
his
code
before
I
know,
there's
other
instances
of
creating
some
like
his
ex
creating
dense
matrices
of
the
meshes
at
some
point.
So
we
gotta
watch
that
out
watch
out
for
that,
but
yeah
I
mean
I
think
there's
a
little
bit
simpler
way
to
implement
the
gradient
operators
that
I
have
that
have
done
in
my
own
stuff
so
or
at
least
the
to
use
arbitrary
gradient
operators
inside
that
cost
gradient
operation,
yeah.
C
F
A
E
E
H
E
F
E
Happy
to
help
and
I
do
have
some
example,
and
it
seems
like
all
your
LP
inversion.
It's
like
a
it's
all
kind
of
everything's
supposed
to
combine
together
well
to
to
have
a
good
performance.
It
seems
my
feeling,
like
I'm,
basically
playing
with
your
codes
and
getting
some
good
reserve
in
Ian's
problem,
but
yeah
I
wasn't
quite
sure
anyway.
So
it's
kind
of
curious.
Do
you
have
a
tutorial
or
some
some
example.
F
G
F
G
F
A
Let's
maybe
start
if
you're
game
to
start
an
issue
discuss
this
because
I
think
we
can
think
through
this
I
want
to
make
sure
that,
like
pip,
install
SCIM
peg
still
continues
to
work
easily
for
everyone.
So
we
might
need
to
like
think
through
what
the
installation
process
looks
like
or
if
the
Fortran
piece
is
actually
sort
of
a
separate
repo
that,
like
you,
can
install
and
if
it's
installed,
then
we
use
it
and
if
not,
we
use
the
Python
code,
but
I.
Think
yeah.
A
F
E
F
E
E
A
A
B
Okay,
yeah
so
yeah
just
applying
it
to
our
data
sets
in
the
industry.
Here.
First
off
our
DC
and
IP
we've
been
using
pretty
extensively
mostly
tensor
mesh,
because
that's
the
one
I've
convinced
everybody
on
so
far
and
then
clients
have
already
come
back
to
us
saying
how
much
they've
enjoyed
the
inversion
of
the
models
and
very
good
stuff,
and
hopefully
s
IP
soon.
I
gotta
get
Charlotte
and
touch
a
bit.
B
Soggy
they're
here
right
away,
cool
tensor,
pretty
much
a
lot
of
stuff
that
Dom's
been
doing
just
piggybacked
off
of
him
and
got
some
great
results
there
and
then
magneto
clerks
yeah,
just
joining
a
whole
bunch
of
different
different
styles.
So
first
yeah
so
early
on
I
managed
to
get
the
the
tensor
mesh
going
pretty
quick
for
everybody
and
here's
an
example
of
some
of
the
larger
inversions
it
took
us.
Seven
million
sell
tensor
mesh
like
seven
hundred
thousand
data
points,
eight
hundred
injections
yeah.
B
It
was
pretty
big,
takes
about
a
week
to
run,
but
you
know
it's
pretty
doable,
not
too
bad
yeah.
So
I
guess
you
can
see.
You
know
I'm,
just
gonna
move
that
there.
So
in
the
corner
here
is
the
chargeability.
Then
I've
got
our
chargeability
and
resistive
structures
they're
mostly.
What
we
do
is
just
just
a
little
bit
of
surface
waiting,
maybe
play
around
with
the
errors
a
little
bit.
B
We
have
a
little
bit
of
a
calculations
that
we'll
do
ourselves
just
see
how
quality
the
signal
was
and
yeah
we've
been
really
getting
really
good
results.
So
the
next
thing
I
was
doing
was
getting
octree,
hopefully
going
because
to
run
one
of
these
inversions,
we
need
to
put
it
on
one
of
our
big
computers.
That
has
you
know.
B
Eighty
cores
run
it
for
quite
a
while
so
coming
here
you
can
see
we
like
to
include
the
potential
and
current
remotes,
so
you
can
tell
if
that
air
you
can
see
that
there's
a
tensor
mesh,
it's
gonna,
it's
really
gonna.
