►
From YouTube: SimPEG Meeting May 14
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
Oh
feel
free
to
jump
in
there.
If
there
are
things
things
you
want
to
go
over,
but
first
of
all,
first
order
of
business
is
to
welcome
theatres
here,
it's
nice
to
nice
to
happen
here
and
he'll
be
filling
us
in
on
some
of
the
exciting
work.
I've
been
doing,
building
up
a
multi
grid
solver,
so
I'm
really
excited
to
see
this,
but
Before
we
jump
in
there.
Let's,
maybe
just
do
if
there's
quick
updates
from
anyone,
so
you
can
see
folks
is
there.
C
A
C
Young
I've
literally
got
lots
of.
It
creates
that's
from
to
explain
what
geophysics
is
two
kids
and
a
large
public
and
it's
called
go5.
What
and
now
we
made
the
first
video
actually
explained
just
our
geophysics,
with
some
Street
interview
asking
people
whether
they
know
when
to
you,
physics
is
and
stuff
like
that,
and
you
can
check
it
out.
I
can
put
the
link
to
the
blog,
in
slack
very.
A
If
you
were
following
along
in
the
Python
channel
in
slack,
there's
a
conversation
that
was
going
on
there
with
mostly
myself,
Franklin
and
Rowan,
want
to
figure
out
like
what
the
data
class
should
look
like
and
from
that
I've
sketched
out
an
issue
comment
on
like
what
some
of
my
thoughts
are
and
so
I'd
appreciate.
If
anyone
wants
to
jump
in
and
take
a
look
and
chime
in,
that
would
be
great
and
yeah
so
sort
of
feedback
in
the
next
sort
of
day
or
two
and
then
we'll
get
cracking
on
an
implementation.
So.
D
E
Yeah,
nothing
right
now,
I'm,
just
kind
of
working
actually
I'm
working
on
the
Meg
gradient
stuff
they're,
just
trying
to
get
it
running
so
I
should
have
actually
some
runs.
I
got
so
I
have
my
hands
on
some
actual
data,
so
I'm
trying
to
invert
it
right
now
so
about
a
week
or
so.
I
should
actually
have
something
to
show
excellent.
D
D
Very
exciting,
very
inspiring,
very
transformative,
I,
don't
know
there
were
I
think
about
26,
and
what
surprised
me
the
most
is
there
were
a
lot
from
industry
folks
for
shell,
a
queen
or
del
slumber,
say
doc,
oxy
I
think
so.
Maybe
I
forgot
someone
well
yeah.
So
there
was
a
pretty
broad
range
of
companies
present
as
well.
Next
to
all
academia
and
different
sort,
I
tried
to
science
was
there
of
course,
yeah
I.
Think
that's
it,
but
ya
know.
D
The
basic
idea
is
that
you
can,
you
can
go
to
smaller
grids,
so
you
keep
the
dimensions
equally
white,
but
instead
of
saying
having
24
cells,
you
just
go
down
to
two
massive
cells.
You
solve
the
system
there
and
then
you
upscale
it
and
and
try
to
solve
it
there
and
you
jump
between
different
levels
of
the
grid
and
the
idea
is
to
one
on
the
lower
grades.
Basically,
you
capture
the
feature
that
our
lower
frequency
in
some
sort
and
then
in
the
finer
mesh.
D
You
capture
the
parts
that
are
from
like
a
higher
frequency
content
and
the
fact
that
you
can
do
lots
of
stuff
from
lowered
rates
means
you
can
save
a
lot
of
energy
and
if
you
save
a
lot
of
memory
and
also
run
time,
that's
kind
of
the
basic
idea,
I
would
say,
but
it's
an
iterative
solver.
So
you
jump
through
grade
levels
and
then
the
residual
from
from
you
calculate
left
hand,
side
right,
hand,
side
that
the
residual
is
your
new
source
term
in
the
new
grid.
D
D
D
A
D
D
If
there
is
strong
and
isotropic
or
let's
say
strong
variations,
then
the
pure
multigrid
code
can
struggle,
and
so
but
I
probably
show
that
by
the
way
too,
there
are
some
modes
where
you
do.
There
are
kind
of
two
tricks
one.
Is
you
only
get
coarser
in
some
directions,
mm-hmm
and
you're,
not
getting
closer
in
one
for
one
cycle,
and
then
you
change
that.
So
you
can
alleviate
some
of
those
those
problems
and.
D
The
other
thing
is
instead
of
not
coursing
in
between
each
step,
we'll
see
that
as
well.
