►
From YouTube: SimPEG meeting April 22
Description
No description was provided for this meeting.
If this is YOUR meeting, an easy way to fix this is to add a description to your video, wherever mtngs.io found it (probably YouTube).
A
B
B
Trying
to
find
ways
to
caramelize
it
or
the
calculation
fields
I'm,
actually
finding
that
when
I'm
doing
it
in
parallel,
it's
not
that
much
faster
that
if
you
solve
it
in
sequential
yeah,
it's
quick.
It
takes
like
a
couple
like
five
seconds
to
solve,
whereas
if
I'm
just
parallelizing
with
Dax,
sorry
to
ask
it
I'm
only
getting
like
a
couple
like
maybe
10%
speed,
oh
wow,.
D
B
B
B
A
B
C
B
C
A
D
Yeah
I
can
so
the
decor
request
I
created
last
week
for
for
the
surface,
to
end
topo
on
discretize,
so
I
thought
I
created
a
new
test,
I
kind
of
moved.
The
test
I
was
in
synthetic
and
never
meant
to
do
so.
The
poor
request
is
really
for
review.
It
passes
all
everything
just
need
someone
to
to
approve
it
and
I
expanded
it
for,
for
all
my
shows,
or
should
work
for
tensor
archery
cylindrical
and
from
2d
to
1d,
from
3d
to
Wendy,
so
I
think
we're
we're
pretty
covered
there.
D
So
I
would
think
if,
if
this
goes
into
disco
toys
we
should
just,
we
should
try
to
just
change
all
that
all
the
code
in
incent
bag
that
uses
the
surface
gentoku,
currently
insane
flag
and
just
reroute
to
a
discretized
version
and
through
our
warning
you
know,
because
I
renamed
it
to
something
more
something
a
little
better.
I
think
so
yeah
because
that
bridge
once
you
once
you
guys
bring
it
up.
C
E
Instead,
I
have
enabled
it
to
be
like
a
solver
option
that
you
can
use,
so
it's
not
tied
to
the
simulation
class,
it's
tied
to
it
as
a
solver
up
solver
option.
It
took
a
little
bit
of
thinking
to
get
that
working
for
the
party,
so
the
solver,
because
the
way
that
asset
mko
number
of
threads
works
is
it
sets
it
for
all
of
em
Kayla
processes.
E
So
every
time
you
call
it
it
sets
it
for
every
MDL
process
that
gets
called.
So
it's
more
like
a
global
property
than
something
that's
local
to
each.
That
would
be
local
to
each
solver.
Object
itself
had
to
think
a
little
bit.
I
didn't
meant
it
like
a
global
way
in
terms
of
the
Python
solvers.
So
now
it's
so
if
you
see
a
little
bit
of
weirdness
in
there,
that's
that's
why?
If
you
start
looking
through
it,
if
you
want
look
to
the
pull
request,
I
created
one,
the
other
part
of
that
is
I.
E
It's
hard
like
we're.
Not
getting
tests
passing
properly
on
Travis
and
things
do
to
monk's,
usually
failing,
and
what
I'd
like
to
do
is
remove
that
part
of
the
solver
interface
and
pie
that
solver
for
now
and
we're
gonna
be
working
I'd
like
to
work
towards
getting
psyphon
interfaces
to
the
mumps
package
through
Conda
Forge,
because
it
is
available
through
there.
E
So
we
can
hopefully
optimize
with
the
solver
solver
a
little
bit
better
and
use
that
and
take
advantage
of
some
of
those
up
these
little
parameters
that
we
can
tweak
cuz
right
now,
there's
not
really
much!
You
can
optimize
with
the
party
so
solver,
you
know
not
many
knobs.
You
can
turn
right
now
with
it
so
I'm
working
towards
that.
Yes,
I.
C
E
E
E
He
asked
for
all
the
fields
in
the
way
that
you
did
and
if
it
was
a
simply,
it
was
a
delayed
function.
