►
From YouTube: SimPEG Meeting August 4th
Description
Weekly SimPEG meeting from August 4th, 2021
A
So
hello,
everyone,
nice
to
see
you
all
again,
as
always,
I've
stuck
to
the
notes
if
you
like
get
started
on
a
few
agenda
items
here.
So
after
we
get
through
the
quick
reports
on
things,
I'd
like
to
do
a
pr
status
update,
so
we
can
go
through
and
look
at
all
the
pr's
and
if
they
need
assignments,
we'll
review
our
assignments,
we
can
have
people
on
sale
that
volunteer
to
do
that
and
get
things
going.
A
Also,
some
announcements
tebow
is
going
to
be
doing
a
seminar
soon
for
the
for
august.
Tivo
was
the
date
that
we
decided
on
for
that.
A
B
On
that
note
actually,
so
I
think
I
don't
think
she
sends
a
newsletter
yet
and
that's
gonna
be
part
of
the
newsletter,
but
she
also
asked
joe
and
I
that,
if
you
guys
have
updates
that
you
would
like
to
share
and
contribute
to
the
newsletter,
just
send
a
message
on
slack
or
to
lindsay
directly.
So
it's
added
to
the
newsletter.
C
C
Oh
well,
I
might
be
mistaken
just
that
somehow
I
kind
of
thought
we
discussed
you
know
some
time
ago
about
you
may
be
taking
over,
and
then
you
were
concerned
about
it
being
a
little
bit
time
consuming
and
then
lindsay
was
trying
to
convince
you
that
you
know
basically
it's
just
providing
a
format
and
people
providing
stuff
and
but
somebody
to
coordinate
and
send
it
out
on
a
regular
basis.
Anyway.
C
A
A
Let's
move
on
to
some
quick
reports
here,
so
I've
basically
finished
going
through
the
discretized
talks
to
finish
that
this
morning,
just
was
writing
up
here
before
the
meeting
gets
started
to
push
up
there.
So
I'll
have
noted
things
and
they're
just
kind
of
like
conventions
that
should
try
to
stick
to
some
notes,
but
right
so
this
is.
It
was
like
kind
of
my
running
track
of
things
that
I
had
seen
like
just
okay,
well,
to
make
a
consistent
style
for
that
work
that
we're
using
so
it'll,
be
nice,
they're.
A
E
D
One
one
thing
that's
also
gonna
speed
up
working
with
simpeg
is
getting
getting
that
similar.
I
guess
extensions
and
formatting
and
all
that
stuff
so
that
that
website
will
build.
A
D
Yeah
we'll
see
how
much
needs
to
really
be
revisited,
because
the
descriptions
of
a
lot
of
functions
functionality
is
really
it's
a
lot
better.
Now.
D
E
D
Sort
of
what
those
tutorials
need
to
be,
we
don't
we've
now
kind
of
figured
out
well
yeah
in
the
in
the
doc
strings.
We
now
have
explained
what
the
syntax
is
for
constructing
the
stuff
and
applying
it,
and
that
doesn't
really
need
to
be
in
there
anymore.
A
D
Yeah,
okay,
I
guess
I
was
thinking.
I
wouldn't
well
you.
Actually
you
want
to
wait
until
the
api
stuff
is
done
before
we.
You
don't
really
want
to
add
that
to
the
the
pr.
Do
you
no?
No,
no!
Okay,
so
I
mean
you
can
look.
I
don't
even
know
if
I'd
bother
looking
at
it
yet
because
it's
it's
gonna
be
a
few
weeks
away.
A
A
A
G
Yeah,
nothing
too
much
yeah!
I
didn't
get
to
the
pull
request
stuff
this
week.
I
finally
got
my
cluster
run
up
and
running,
so
I've
been
playing
with
that
yeah
having
some
good
success
there,
joe
yeah
you're.
Definitely
right
about
it's
just
that
that
simulation
object,
even
if
you're
just
storing
the
operators.