You
do
smoothing
basically
one
or
two
cows,
Seidel
iterative,
solvers
or
just
two
steps.
So
just
do
some
averaging
website
and
you
do
that
enough
for
the
whole
problem
at
once,
but
you
go
from
one
node
to
another
node.
So
the
only
way
he
always
just
include
all
eight
neighboring
cells
and
one
of
the
ways
to
address
this
problem
with
anisotropies
like
then
you
can
do
this
smoothing
in
one
direction
completely
at
once.
D
D
D
D
Two
to
the
power
of
n
are
best
suited
because
you
can
coerce
them
down
to
all
the
way
to
two,
but
you
can
also
use
in
between,
let's
say
3
times:
2
2
times
2
times,
2
or
4
times
there
5
times
2
times,
2.
Well,
yeah,
it
should
be
a
number.
If
you
take
a
high
prime
number,
then
you
can't
coerce
on
it
because
you,
you
cannot
divide
all
white
cells
by
2
right
mm-hmm.
This
Cobra
will
be
generalized,
but
it
isn't
so
yeah
just
to
keep
in
mind
and
I.
D
Just
a
simple
models
of
10
Hertz
here
I
make.
This
is
a
homogeneous
model,
so
I
just
have
here.
The
resistivity
in
X
is
1.5
tested
in
y1,
potato
st
inset
3.3
and
then
I
defined
the
source
field.
In
this
example,
where
I
did
it
compared
it
to
same
peg,
you
can
always
define
also
just
in
in
simp
X.
So
this
is
nothing
all.
This
is
nothing
new,
let's
say,
and
then
what
is
news
to
you
now,
the
solver?
So
it
just
quickly
run
it.
So
you
can
see
first,
what
it
means
by
multi-rate.
D
So
we
see
here,
the
original
grid
is
48
cells
by
32
by
32
so
and
the
courses
street
is
3
by
2
by
2,
so
we
can
course
on
it
4
times
in
each
dimension,
right
we
can
go
from
48
to
go
to
24,
12,
6,
3
and
similarly
from
32
it
goes
to
8
to
16
8
4
to
that
kind
of
the
the
cordoning
procedures.
So
we
start
off.
D
D
So
at
the
beginning,
here
I
have
a
little
QC
plot.
So,
and
this
is
kind
of
for
the
offensive,
then
you
read
up
on
multigrid.
You
have
different
plots
that
look
like
WS
or
V's
or
similar
like
this,
that
this
is
the
coarsening
procedure,
so
we
go
all
the
way
down
and
here
on
the
bottom,
we
have
a
model
from
this
three
by
four
by
four
cells.
D
We
do
smoothing
here,
which
is
once
or
two
steps
of
the
gauss
seidel
interative
solver,
and
so
it
just
took
two
iterations,
so
it
would,
it
will
not
solve
it
until
it.
It
fix
the
tolerance
it
just.
How
do
you
describe
the
past?
It
just
has
some
iteration,
so
it
that's
why
it's
called
smoothing
right.
It
tries
to
adjust
the
parameters,
but
then,
instead
of
going
all
the
way,
it
goes
one
level
up.
D
So
then,
here
we
will
have
six
by
four
by
four
cells
and
it
was
smoothing
again
and
then
it
goes
to
the
coarsest
level
again
and
so
like
this.
It
just
moves
through
the
cycle
and
tries
to
solve
here
for
the
lowest
frequency
content
and
then
on
higher
level.
It
tries
to
fix
higher
and
higher
frequency
content
of
your
problem
and.
D
That
is
defined.
This
is
the
sometimes
you
really
does
full
multigrid
cycle,
so
we
start
with
the
coarsest
and
then
go
subsequently
up
and
down.
I
think
the
original
cycle,
it's
just
a
V
cycle,
so
you
go
all
the
way
down
and
go
all
the
way
up,
but
multigrid
in
cycle
you
rarely
solve
it.
So
you
see,
in
this
case
I
run
eight
of
those
cycles.
D
D
D
D
G
D
Just
instead
of
running
it
straight
away,
we
do
multi
grade
cycles
as
preconditioning,
and
now
you
see
it
requires
three
big
cheese
stuff
cycle.
If
I
take
multigrid
completely
out,
then
probably
50
directions
which
I
said
to
the
maximum
is
not
enough
to
solve
the
problem.
I
would
have
to
do
many
more
big
cheese's
I,
so
multigrid
can
speed
up
other
solvers.
Quite
significantly,
that's
another
possibility,
let's
say.