Add
up
quick.
However,
we're
so
running
to
off
tree
right
away.
The
IPE
model
actually
really
well
not
too
much
of
a
problem
there.
It
was
when
we
got
to
the
DC
the
tensor
mesh
kind
of
trims
everything
around
here.
B
You
can
see
my
my
cursor,
that's
usually
how
how
far
it'll
pull
deep,
whereas
the
awk
tree
it
seems
that
everywhere,
at
the
levels
it
just
kind
of
heard
of
kind
of
goes
through
all
of
them
and
just
like
pulls
everything
down
deeper.
You
know
we'll
be
cutting
off
the
survey
right
there.
It's
just
a
lot
of
my
colleagues
and
stuff
we're
kind
of
concerned
why
it's
doing
so
much
at
depth.
B
However,
in
the
area
of
concern.
These
three
guys
here
for
this
example
here
correlates
to
these
little
three
guy
Nats
of
of
interest
to
the
geologists
that
we
working
with.
So
all
in
all
the
Autry
is
working
pretty
well
getting
results
that
we
want
just
a
little
concern
on
what's
going
on
at
depth,
why
it's
building
stuff
and
also
at
the
interfaces
you
can
kind
of
see
the
model
kind
of
there's
a
little
bit
of
a
distortion
here
and
there,
however,
I
talked
to
Tom
about
this
and
then
so.
B
B
B
F
B
F
G
E
A
Well,
is
that
something
I
think
we
talked
about
just
like
the
potential
confusion
of
of
that
step
and
potentially
wanting
to
actually
have
cell
weights
in
there?
Oh
sorry,
the
cell
volume
is
always
in
there
and
then,
if
users
want
to
include
some
weights,
have
that
be
separate?
Is
that
still
something
that's
on
your
radar
loose?
Think
that
that's
important
or
is
that
that
conversation
for
later.
F
The
if
here
the
main
issue
is
not
the
cell
volumes
is
the
way
the
gradients
are
calculated
for
the
active
tree
for
the
tree
mentions
right.
So
the
main
difference
between
the
two
here
is
that,
in
the
simple
case,
we're
doing
gradients
for
each
of
the
hanging
phases
to
our
particular.
It's
a
single
gradient
over
four
phases
and
that
that
changes
that
changes
the
solution,
that's
the
main
difference.
Yeah.
E
E
G
B
Cool
that
was
the
other
thing,
one
of
the
other
things.
This
is
a
like:
a
combined
data
set
with
pull
dipole-dipole
and
pull
pull,
and
it
seemed
that
this
was
kind
of
a
rising
more
in
the
pull
pull
data.
But
I
guess
that's
because
it's
trying
to
do
something
more
depth,
whereas
before
you
know
there
was
nothing
there
sensitive
to
it.
So
we
didn't
see
it
yep.
A
No
Doug
right
now
we
don't
at
least
it's
not
in
the
iteration
printers.
This
is
something
that
we
actually
should
allow.
Some
flexibility,
I
think
soggy
froze,
which
is
why
I'm
jumping
in
so
right.
Now
this
the
printers
are,
we
do
like
Phi,
D
and
Phi
M.
If
you
like,
you
can
write
a
directive
that
would
amount
because
we
do
have
access
to
all
of
that.
Actually
T
Bo,
you
might
have
done.
You
might
have
done
some
of
this
already.
Have
you
done.
You
changed
that
like
what's
printed
to
this,
we.
D
Is
whether
it's
big
or
small
actually
does
define
the
volume
that
cure
that
you're
working
with
so
alpha
s
is
10
to
minus
3
might
actually
be
a
pretty
small
number
if
you've
got
something.
That's
only
couple
hundred
meters
across,
but
if
you've
got
something
as
a
couple
of
kilometers
and
to
have
the
same
kind
of
relative
weighting
for
that
those
components.
D
Rotation
now
this
might
have
weakened
the
minor
in
a
spot
and
an
easy
way
of
just
kind
of
giving
you
first-class
assessment
of
this.