It
just
seems
that
when
you
get
to
the
big
enough
problem,
it
still
is
a
it's
a
big
graph
like
it's
you're,
passing
around
a
lot
of
data.
So
what
I've
done
is
I've.
G
Just
I've
made
pretty
much
made
like
a
full
or
I
guess,
like
a
sub
simulation
like
the
tile
simulation.
I
just
made
a
light
version
of
it.
That's
just
pretty
much
like
a
json
file
that
sends
and
then
it
each
worker
loads
up
it's
its
tile.
So
it's
I've
kind
of
changed
the
tiling
so
that
it's
now
its
own
worker
and
it's
just
concerned
with
its
own
data
and
it
loads
it
up
on
that
other
server
or
that
other
node
computer.
G
Each
node
is
just
a
tile
and
you
just
send
it
or
what
I'm
working
on
right
now
is
you
just
send
it
that?
Should
that
change
in
the
model
vector
so
that,
if
you
ever
need
to
call
it
the
fields
or
you
want
to
calculate
j
kind
of
thing,
you
just
send
it
to
those
workers
and
they
they
spin
it
up
and
kubernetes
is
actually
really
nice.
For
that.
I
was
trying
to
do
a
simple
version
with
ssh
just
ssh
into
them
and
running
individually,
but
I
yeah
they're
they're.
G
I
swear
that
I
I
I
hooked
up
every
or
set
up
everything
that
it's
identical,
but
there's
still
like
it's
not
finding
certain
stuff
in
sim
tag.
It's
saying
that,
like
the
das
get
field
isn't
there.
But
when
you
look
at
the
thing
it's
clearly
there,
so
kubernetes
is
just
a
lot
nicer
because
then
you
just
spin
up
an
image
and
there
you
go
everything
works
but
yeah.
Hopefully
I
was
maybe
gonna
do
maybe
or
maybe
october.
G
I
should
have
quite
a
bit
of
stuff
to
show
maybe
I'll
sign
up
for
that
one
and
expand
on
this
using
distributed
computing
for
the
inversion.
Oh.
G
Like
a
lot
to
do
it
that
way,
yeah,
it's
not
only
speeding
it
up,
but
well.
I
guess
it's
still
using
a
lot
of
ram
but
yeah
it.
So
far.
I've
got
like
a
two
time
speed
up,
but
I've
only
got
three
nodes
right
now:
hooked
up.
H
G
Yeah,
it's
still
like
when
you,
when
you
want
to
send
it
to
those
workers,
it
still
has
to
send
that
whole
simulation
object,
even
though
that
it's
concerned
with
that.
Well,
I
guess
I
guess
I
have
a
little
bit
of
a
different
problem
where,
when
we
were
doing
it,
we
were
able
to
do
the
like
submeshing,
we
didn't
send
a
global
mesh
to
it
all
the
time
and
with
with
diaz,
I
have
to
send
the
global
mesh
so
you're,
not
really
it's
a
big
size.
H
G
Sorry
yeah,
I
guess
in
that
case,
even
with
the
storing
the
operators
like,
I
tried
the
full
that
full
data
set
that
we
had
from
yeah
that
one
that
we
worked
on.
I
don't
know
if
I
can
pass
on
the
name
or
not
yeah.
I
tried
that
whole
one
again
on
the
new
simulator
on
the
news.
Tiled
simulations
and
the
graph
is
just
huge.
It's
like
40
50
gigabytes,
even
though
we're
only
storing
the
operator
operations,
somehow
that's
getting
sent,
or
it's
still
being
generated.
G
Maybe
it's
on
that
worker
over
there,
but
something
coming
back.
Yeah,
that's
kind
of
where
I'm
at
right
now
is
just
kind
of
mapping
where
everything
is
coming
from
and
where
everything's
going
but
yeah
so
far,
I've
found
is
yeah
we're,
depending
on
sending
a
lot
to
those
workers.
Still
even
the
another.
Funny
thing,
though,
is
it
seems
that
the
frequency
domain
stuff
is
working
better.
G
It
doesn't
actually
make
as
big
of
a
graph
which
kind
of
confuses
me,
but
again
I'm
looking
into
it
and
see.