D
Yes,
that's
one
of
the
things,
so
it's
it's
implement
for
frequency
domain
and
just
for
electromagnetics
mm-hmm,
the
source
term,
which
sort
of
makes
it
just
feasible
right
now
for
yes,
M,
nothing
I'm,
not
even
sure,
because
the
magnetics
is
not
explicitly
involved,
I
mean
they
are
linked,
so
you
could
somehow
get
the
magnetic
field
out
of
it.
But
right
now
it's
just
implemented
for
electric
source,
electric
receiver,
that's
how,
by
default
it
works.
A
G
D
I'm
gonna
use
the
same
thing
because
at
the
moment
I
mostly
just
take
the
correct
locations,
which
is
the
edges
where
the
field
is
located
at
that
moment,
so
I
or
I
interpolate
it
like
you,
do:
okay,
but
yeah.
It's
just
I
mean
the
whole
discretize
model
is
the
same,
so
the
calculation
happens
at
the
same
point
to
say,
and
also
that
this
how
the
source
is
defined,
I
think
that's
all
the
same
as
in
same
peg,
because
if
I
don't
did
you
have
time
to
have
a
look
at
this
one?
No.
D
F
D
How
you
usually
yeah
this
is
the
model
I
think
you
defined
the
model,
so
it's
just
here,
err
see,
there's
a
little
a
little
target
in
the
middle
and
and
the
rest.
This
is
a
homogeneous
background
and
then
let's
jump
the
multi
read
code
and
go
here,
so
you
defined
the
the
PDE
mm-hmm
the
receiver
locations,
all
the
source
server
you
pair
them,
and
then
you
run
it
make
changes
take
place
to
performance.
It's
actually
not
too
long.
So
let's
run
it
so
in
this
notebook
I.
D
D
A
D
A
Sorta,
how
how
much
is
the
solver
you
need
to
know
about
the
physics,
I
guess,
and
so
well
maybe
what
I'm
getting
at,
and
this
will
be
more
evident,
I
guess
when
I
have
a
chance
to
dive
into
the
code
and
when
others
do
too
but
like
how?
How
in
how
entangled
I
guess
are,
is
the
physics
with
the
salt,
I.
D
Guess
have
two
levels:
one
is
the
course
inning
and
the
prolongation
there.
You
just
have
to
make
sure
that
all
the
properties
you
need
are
getting
coarsened
or
prolonged,
as
they
usually
call
it
also
made
one
level
down,
or
one
interpolated
or
or
or
summed
up
right.
This
is
one
point.
This
is
more
about
the
physical
properties
which
is
kind
of
more
mass,
related,
I,
guess
and
then
the
soul.
We
need
to
know
specifically
in
the
smoothing
steps
it
does
calculate
the
whole.
D
D
Ya,
I,
don't
think
it's
getting
clearer
just
from
quickly.
You
seem
like
when
you
go
down
here.
This
little
V
is
right
on
you
new
one
new
course.
New
to
this
is
this
is
called
the
three
smooth
thing.
Then
the
smooth
thing
on
the
coarsest
level,
which
is
kind
of
a
solver,
and
then
you
have
to
post
smoothing.
So
in
all
these,
you
do
iterative
solutions
to
your
problem
and
all
these
need
to
know,
but
it's
always
the
same
function,
but
this
function
needs
to
know
your
your
problem.
E
D
A
Dug
into
it
a
bit
most
of
it
was
written
by
Rome
like
at
a
conceptual
level
and
then
I,
dug
into
the
implementation
too.
So
I'm
happy
to
chat
through
that.
It
seems
like
this.
Might
it's
not
clear
entirely
like
where
it
would
sit?
It
might
integrate
better
to
be
imported
from
sim
pack
just
because
it
needs
to
know
so
much
about
the
Oracle
beam.
A
But
that
said
like
there
might
be
some
things
that
we
can
do
to
abstract,
like
which
physics,
it's
solely,
because
we
store
models.
So
similarly,
across
all
of
the
different
physics
implementations
like
basically
sort
of
like
whatever
you
need
to
could
like
in
whatever
you
need
to
prolong,
should
be
relatively
straightforward,
yeah
but
they're
like
we
can.
We
can
dive
into
the
implementations
a
bit
more
offline,
I
guess
yeah.
F
A
G
A
A
H
D
What
is
my
next
topic
is
to
make
some
rappers
to
do
time
domain,
but
still
calculating
in
the
frequency
domain,
but
we
have
some
ideas:
how
we
can
probably
bring
the
phrases
down:
24
a
big
3d
model
to
have
all
time
domain
solutions,
but
then
you
can
start
with
one
frequency
and
then
this
result
you
can
give
as
input
to
the
next
frequency.