It
just
say
like
okay
in
the
regularization
function,
just
print
out
each
component,
and
now
we
find
that
that
smallest
model
component
is
ten
times
as
large,
bring
in
terms-
and
you
know
like
it's
just
a
quick
diagnostic
yeah.
H
A
What
would
be
nice,
if
someone's
game,
to
give
this
a
go?
It
actually
should
be
reasonably
well.
Scoped
is
actually
in
the
optimization
code
to
perhaps
like
just
change
the
verbosity
and
so
be
able
to
say,
like
you
know,
right
now
is
the
default
is
1
or
whatever
and
be
able
to.
You
know,
set
that
to
2
and
how
that
actually
outputs
all
the
components
of
the
regularization.
A
A
B
B
Yeah
well,
I
guess
next
is
kind
of
the
stuff
that
mostly
I've
been
trying
to
do
here.
At
first
is
yeah
using
Zara
to
storge
I
mean
using
desk
I
can
create
Jay.
The
only
thing
is
I'm
finding
that
since
okay,
so
how
I
make
it
and
ask,
is
I
make
a
list
of
all
the
sensitivity
matrices
from
each
injection
for
each
source,
receiver,
I
guess
each
source
will
have
its
own
matrix
and
it's
since
it
changes
every
for
every
injection.
That's
where's
ours.
B
Having
problems
that
I'm
finding
here,
let
me
go
up
to
where's
my
get
j,
so
I
bring
it
into
here
and
I
have
experimented
so
if
I
make
a
desk
array
or
a
desk
object.
That
just
was
a
bunch
of
vectors
of
all
the
same
length.
X
amount
of
vectors
I
can
write
bizarre,
no
problem,
but
as
now
that
it's,
since
it's
they're
all
different
matrices,
I
still
form
I
guess
I
do
have
V
stack
of
the
desk
operator
and
it
all
my
shape.
B
B
F
B
G
F
B
B
F
B
G
B
Cool,
so
yeah
I
should
be
able
to
have
a
examples
here.
You
recondition
air
with
a
pretty
large
PC
circle.
I'll.
Try
that
out
the
very
least
and
I
guess
after
that,
I'd
like
to
try
the
tiled
inversion,
because
that
was
really
successful
in
the
full
tensor
Meg
stuff.
It's
it's
been
amazing.
I
can
invert,
so
many
so
many
data
said
so
I'm
like
a
laptop.
B
F
B
For
using
surface
weights,
applying
it
to
the
Audrey
problem
for
DC
yeah
for
DC
yeah
yeah
I'm,
trying
to
how
can
I
word
this,
it
almost
seems
that
if
I
don't
have
a
really
good
DM
that
covers
the
whole
mesh,
I'm
gonna
have
really
big
blocks
on
this
side
and
those
blocks.
If
they're
associated
with
a
surface
where.
C
B
Start
the
surface
waiting
I,
don't
know
that.
Just
to
me
that
seems
like
where
things
are
going
wrong.
So
is
it?
Is
it
maybe
the
way
that
I'm
formulating
how
to
I
guess
how
to
use
the
cell
weights
or
is
it
may
be?
The
difference
in
cells?
Do
I
have
to
take
into
account
the
size
of
the
cells
themselves
when
they're
doing
apply
did.
F
B
G
B
B
Karen's
and
yeah,
so
we
made
it
so
that
you
can
just
pull
in
like
a
CSV
file
of
all
your
electrodes
and
then
you
can
pull
in
a
topography.
So
when
I
guess
it's
meant
more
for
planning
a
survey,
and
we
can
use
this
to
you
just
throw
in
your
electrodes
tell
it
what
kind
of
spacing
you
want,
how
many
injections
you
want
and
yeah
it
just
uses
the
simulation
class
to
produce
a
forward
model
yeah
on
anything.
You
want
any
terrain.
Any
electrode
arrangement.
A
That's
awesome,
yeah
I
mean
if
you
wanted
to
start
start
a
pull
request
on.