G
Yeah
yeah,
the
the
problem
I'm
dealing
with
right
now
is
the
dc.
I
find
that
the
the
frequency
domain
stuff
is
working
really
good,
even
on
a
single
machine,
local
cluster,
it
runs
pretty.
It
runs
comparable
to
what's
out
there
for
sure,
especially
3d,
but
yeah
the
dc
code.
For
some
reason
it
just
doesn't
want
to
parallelize
not
on
a
local
machine
anyways.
It
there's
too
much
thread.
Even
if
you
try
to
put
everything
to
a
single
thread,
there's
still
some
crosstalk
going
on
or
something
or
it's
blocking.
I
Maybe
I'm
not
sure
the
only
difference
that
I
can
see
the
dc
code,
it
the
it
passes
when
you're
passing
the
right-hand
side,
you
pass
all
the
vectors,
so
it
actually
generates
a
large
matrix
and
then
apply
the
a
in,
whereas
the
frequency
frequency
so
potentially,
I
think
that's
for
a
small
problem
that
actually
makes
a
speed
up
but
yeah
but
you're
when
you're
dealing
with
a
really
large
problem
that
I
think
d,
that
right-hand
side
at
the
end
like
it
actually
changed
it
to
a
dense
matrix
and
then
yeah
then
applied
so
potentially
that
could
be
a
place
that
that
could
be
a
problem.
G
I
Yeah
right
then
white
you
can
play
loop
over,
but
anyway
it
is
generated
and
stored
in
each
source
class.
So
yeah,
I
I'm
not
sure
it's
the
best,
but
but
but
that's
the
same
case
even
frequent
domain.
It
is
stored
in
a
source
class.
So
if
that's
causes
a
problem,
you
probably
just
generate
them
on
the
fly
but
yeah
it's
yeah.
It's
just
a
just
a
point.
G
J
A
E
D
Yeah
sure
so
I
guess
there's
a
couple
of
things.
One
was
these
validations.
I
was
doing
for
the
time
domain
codes
and
I
mean
lindsay,
took
a
look
at
my
some
of
the
scripts
that
I
had
and
then
she
just
did
a
simple
step
off
and
that
result
seemed
to
be
stable,
but
I'm
not
doing
anything
unreasonable
with
time
stepping
and
it
should
be
stable
because
it's
just
using
backward
oiler.
So
there's
something
in
there.
That's
kind
of
suspect
and
we
haven't
really
investigated
it
yet.
D
But
that
kind
of
opened
up
to
me
just
looking
at
pieces
of
our
our
time
domain
em
code
and
our
em
code
and
kind
of
looking
for
ways
to
improve
it
and
looking
for
ways
to
kind
of
fill
in
all
the
different
kind
of
sources
and
receiver
types
you
could
use
with
all
of
the
different
choices
and
discretization.
D
So
I'm
kind
of
going
through
right
now
and
and
documenting
the
theory,
because
we're
going
to
need
that
for
for
simpeg
when
we
get
to
the
the
api
for
that,
but
yeah
I've
kind
of
been
been
going
in
and
doing
that.
So
I'm
sort
of
in
the
process
of
testing
stuff
trying
to
get
the
theory,
seeing
if
there's
a
better
way
to
do
anything,
see
if
anything's
missing.
That
would
be
really
useful.
D
Everything
was
fine
and
then,
when
it
got
to
below
a
certain
order
of
magnitude,
it
was
unstable,
so
that
and
that
kind
of
leads
back
to
maybe
what
what
joe
was
saying
about,
checking
the
accuracy
of
the
solve,
because
if
you
have
a
unit
step
off
the
magnitude
of
the
response,
I
think
is
going
to
be
it's
going
to
be
higher
than
for
a
square
pulse
and
it's
going
to
be
larger.
D
D
D
So
I
have
you
when
you
turn
when
you
have
the
step
on
you're,
using
kind
of
some
more
course
time,
stepping,
because
that's
not
impacting
your
response
as
much
as
right
before
the
step
off.