You'd
get
to
give
you
a
delayed
function
back,
but
you
wouldn't
really
know,
and
then,
if
you
pass
that
to
like
Jay,
Beck
and
J
transpose
that
it
would
recompute
the
fields
every
single
time,
because
it
would
compute
it
would
recognize
like
the
J
bike
and
JT
leg
of
the
night
with
the
fields
a
delayed
function,
it
would
call
compute
on
them,
so
called
compute
on
them.
C
E
E
E
E
E
A
You,
oh
yeah,
I,
know
and
I
totally
agree.
Removing
mumps
seems
like
a
reasonable
step
at
this
point.
We
can
just
do
a
minor
release
of
climax,
Oliver
and
say
it's
gone.
We're
gonna
work
on
getting
it
back
in
a
better
fashion,
but
if
you
were
relying
on
it
use
a
older
version
of
high
mass
over
yeah.
A
F
F
Well,
one
thing
is
I'm
fighting
with
the
50
gradients,
as
you
know,
but
another
thing
is
what
might
be
interesting
and,
as
many
of
you
I
think,
Dom
shown
as
well
have
talked
I
get
the
heavy
role
in
graphical
user
interfaces,
development
and
sort
of
the
last
two
weeks.
I
went
the
other
direction.
I
wrote
a
lot
of
terminal-based
user
interfaces
so
within
the
consortium
there's
this
joint
conversion
framework,
where
they
use
gravity
day
in
the
magnetic
data
tomography
in
versions,
and
they
all
belong
to
different
companies.
F
So
they
have
their
central
servers
and
the
joint
inversion
framework
that
calls
out
different
methods
to
run
those
and
then
combine
them
again.
So
all
that
has
to
be
able
to
run
with
command
in
the
terminal
and
configuration
files.
So
I
learned:
how
can
my
code
be
called?
It's
just
a
command
and
configuration
files
which
it's
interesting
because
it
needs
quite
some
different
thinking
and
then,
when
you
work
in
a
notebook
or
or
or
in
an
I
Python
shell
and
have
everything
available,
it's
kind
of
a
reduction
workflow.
F
A
Don't
have
a
command
line
interface
at
the
moment,
Dom
has
written
sort
of
like
driver
files
and
lightweight
wrappers
that,
basically
you
just
do
like
Python
run
my
run.
My
file
yeah,
but
I
think
that
this
would
fit
really
nicely
when
we
start
to
think
about
like
a
higher
level.
Api
for
simple
is
to
have
basically
like
mechanisms
to
make
some
opinionated
choices
around
mesh
design
default
parameters,
those
sorts
of
things
and.
A
And
so
with
that
sort
of
layer
of
infrastructure,
once
we
get
that
in
place,
then
I
think
it
starts
to
make
more
sense
to
like
figure
out
what
should
our
command
line
in
your
face
be
yeah,
because,
right
now
it's
we're
still
a
couple
steps
away
from
sort
of
like
being
able
to
really
just
say,
like
running
and
working
with
with
this
file
and
with
the
properties
in
the
serialization
work
that
we've
been
doing
it's.
It
should
definitely
be
getting
closer,
but
yeah,
not
quite
there.
A
F
D
G
A
G
F
F
F
There
you
have
the
basic
information:
it's
an
european
funded
project
in
a
wider
and
Maritim
research
umbrella
that
that
that
spawns
of
many
projects
that
have
anything
to
do
with
meritum
technologies
and
in
our
project.
It's
there
are
two
terminal:
openness,
its
tech
and
terraces.
They
are
both
more
or
less
in
the
greater
area
of
humble
in
German
and
then
involved
is
Gilmar
in
kill
a
German
Research
Institute.
There
is
a
Chiefs
at
Potsdam,
also
in
Germany,
and
then
it's
me
from
tu
Delft
and
the
other
four
partners
with
kind
of
changing
partners.
F
They
had
already
two
or
three
previous
European
research
projects
that
are
mainly
focused
on
hydrates.