So
it
doesn't
start
with
zero
or
with
there's
nothing.
So
it
can
go
as
usually
the
result
changed
smoothly
over
frequencies.
D
D
But
yeah,
that's
the
thing
is
I,
think
new
methods
I
think
it
would
be
possible.
It's
not
on
my
list,
but
I
would
be
very
happy
to
or
at
least
discuss
their
very
should
have
to
take.
In
my
points.
The
next
points
are
time
domain
and
general
anisotropy
at
the
modus
tri-axial
and
I.
Such
a
big
I
want
to
see
how
difficult
it
would
be
to
extend
that,
because
that's
part
of
the
to
lose
of
the
project
she's
focus
on
those
and
time-domain.
E
A
D
D
A
Say
so
in
simple
right
now,
I
believe
that
the
the
main
implementation
is
primary
secondary.
We
can
go
in
and
change
the
boundary
conditions,
I
think
that's
only
working
for
a
single
implement,
eight
like
a
single
PDE
at
the
moment,
but
that's
that's
something
we
could
play
around
with.
If
that's
of
interest
to
you
yeah.
D
D
B
Gonna
say
if
you
wanted
a
few
ideas
for
different
ways
to
solve
the
problem:
the
e3d
MT
so
I
like
a
Jif
like
Fortran
code,
the
theory
for
that
is
actually
online,
and
so
there
has
been
some
time
spent
to
to
go.
Think
about
implementation
of
1d
and
3d
boundary
conditions,
and
so
yeah
I
could
share
that
URL.
For
for
that,
so
the
theory
section
is
yeah
I've
been
reviewed
by
Roman
Schekman,
who
who
works
with
our
group
so
that.
F
H
I'm,
just
like
a
question
for
general
interests,
but
like
this
beauty,
queen
makeups
like
so
right
now,
we
do
that
on
a
tensor
match
and
you
and
you
divide
the
mesh
by
two
everywhere-
how
you
work,
for
example,
on
the
node
tree,
where
you
only
be
fine
in
your
area
of
interest
instead
of
refining
everywhere.
You
do
not
really.
Finally,.
D
H
D
A
D
Don't
mind
yeah,
I,
think
I
think
it
should
it
be
to
the
figure
because
most
of
the
internal
parts
wouldn't
change
it's
just
for
the
three
mesh.
We
would
have
to
extend
it
on
how
aids,
cautioning
and
refining
the
grid
basically,
and
so
we
would
have
to
look
into
three
meshes
how
this
is
done,
so
it
sort
of
keeps
track,
because
not
now
at
some
point
like
in
Python,
it's
very
easy
just
to
the
semicolon
semicolon
to,
and
so
it
takes
every
second
no
time,
and
you
have
a
course
already
right.
D
H
D
Probably
development
then
I
also
be
like
at
the
moment
the
iterative
solver.
It's
probably
one
of
the
reasons,
also
why
it
doesn't
use
so
many
memories.
It's
just
solving
for
one
note
if
it's
adjacent
eight
cells
and
then
moving
through
the
whole
3d
cube
right,
and
this
will
probably
also
change
if
you
have
a
tree
mesh,
because
you
wouldn't
have
a
regular
mesh
to
go
through,
it
will
probably
depend
here
how
a
tree
mesh
is
set
up,
I,
not
sure
about
that.
It.
A
Actually
should
be
very
natural
for
the
tree
match,
because,
basically,
how
we
think
about
creating
and
storing
the
tree
mesh
is
like
conceptually.
It
just
starts
from
a
single
cell,
and
then
we
just
progressively
refine,
and
so
actually
the
entire
refinement
structure
is
basically
like
how
the
tree
mesh
is
stored.
So
all
you
basically
would
have
to
do.
Is
you
like
walk?
You
walk
up
that
tree,
so
so
that
should
be
like
a
pretty
natural
extension.
If
you
want
to
play
around
with
like
a
good.
A
You
wouldn't
necessarily
loop
over
X,
but
you
would
we
can
loop
over
like
refinement
levels
of
the
mesh.
If
that
makes
sense,
you
sort
of
have
to
do
everything
at
once.
You
can't
just
like
walk
right
across
X,
because
it's
not
a
tensor
grid
like
it's.
You
can't
just
delineate
what's
going
on
in
only
the
X
direction
and
sort
of
ignore
the
rest
of
the
dimensionality
of
the
mesh
right
yeah,
but
we
can
sort
of.