So
it's
kind
of
like
coarser
time
steps
you
get
the
step
on
and
then
it
gets
finer
and
finer
and
finer
step
off
you're
still
using
fine
and
then
getting
coarser.
E
A
D
I
The
only
only
case
is
like
when
you
got
very
short
perks,
so
actually
likely
that
rampant
makes
a
large
impact
into
your
off
time.
That's
probably
where
you
need
to
do
what
joe
suggested,
but
it
seems
like
you're
trying
to
emulate
the
step
off
with
the
large
duration
of
the
pulse.
Then,
if
that's
the
case,
I
yeah
it's
it's
unnecessary
in
general,
I
think
that's
what
I
have
tested
when
I
was
just
testing
like
how
to
generate
like
a
just
using
the
long
period
of
step
to
approximate
the
step
off.
I
Confused
like
if
step
off
works
like
why,
like
I
do,
have
a
problem
with
just
putting
a
like
a
certain
period
and
then
see
that's
matching
with
this
step
off.
Is
that
is
that
something
that
you're
having
a
problem.
D
D
I
I
So
I
like,
I
think
we
made
sure
in
that
just
a
conductivity
case.
We
make
sure
that,
but
I'm
not
sure
in
when
you
actually
get
when
you're
actually
applying
like
a
large
magnetic
permeability,
I'm
not
sure
that's
actually
well
tested.
I
don't
know.
I
think
if
that
seems
that
the
case
that
you're
trying
so
a
potential
place
where
we're
getting
a
large
error
is
that
could
be
the
place,
and
I
think
that
the
ubc
code
does
extra
things.
I
I
guess
to
correct
for
that
to
make
sure
that
they'll
divergence
of
b
is
equal
to
zero.
I
think
I'm
not
sure
exactly.
I
think
I
remember
hd
hd
td
dude,
like
that
kind
of
things,
extra
extra
sort
of.
D
It's
true:
there's
there
was
a
little
there.
There
is
a
little
matrix
that
gets
added.
I
guess
to
to
do
that
and
then
all
it
says
it
has
something
to
do
with
hanging
nodes,
but
there
was
some
additional
consideration
in
the
in
the
ubc
codes.
Yes,.
E
D
To
work
the
best
for
for
high
magnetic
permeability
but
version
the
version
two
code
which
uses
an
electric
field.
Formulation
works,
super
super.
Well,
so
I
mean
in
the
end
I'm
kind.
These
are
kind
of
all
the
things
I'm
diving
into.
So
what
is
the
best
discretization
for
solving
certain
types
of
problems?
It's
kind
of
what
we
want
to
know
so
that
we
can
lead
people
down
the
right
path
when
they
start
using
this
package.
A
D
Yeah
but
then
I
mean
won't
if
I,
if
I
discretize
it
coarsely
or
more
coarsely
at
that
time,
then
it's
basically
like
saying:
oh
I'm
just
kind
of
doing
more
of
a
ramp
as
opposed
to
a
true
step
on
and
if
you're
far
enough
back,
it
impacts
the
the
the
data
much
less
then,
as
you
get
closer
to
the
step
off.
D
A
I
Try
that
I
think
I
still
think
that
you
don't
have
to
because
the
this
pdf
that's
the
feature
of
the
bdf
like,
even
if
you
got
actually
wrong
initial,
like
a
timestamping
like
you,
do
a
very
poor
job
on
initial
timestamping,
but
actually
accuracy
gets
converging
as
you
go
later
time,
so
I
don't
know
try.
I
think,
I'm
not
100
sure.
That's
that's
my
intuition,
so
you
can
check.
D
D
I
It
yeah,
so
we
do
sort
of
that
by
using
the
vector
potential
when
they're
generating
the
source
term
but
yeah,
but
then
it's
not
necessarily
still
like
in
in
when
you're
calculating
the
b
field
by
solving
the
pde.
We
don't
really
make
sure
that
so
yeah,
I
don't
know,
what's
what
could
happen
with
the
large
permeability.