They
also
had
more
chemistry
focus
and
now
it's
more
a
geophysical
focus,
and
in
this
one
we
have
sort
of
three
areas:
three
zones
as
the
the
black
sea,
which
are
hydrates,
there's
a
bar
on
food
or
grass
and
then
there's
a
closed
vitally
there's
kind
of
deep-sea
mining.
Sulfites
and
the
whole
idea
is
just
this
joint
integration
of
multiphysics.
If
you
want
so
they
have.
This
terraces.
Has
this
framework
where
to
run.
E
F
G
F
I
want
to
have
a
gradient
for
that,
and
then
this
server
like
say
that
Kumar
is
setup,
that
it
always
listens
to
incoming
new
resistivity
models
and
as
soon
as
there
is
a
new
model,
it
runs
one
iteration
well
one
forward
model:
it
calculates
the
gradient
and
writes
it
back
to
file
and
lets
the
central
server
know
that
there's
a
new
gradient,
so
we
can
continue,
and
that
happens
with
with
each
of
the
methods
now
with
my
code.
That'll
be
definitely
because
my
code
is
open
source.
F
F
H
I'm
not
too
much
so
I've,
more
or
less
implemented
the
recursive
solution.
So
thanks
to
Joe
for
giving
me
some
assistance
with
that
been
doing
my
I
guess
the
adjoint
test
and
looking
at
the
sensitivities
and
they're
they're,
very
accurate
for
the
real
component
and,
if
I'm,
with
the
sensitivities
of
apparent
resistivity
data,
but
there's
some
something
interesting
happening
with
the
the
imaginary
component.
H
So
gonna
need
to
investigate
that
so
I'm
wondering
if
there's
an
operation,
that's
maybe
truncating
it
or
we
haven't
defined
some
array
to
keep
enough
of
the
of
the
digits
but
yeah
something
something's
going
on
there.
But
it's
almost
there.
That's
that's
pretty
much.
What
I'm
trying
to
finish
up
in.
A
A
Just
gonna
say
it
might
be
constructive
to
go
through
the
frequency
domain,
J,
Beck
and
JT
Beck.
The
handling
of
real
and
imaginary
components
is
a
touch
subtle,
especially
in
the
graduate,
so
it
might
be
worth
just
kind
of
reading
through
that
code
and
and
seeing
if
there's
any
insights
that
you
can
gain
in
in
parallel
with
what
you've
done
all.
H
Right,
yeah,
well,
I,
guess
yeah.
Basically,
what
I've
done
is
we
that
we've
formed
the
the
sensitivity,
but
it
hasn't
been
broken
into
real
and
imaginary,
and
so
then
I'm
just
taking
the
real
component
of
that.
The
sensitivities
is
my
my
sensitivity
to
the
real
data
and
the
imaginary
component
of
that
for
the
matter:
the
components
of
the
imaginary
sensitivities,
so
I'm
separating
them
that
way
right
because
I
don't
want
to
have
the
sensitivity
matrix
with
imaginary
numbers
floating
around
otherwise
you're
yeah
updating
the
model
in
the
complex
plane.
H
So
I've
done
that,
but
you're
saying
that
maybe
it's
it's
has
something
to
do
with
how
j
j--
10
j
transpose
avec
is
doing
the
computation
cuz.
I
noticed
that
it's
also
just
j
back
as
well.
I'm
individually
looking
at
the
Opera
yeah
I'm,
looking
at
the
operation
of
J,
vac
and
I'm.
Comparing
this
also
against
the
what
we
call
it:
the
the
finite
difference,
sensitivities,
okay,
so
I'm,
noticing
that
there's
a
discrepancy
only
in
the
imaginary
component,
so
I'm
I'm
a
little
curious.
Why
this
is.
A
H
Be
good
to
know,
I'm
I'm
sure
it's
one
of
those,
it's
a
just,
a
tiny
thing
that
needs
to
be
found
and
then
it'll
work
perfectly
so
details,
but
I
guess
after
that
then
I
was
just.