We
should
figure
out
like
what
the
name
of
that
like
style
of
module,
is
because
I
think
that
we
talked
a
bit
about
this
in
Montreal,
but
the
idea
of
like
sort
of
rappers
or
drivers
or
simply
Express,
or
something
like
that.
A
That's
basically
sort
of
like
a
high-level
API
that
sits
on
top
of
synthetic
that
if
you
have
set
up
your
inputs
in
sort
of
a
specified
way
like
what
you're
saying
with
the
CSV
file,
then
we
can
just
take
care
of
so
yeah
I
mean
if
you
want
to
start
a
pull
request
and
any
like
name.
Ideas
for
what
this
module
is
called
would
be
would
be
great,
and
that
goes
for
any
one
is.
B
Yeah,
so
for
DC,
that's
practically
what
I
got
there
next
is
the
full
tensor
can't
quite
talk
too
much
about
this,
because
it's
a
very
hard
data
set,
but
here's
the
pictures
of
it
yeah
big-time
mesh
I,
ran
this
on
my
Mac.
It's
about
4800
stations
after
downsampling,
but
yeah.
It
worked.
Oh
pretty
well,
yeah
because
of
the
tile
dr.
Ian
Lazar,
that's
all
Dom
right
there,
no
superfast,
that's
great
yeah,
and
then
the
Cartesian
and
the
spherical
I've
been
really
enjoying
that
too
the
difference
or
being
able
to
use
both
of
them.
B
As
you
can
see
here,
it
seems
that
the
spherical
again
seems
to
reduce
the
the
spread
yeah
to
go
into
more
example.
Here.
This
I
guess
it's
kind
of
up
here.
It's
this
little
area.
Here
was
the
area
of
relative
interest,
and
this
is
what
they
were
producing
and
most
of
their
results
were
the
one
on
the
right
here
and
they
were
kind
of
questionable.
B
Well,
this
right
here
like
how
how
real
it
is,
or
you
know
how
is
it
something
that
should
be
investigated
and
as
we
go
into
like
the
spherical,
then
using
TMI
with
the
gradient
components,
it's
actually
showing
that
you
know
it
might
not
be
the
most
important
part
that
they're
looking
for.
So
they
were
really
really
happy
to
see
a
discrimination
between
those
two
loans
right.
There.
G
B
Yeah
the
way
that
yeah
I
haven't
really
developed
a
way
to
assign
the
weights
so
much.
Oh
I,
guess
I'll,
just
the
residuals,
yeah,
just
kind
of
by
Rhian
vert
inversion
every
time,
just
kind
of
seeing
what
the
residuals
are
afterwards
I'll,
either
bump
up
or
bump
down
the
components.
I
guess
that's
probably
another
thing:
I
can
mr.
rack
Dom's
marine
or
anybody
else
is
like
this
using
that's
kind
of
like.
How
can
this
be
automated
instance?
B
We
can
afterwards
do
an
assessment
and
kind
of
narrow
in
because
it
took
me
about
four
or
five
like
those
at
it
before
I
found
a
nice
right
right,
threshold,
I
guess
maybe
just
say,
maybe
there's
a
way
to
be
a
little
bit
more
effective.
Maybe
this
was
kind
of
similar
to
what
Sean
Walker
was
talking
about
a
while
ago.
B
D
G
B
Magnitude
of
the
diagonals
cool,
all
right
yeah,
so
that
one
also
I
have
started
bringing
this
over
or
I've
taken
the
3d
example,
loans
MPEG
there
and
have
started
bringing
that
over
to
the
simulation
class.
Just
to
try,
however,
on
when
I
run
it
my
first
iteration
they
crashes
after
the
first
iteration,
because
both
my
Phi
D
and
my
F
come
out
as
any
ends,
but
if
I
run
it,
if
I
run
without
simulation,
if
I
use
the
problem
class,
everything
seems
to
work.
Okay,.
G
B
A
B
A
B
A
So
that's
a
first
thing
to
do:
okay,
there's
some
example.
So
a
lot
of
the
tests
you
can
have
a
look
at
basically
enter
the
frequency
domain
or
the
MT
tests.