K
H
Yeah,
so
I've
been
exploring
ways
to
be
able
to
go
from
surfaces
like
a
dxf
to
a
black
model.
It's
an
operation
that
we
have
to
do
fairly
often
right
and
I
was
relying
on
gocat
before,
and
I
was
trying
to
look
for
an
open
source
way
to
be
able
to
do
this.
So
we
had
vtk.
If
you
guys
remember
right.
H
Tkc
were
using
btk
for
this,
but
there
were
some
weird
behaviors
that
if
the
surface
were
not
closed,
and
so
I
found
this
try
tree
mesh
package
that
does
ray
tracing
and
it's
actually
quite
quite
fast.
So
I
recommend
you
guys
to
look
at
it
if
you
have
to
do
that
operation
and
then
otherwise,
I've
been
diving
back
into
the
the
desk
potential
fields
modules,
because
we
need
to
deliver
things
and
it
was
not
giving
the
same
result
as
before
so
some
efficiency
there.
H
So
john,
I
don't
know
if
you,
if
you're
touching
our
tiled
simulation
branch,
but
I've
been
working
on
our
fork
and
now
I
would
have
some
pretty
good
updates
to
to
push
back
yeah.
G
H
Yeah,
so
I
will
do
then
a
pull
on
my
branch
to
see
what
you
did
and
then
yeah
send
it
back
cool
it
should
fix.
You
should
fix
some
good
stuff
for
grab
mag
cool.
H
Well,
we
promised
that
we
would
give
access.
You
know
an
analyst
hookup
to
to
the
latest
simpag.
You
know
basically
update
from
the
old
version
to
zero,
14
and
plus
by
the
end
of
I
think
by
october
we
said
so.
I
just
need
to
make
sure
that
simp
x014
runs
fast
and
gives
the
output
that
we
used
to
have
right
so
yeah.
That's
that's
deliverable.
H
Hey
for
sure,
I
need
to
switch
off
joe,
but
paying
me
man
if
you
want
me
to
be
in
an
editor
on
any
of
these
guys,
just
flag.
A
I
think
the
only
one
I
wanted
to
ask
you
about
really
quickly
was
the
xiao
long
have
the
celebrate
or
the
cross
radio.
H
Right
you
want
to
do
this
on
this
pr
or
we.
I
would
just
probably
split
that
right.
A
H
I
don't
remember
either,
but
I
would
say,
let's
just
make,
you
know,
make
a
decision
for
this
one
in
line
with
what
we
want
to
do,
basically
always
normalized
by
volumes
and
then
and
then
we
can
readjust
the
other
regs
after
in
a
separate
pr.
That.
H
A
Okay,
so
I'm
just
gonna
start
at
the
most
recent
one
and
go
from
there,
so
the
number
1027
replace
mp
matrix
by
regular
nd
arrays,
I'm
not
sure
where
so
is
it
just
failing
on
the
code
coverage
right
now
it
looks
like
it's
just
failing
on
code
coverage
for
the
diff.
This
was.
A
B
Yeah,
I
think,
mike
your
comment.
Yeah,
it
seems
like
this
code
is
not
like
the
part
of
the
code
that
I
wrote
that
was
modify
is
like
not
not
tested
and
yeah
like,
so
that
was
like
the
analytic
for
the
mt
1d
that
we
I
transferred
from
my
from
one
of
my
very
first
projects
in
simpek
to
to
compare
and
with
the
work
that
soggy
and
they
did
with
the
leading
edge.
But
I'm
not
sure
it's
been
used
since
then,
and
we
didn't
write
any
tests.
A
Okay,
yeah
from
what
I've
seen
it
looks
like
it
was
done
like
the
over
was
done
correctly
from
matrix
to
array
there.
I
couldn't,
I
didn't
see
anything
else
so.
D
A
A
B
That's
a
draft,
I'm
not
sure
if
we
want
actually
to
do
that,
but
I
know
that
yeah
that
you
can
write
things
as
draft
in
github.
So
I
actually
that's
what
I
I
don't
worry
about.