It
was
gonna,
go
and
turn
what
I
have
into
some
tutorials,
because
it
follows
the
norm,
the
the
framework
that
we're
going
to
be
using
and
and
then
have
those
ready
for
any
time.
You're
gonna
bring
the
stuff
into
the
base.
Emulation
fresh
sounds.
A
Good
one
thing
we
can
think
about
and
I'd
appreciate,
Joe's
take
on
this
too,
when
that
pull
request
is
ready,
is
actually
if
the
simulation
engine
should
sit
in
Geo,
Anna
or
if
it
should
sit
in
simple
and
I
I'm
inclined
to
put
it
into
G
Oh
Anna
because,
like
with
insipid
we've,
really
sort
of
focused
on
sort
of
the
finite
blame
approaches,
but
that's
something
we
can.
We
can
think
about.
I
just
want
to
pose
it
as
a
question
so
that
when
we,
when
we
do
like
think
about
merging,
we
just
think
that's
you.
E
Yeah
in
the
in
the
e/m
have
a
static,
resistivity,
1d
solution.
It
basically
does
a
recursive
solution
like
thanks,
Auggie
implemented
crystal
solution
it,
so
that's
similar
to
kind
of
thinking.
So
does
the
recursive
solution
ever
uses?
Those
Hinkle
transforms?
It's
transformed
it
afterwards
so
yeah.
A
H
H
D
C
H
C
H
Sorry
good
I
was
just
gonna
say
what
sounds
like
the
the
class
for
the
em1
DFM
problem
and
I.
Don't
know
if
you've
done
e-m1
DTM
as
well.
Those
are
living
on
the
base,
simpang
branch,
so
they're
they're,
using
the
old
syntax,
and
so
really
it
just
just
the
process
of
bringing
them
into
simulation.
No.
D
D
H
It
doesn't
seem
like
it's
super
high
level,
so
yeah
I
can
I
can
get
on
that
and
then
once
we
have
that
stuff
in
there,
then
we
can
go
and
figure
out
how
we
solve
the
universe
problem
or
how
we
create
a
regularization
class
for
that
yeah.
H
Well,
I
like
having
like
a
simulation
one
D
class
within,
like
it's,
the
frequency
domain
method,
to
keep
it
separate.
It's
working
so
much
differently.
Then
then,
like
you
know,
solving
Maxwell's
equations
for
in
3d
doesn't
seem
like
you're
you're
able
to
use
much
here
in
class
and
that
yeah
yeah,
so
you
might
as
well
just
keep
it
separate
and
it's
a
much
I,
don't
know
it
seems
to
keep
things
organized
that
way.
So.
D
I
D
D
I
A
H
A
Ok
because
because
the
other
option
and
I
don't
in
some
senses,
this
might
be
simpler
and
Joe
feel
free
to
chime
in
here
to
you,
but
is
to
actually
just
take
the
the
currents
in
Peggy
m1d
repo
and
create
a
simulation
branch
on
that
and
then
do
the
refactoring
there.
And
then
we
can
figure
out
what
the
best
structure
is
within
within
simple
for
a
lot
of
these
1d
codes.
H
A
E
H
E
I
What
interesting
question
actually
for
that
empty
inversion
is
so
you
have
a
two
different
types
of
data
that
you
can
think
of:
impedance
real
imaginary
or
apparent
even
face
yeah.
So,
like
kind
of
like
having
a
bit
of
motivation,
it's
like
a
stinging
one.
These,
it's
not
very
like
exciting,
but
if
you
have
a
couple
of
questions
like
that,
Oh
which
data
type
is
better
or
what
what
what
we
need
to
use.
I
D
D
H
And
I
didn't
find
it,
and,
and
so
I
thought
there
was
definitely
space
for
us
to
to
add
it,
and
I
also
feel
that
it
might
work
even
better
than
the
em1
DFM
codes.