The
frequency
domain
might
be
slightly
easier
to
sort
through
military
shirt
yeah
because
that's
that's.
The
first
thing
to
check
is
just
make
sure
that
we
are
walking
in
the
right
in
the
right
direction,
so
the
sensitivity
should
converge
second-order
and
if
not,
then
that
means
that
there
is
something
we
need
to
go
in
and
hunt
down.
A
A
G
H
B
B
A
B
Yeah
yeah
and
then
yeah
I
just
ran
a
chipper
one
over
the
same
situation,
just
kind
of
tipper
in
the
err
just
move
the
block
around
and
things
seem
to
work,
pretty
good,
perfect
case,
but
it
seems
to
work
pretty.
Well,
that's
good.
The
the
cone
activities
are
a
little
bit
higher,
but
not
not
too
much
more.
It's
it's
pretty!
Well
thanks!
B
So
then,
there
yeah,
so
what
I've
found
the
parallelization
I've
used
desk
and
I
tried
multi
processing
at
the
same
time,
doing
it
over
frequencies
and
multi
processing
is
about
you
know.
Twenty
seconds
faster
task
is
about
yeah
I'm
a
few
seconds
faster
than
that.
It's
where
the
plug
is
is
doing
the
a
inversion,
because
every
frequency
you
need
to
run
the
solver,
and
that
seems
to
slow
everything
right
down.
B
A
There's
a
couple
things
going
on
here,
so
the
serialization
and
the
D
serialization
shouldn't
affect
a
n
that
should
affect
the
when
you
call
fields
yeah,
that's
where
the
serialization
deserialization
of
the
fields
objects,
the
the
AF
term.
So
the
question
I
have
here
is:
are
you
using
like?
Are
you?
Is
this
multi
threaded
or
multi
core?
Actually.
A
And
not
still
because
so
a
multi
threaded
I
could
see
that
being
problematic
is
cuz
like
the
solver
is
basically
going
to
try
and
go
and
like
grab
all
the
threads
and
then,
if
you're,
trying
to
run
multiple
frequencies
at
once.
No
matter
how
your
parallelizing,
the
solver,
is
parallelized
differently,
and
so
we
sort
of
don't
have
control
at
that
level.
G
A
Maybe
the
thing
to
do
is
because
I
think
this
is
gonna,
take
a
little
bit
for
for
all
of
us
to
sort
out
but
I.
Think
there's
a
few
of
us
who
are
perhaps
working
with
this
right
now
so
I'm
doing
work
running
simulations
on
nurse
using
dots,
so
I've
been
successful
in
just
like
basically
sending
jobs
to
different
cores
and
getting
simples
back
that
that's
sort
of
this
simpler
version.
G
A
E
E
G
E
G
E
E
E
E
And
also
John,
there's
a
like,
rather
than
using
Purdue,
so
there's
a
iterative,
solver
I
think
it's
a
by
CG
Jacobi
or
something
like
that
in
in
pinus
solver.
Try
using
that
cuz
like
you'll
see
you
know:
okay
yeah
in
that
case
they're
supposed
to
some
some
improvements.
I
guess,
even
if
you
don't
store,
am
sorry.
B
E
That'll,
like
that'll,
cut
like
the
memory
shoe
for
a
yam
problem,
particular
auntie
I
think
that's
it.
That's
actually
not
a
well.
If
you
have
really
large
cell
like
a
in,
will
be
expensive,
but
mostly
it's
filled
so
yeah.
If
you
can
store
that
and
that
could
be,
it
could
also
save
memory
and
potentially
speed
up
as
well.
Okay,.
G
A
E
G
A
A
A
D
B
Would
be
great,
I
was
trying
to
work
on
trying
to
get
some
other
colleagues
to
you
know
work
on
some
of
the
analytical
solutions
so
that
we
get
at
least
compared
to
something
there,
but
yeah
we're
essentially
trying.
It
would
be
a
little
bit
easier
to
to
sell
to
the
bosses
and
stuff
if
we
could
yeah
have
a
comparison
to
something.