B
M
A
That
one's
fine,
okay,
total
vector
map
dom
mark
that.
B
Is
in
progress
so
yeah
we
we're
still
not
very
happy
with
what
we're
getting
it's
kind
of.
There
is
some
like
yeah,
some
weird
things
happening
like
this.
It's
a
difficult
problem.
B
Okay,
I
think
he's
implementing
what
he
saw
in
that
from
that
man
and
also
where,
like
it's
mvi,
but
you
regular,
it's
cartesian
mvi,
but
you
regularize
on
the
unfitted.
A
All
right:
okay,
let's
see,
there's
a
remove
deck
coat
and
somebody's
pretty
small.
F
A
I
Yeah
the
generating
like
where
we
used
that
to
generate
the
current
density,
like
do
you
remember
mike,
was
asking
us
to
provide
like
oh.
How
do
we
generate
that
wire
loop,
like
a
wire
wire
line,
source.
I
Yeah,
so
do
you
like
it?
You
kind
of
now
confident
that,
like
because
there
were
a
bit
of
issues
when
you
first
implemented
that
into.
A
F
It's
also
an
inherit
problem
that
we
have
with
voxel-based
code
right
because
it's
an
it's
an
average
of
our
cells
anyway.
So
it's
if
you're,
not
if
you're,
too
close,
it's
not
precise,
probably
in
indifference
to
to
finite
elements
where
you
can
neatly
form
your
cells
along
the
source.
In
our
case,
that's
just
impossible.
F
A
Okay,
okay,
so
it's
moving
on
fix
for
issue
771,
I
think,
is
this
something
still
to
do
with.
A
A
B
Yeah
actually
joe,
so
we
like,
we
haven't,
we
have
an
even
older
pr
for
that
771
issue.
Yeah.
E
B
I
A
I
Maybe
actually
it's
probably
worthwhile
to
deprecate
that
gaussian
newton.
I
think
because
yeah,
I'm
not
sure
why,
like
like
you,
can
basically
use
it
in
exact.
An
exact
gosh
newton,
always
sure
you
like
make
sure
you
increase
number
of
cg
to
make
sure
you
solve
well
and
then
then
what
you
get
should
be
almost
equivalent
to
the.
A
Okay,
gaussian,
we
I
mean
we
should
definitely
do
something
with
that.
I'm
not
sure
what
we
can
do
right
now,
either
way
like
both
those
things
I
don't
think
are
good
solutions
to
it.
Looking
at
them,
the
next
one
extend
base
source
for
poland
dc
or
minor
extensions
soggy.
Would
you
mind
looking
through
this
one
number
10
7.
A
M
A
E
G
I
just
gotta
write
some
more
tests,
just
the
yeah,
the
code
coverage
is
just
failing
still
so
there's
I
started
to
go
through
it
a
little
bit.
I.
G
So
yeah
I'll
work
on
it
for
sure,
okay
cool!
No
I'm
just
don't
worry.
G
A
A
A
F
But
you
could,
you
could
just
put
the
one
liner.
He
has
put
that
here.
A
F
A
A
F
F
F
A
I
E
A
So
then,
the
next
one
there
is
the
representation
for
survey,
sources
and
receivers.
I
know
we
kind
of
let
that
one
get
off.
F
A
I
I
don't,
I
think
I
might
have
pushed
a
change
or
two
in
there
you'll
put
some
changes
here.
Yeah.
I
think
I
was.
A
Okay,
so
there
we
have
this
examples
mt1b
from
about
two
years
ago,
now
we're
almost
to
the
end
of
them.
What
do
we?
We
have?
We
have
one
tmd
examples
in
there
right
now,
because
it's
no
longer
needed
because,
right
now,
it's
just
a
it
was
like
literally
a
tutorial
or
example.
I
Yeah,
I
think
it's
just
probably
worthwhile
to
replace
like
not
all
of
them,
but
just
put
simple
like
I
want
example
using
the
analytic
analytic
function.
I
guess
that's
pretty.