Just
because
I
feel
the
data
are
much
more
sensitive
to
what
is
directly
below
compared
to
the
em1
DFM
I.
Don't
think,
there's
as
much
horizontal
talking
with
the
empty
problem
as
there
is
in
the
the
e/m
problem.
There.
E
I
No
yeah
that
no
I
mean,
if
you
think
about
the
footprint,
it's
like
it's
much,
much
bigger
and
then
I
think
I
think
I'm
a
little
bit
pessimistic
because
you
probably
think
the
other
way
cuz.
If
because
one
DMT
is
such
as
like
a
a
cheap
solution
and
if
never,
if
other
people
haven't
done
that,
then
oh
that's
probably.
F
F
G
I
think
in
the
scheme
of
things
it's
just
like
that
is
something
used
to
be
done
in,
like
the
nineteen
seventies
type
of
version,
nineteen
1980s
and
then
then
the
default
went
to
the
2d
because
the
2d
algorithms
kind
of
came
out
there.
Actually,
you
know
moderately
robust,
so
you
could
get
both
te
and
TM
inversions
and
that's
basically
where
that
default
lays
I.
Think
I
mean
there
definitely
are.
As
dieter
says.
You
know,
stitched
1d
algorithms
around
but
yeah
they
haven't,
they
haven't
really
been
utilized
since
the
TV
became
so
prevalent.
I
Yeah
I
don't
know
that
because
it
I
think
you'll
have
even
like
a
hard
time,
even
fitting
the
data.
If
you
just
simply
use
one
decode,
so
you
need
a
lot
of
other
processing
like
other
people,
are
seem
to
actually
use
one
decode
to
invert
actual
MT
data,
so
yeah
I
think
it's.
The
context
is
spit
different
from
the
airborne
here.
I
Yes,
yeah
I'm,
not
quite
sure.
G
Then
inverting
him
in
you
know
one
D
should
give
you
a
pretty
reasonable
answer,
but
if
they
start
to
diverge
them,
not
really
sure
what
you
get
up
actually,
at
that
point,
doing
1d
inversion
for
tensor
invariance
of
the
MP
impedance
seems
to
have
been
a
bit
more
robust
entity
to
to
use
and
and
and
frankly,
if
we're
going
to
use
everybody
stitch
together.
One
DMT,
then
maybe
one
should
use
the
determinant
of
the
impedance
tensor
as
data
I.
E
H
Yeah
I
was
I,
wasn't
proposing
that
we
would
be
inverting
the
Z
x,
y&z
Y
X,
because
yeah
I,
don't
think
I
would
I.
Think
I
would
have
lots
of
trouble
fitting
the
data.
In
that
case,
I
would
almost
propose
that
you'd
you'd
do
something
maybe
like
inverting,
the
the
XY
and
then
inverting
the
the
Y
X
separately,
or
maybe
seeing
if
you
got
different
models
from
from
those
two
components
of
the
impedance
tensor,
but
I
would
think
you'd
be
able
to
fit
the
data.
If
you
did
it
that
way,.
F
Definitely
asked
carry
key,
or
at
least
read
the
Mauri
to
DM
manual
because
he
has
in
Mari.
He
has
all
the
possible
components
you
can
use
for
inversion
through
his
merry
work.
I
think
he
has
quite
a
lot
of
knowledge
about
how
to
join,
convert
e'en
and
empty
and
I.
Remember
that
I
read
some
notes.
What
works
fine
and
what
doesn't
work?
That's
fine,
it
might
be
worth
writing
in
an
email,
is
quite
open
to
Hawaii
what
he
knows
to
share
what
he
knows
about
those
things.
F
H
That
there's
other
parts
of
it
too,
or
at
least
talking
to
some
of
the
people
who
are
trying
to
get
into
using
the
MT
stuff,
is
their
desire
to
try
and
invert
all
of
the
frequencies
at
once,
which
I've
really
said
that
that's
crazy
and
you
can't
be
doing
this.