A
M
I
What
do
you
think
soggy?
I
think
it's
all
come.
Maybe
I'll
come
together
at
some
point.
I
think
that
is
also
working
hard
on
writing
inversion
module
as
like
it's
a
text,
but
I
think
that's
more
kind
of
providing
a
theoretical
point
of
view
of
how
simple
or
how
the
inversion
works.
So
maybe
that
can
be
related
to
this
at
some
point:
change
something
but
yeah.
I
think
we
could
just
let
it
open,
although
that's.
I
Realize
that
you
closed
that
that
em-1d
stuff
do
you
think
I
was
like
good
enough
to
be
merged
or
what
was
sort
of
your
intentions?
I
was
kind
of
revisiting
and
wanting
to
make
sure
like.
K
One
it
was
number:
where
is
that,
let
me
see
it
is
merged.
Thirteen
thousand
thirteen,
oh,
is
it
merged.
I.
I
A
I
Sounds
good,
then,
can
you
reopen
and
yeah
yeah
it's
kind
of
timely,
so
I'm
working
on
writing
a
book
about
em
induction
for
groundwater,
and
I
want
to
actually
use
this
tools
and
as
a
separate
resources
for
the
book.
So
I
want
to
make
sure
all
all
the
functions
of
that
we
have
is
working.
I
think
it's
actually
working.
All
of
them
are
working.
It's
passing
the
test.
I
just
need
to
make
sure
like
update
the
tutorials
that
devin
generated.
A
A
Of
course,
might
just
going
to
miss
click.
Thank
you
for
bringing
that
up.
C
I
Oh,
that
was
my
mistake
like
I,
no
I
they
were
like.
That
was
my.
That
was
my
band
there's.
No,
there
were
no
problem.
Okay,
so
I
had
to
like
I
was
playing
with
like
the
signs
and.
I
There
right
the
the
difficulty
is
now
the
the
current
current
version
has.
Actually
you
use
like
a
you,
find
all
the
unique
pairs
of
sources
at
the
same
height
then.
I
Like
then,
I
then
yeah,
then
I
need
some
sort
of
index
like
that.
Like
that
said,
oh
it's,
a
single
height
like
so
yeah
I've
had
before,
but
now
it's
somewhat
ambiguous.
So
I,
I
think,
all
the
functionality
I
think
it's
working,
but
they
just
kind
of
bit
of
a
logistic
problem.
How
are
we
going
to
set
it's
an
it's
a
it's,
a
problem,
inverse
problem
because
we
can
set
like
number
of
heights
for
each
of
the
source,
but
that's.
I
We
need
to
think
about
a
good
way
to
index.
Okay,
we're
still
inverting
for
single
height
that
I
wasn't
100
clear
how
what's
the
good
way
to
go?
Oh,
it
should
be
pretty
simple
just
using
the.
A
Like
I
just
map
between
the
topography,
I
got
typography
xyz
said
to
there's
to
the
source
location.
I
Like
we
need
something
what
like
similar
to
what
that
did
like,
so
we
need
some
sort
of
reference
height
and
then
everything
is
relative
like
so,
we
got
you're
just
inverting
for
reference
height
and
then
other.
A
A
I
Yeah,
so
I
was
hoping
we
could
deal
with
that
a
little
bit
later,
because,
although
it's
actually
probably
simple
but
need
some
thoughts,
I
think
and
and
then
it's
probably
need
extra
parameters
in
survey
class
like
like.
I'm
not
sure.
So,
if
you,
if
you
have
a
great
idea,
we
could
we
could
just
implement
quickly
and
merge,
but
I
thought
it's
not
necessarily
important.
A
It'll,
just
my
thought
would
be:
okay,
just
use
like
a.
However
you
have
your
mesh
defined
like
you
can
have
like
a
2d
mesh.
If
you
have
them
have
the
sources
in
3d
like
in
on
a
plane
right,
you
just
have
a
2d
mesh
and
then
just
use
an
interpolate
to
source
location.
I
A
I
So
anyway,
so
that's
that's
where
mi.