But
you
know,
if
you
do
this
ditched
1d
inversion,
and
then
you
start
to
get
a
sense.
The
depth
at
which
your
target
is
the
frequencies
at
which
which
are
sensitive
to
that
target.
H
H
I
We're
interested
in
what
they
called
Corcoran
clay.
So
it's
like
a
very
extensive
unit
in
Central
Valley
of
California.
So
it's
a
it's
a
fairly
thin
clay
layer.
So
if
you
just
do
a
smooth
inversion,
you
you
have
like
hundred
meters
thick,
not
chip
layer,
but
that's
not
true,
so
yeah
I'm
just
like
starting
to
use
what
that
prior
information
that
I
can
use.
So
I
was
thinking
about
using
sparse
inversion,
but
the
problem
in
p.m.
I
your
your
mall
is
actually
not
sparse.
So
some
okay,
cuz
gel,
is
complicated,
so
you
can
simply
use
what
them
used
to
do
like
just
putting
a
zero
as
like
a
reference.
So
you
need
to
come
up
with
like
a
pretty
good
background
model
and
then,
of
course,
sparsity
to
get
that
I
got
them
clay,
bad
or
something
yeah.
That's
what
I
made
me
working
on
and
yeah
also
they're,
like.
I
Sometimes
our
board
survey
has
done
multiple
times
in
a
couple
of
years,
so
you're
gonna
get
that
actually
like
a
couple
of
different
system
covering
the
same
area
with
different
sampling
at
different
again
density
of
the
sampling
people,
don't
actually
invert
them
together.
Usually
so
I
was
playing
with
oh
I.
Can
we
actually
invert
them
together,
get
a
common
resistivity
model,
so
structure
is
already
all
of
the
structures
develops,
I'm,
mostly
playing
with
what
we
could
do
for
a
given.
Data
sets
yeah.
I
Really
so
yeah
I'm,
basically
inverting
them
together,
like
using
this
spatial
constraint,
so
not
like
it's
not
an
exact
joint
conversion.
Cuz
model
has
a
different
space,
but
by
putting
a
constraint
in
game,
you
can
still
force
the
invert
like
their
model
to
be
very
similar.
I
thought
that's
actually
a
good
idea,
but
then
it's
actually
quite
applicable
for
various
airborne
system.
So.
D
D
I
That's
what
typically
done
in
en
problem,
so
what
they
call
l1
norm
is
usually
on
the
gradient
so
but
like
that,
doesn't
really
help
right
like
it
it
well.
It
helps
to
have
put
a
boundary
on
it,
but
but
you're
not
really
forcing
the
sparsity
of
the
model,
which
is
yes,
I'm
I'm.
Looking
forward
to
so
I
think
people
haven't
done
that
for
the
M
problem.
I
G
No,
not
really
I
was
really
pleased
actually
hear
about
this
Ataru
Jim
consortium,
which
I
not
hadn't,
heard
of
before
so
it'd,
be
really
interesting
to
kind
of
see.
You
know
what
scope
of
problems
there
they're
attempting
to
do
and
just
sort
of
how
the
superstructure
of
all
of
that
is,
is
working
so
yeah
thanks
for
that,
but
no
I
don't
have
anything.
We've
been
kind
of
consumed
with
scg,
abstracts
and
stuff
like
that,
which
that's
that's
another
meeting
that
might
not
get
held.
He's
taste
that
see.
A
A
Yes
with
that,
and
then
the
other
thing
to
share
I
guess:
will
Joe
and
I
chatted
earlier
this
week,
just
about
the
simulation
and
how
to
bring
that
in
and
do
that?
Well,
so
we're
working
on
starting
a
blog
post
that
we
can
use
to
just
kind
of
announce,
what's
changed
and
be
like
a
high-level
overview,
and
then
we
can
make
sure
the
release
notes
are
pretty
detailed,
but
with
that
one
of
the
things
I'm
going
to
try
and
do
is
sketch
out.
A
So
if
you
actually
have
something
in
mind
and
just
want
to
like
sketch
it
in
a
notebook
and
take
a
screenshot
or
take
a
picture
and
send
it
along,
it
would
actually
be
pretty
cool
to
get
a
sense
of
how
people
are
how
your
conceptually
thinking
about
simpie
and
we
can
try
and
consolidate
that
into
actually
like
one
single
framework
diagram
so
feel
free
to
sketch
in
and
send
pictures,
yeah
and
so
I
think.
That's
basically
it
from
from
my
end
see
anyone
else
have
questions
things.
You
thought
thoughts,
you'd
like
to
bring
up.
I
C
I
So
we
put
together
an
s
e-g
abstract
for
DC
code.
Well,
like
the
motivation
was
helping.
This
set
GW
project,
but
I
thought
like
it's
probably
worthwhile
to
actually
write
a
paper
about
that
code.
So
I
think
now
it's
actually
quite
matured,
though
we
have
like
some
things
on
paralyzation
stuff,
but
I
think
it's
already
I
should
get
good
shape
and
just
extending
and
actually
that
gwe
motivation
is
good
and
we
also
got
kind
of
numerical
support
that
we
can
handle
1d,
2d
and
3d
and
have
a
relatively
good
example
related
to
hydrologic
application.
I
A
I
A
E
B
B
Yep,
okay,
yeah
I'd
yeah
I
can
tackle
that
if
they'll.
C
E
F
A
A
Not
oh
yeah,
no
I
do
have
them
in
the.
A
So
Devon
started
putting
that
together,
so
we
basically,
we
walked
through
sort
of
what
are
some
of
the
big
changes.
We
would
like
to
see
I'm
so
structurally
things
like
right
now.
Actually
the
natural
source
am
has
its
own
simulation,
which
is
distinct
from
the
frequency
domain,
which
doesn't
really
need
to
be
the
case,
but
there's
some
assumptions
in
the
frequency
domain
code
that
just
made
it
easier
with
the
development
to
sort
of
have
them
as
separate
things.
A
So,
for
example,
you
know
the
source
getting
back
to
you
back
years
when
we're
doing
3d,
because
we
need
both
polarizations
for
any
natural
source
simulation,
so
just
those
little
style
of
assumptions,
at
least
for
initial
development.
It
was
easier
to
do
that
on
its
own,
but
it
would
be
nice
to
actually
basically
throw
away
the
an
SEM
simulation
like
the
way
it
calculates
the
fields
and
really
have
the
natural
source
am
only
be
in
the
survey.
The
sources
and
the
receivers.
E
A
With
that,
we
also
talked
a
bit
about
boundary
conditions
and
so
having
right
now.
The
only
implementation
we
have
is
a
primary
secondary
April.
So
it
would
be
nice
to
get
some
boundary
conditions
implemented
so
that
we
can
so
that'll
drive,
hopefully
actually
some
development
in
discretize.
A
If
we
can
absolutely
get
abstract
it
well
enough
and
then
I'm
just
looking
at
the
news,
we
identified
a
few
different
examples
to
work
from
so
we'll
start
with
the
L
block
model
that
the
jiff
codes
have
been
tested
against
quite
extensively
and
then
there's
a
number
of
us
who
have
a
data
set
in
mind
that
we
want
to
work
with
and
so
we'll
sort
of
bring
those
bring.
Those
in.
A
F
A
A
So
things
like
one
of
the
things
that
I
think
would
be
pretty
cool
is
Joe
Hammond
hooked
up
czar
to
a
Redis
server,
so
what
you
can
actually
do
is
be
like
looking
at
results
or
looking
at
pieces
are
the
computation
that
we
typically
wouldn't
have
access
to.
So
something
like
we
could
watch
how
the
gradients
change
in
each
iteration
and
things
like
that.