►
From YouTube: SimPEG meeting October 21, 2020
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
Perfect
well,
I
know
it's
been
well
a
couple
of
weeks
for
some
of
us
because
of
scg
and
the
mirror
event
and
other
things
like
that
yeah,
so
we
can
get
caught
up
on
some
of
those
things
and
maybe
even
a
bit
of
an
update
as
what
was
the
the
last
meeting
that
some
of
us
would
have
missed
so
joe.
Would
you
like
to
start
us
off
yeah.
B
So,
with
last
the
last
meeting
we
ended
up
talking
about
or
the
em-1d
stuff
kind
of
deciding
like
at
least
talking
about
how
we're
gonna
get
that
in
from
what
I
remember,
essentially
from
I've,
been
working
towards
that,
as
well
as
trying
to
get
what
I've
worked
on
into
geogeoanna
and
then
using
geoana
inside
synthetic
m1d.
B
B
I
was
looking
at
doing
the
lagged
convolution
for
the
prediction
error
for
the
for
the
transform,
because,
right
now,
what
it's
doing
is
it's
just
doing
the
normal
digital
linear
filter
the
ankle
transform
where
it's
essentially
computing.
You
know
a
bunch
of
lambdas
for
each
single
point
right
yep,
so
I
was
just
trying
to
look
at
how
it
was
doing
that
and
what
I,
the
issue
that
I
ran
into
is
that
it
had
like
it.
B
B
C
No
it's
just
if,
if
you
do
the
the
thing
is
how
it
goes
into
the
kernel.
If
you,
if
you
do
the
normal
dlf
type,
then
it's
it's
a
matrix,
because
each
frequency
has
a
whole
bunch
of
each
location
has
a
whole
bunch
of
wave
numbers
right.
And
then,
if
you
do
the
lag
convolution,
then
you
have
just
one
array
and
then,
when
you
come
out
of
the
kernel
in
the
transform,
you
have
to
reshuffle
that
and
interpolate
it.
So
that
then,
for
the
actual
dlf
each
location
has
the
array
of
lambdas.
It
needs.
B
So
from
what
I
was
looking
at,
as
I
said
through
your
code
essentially
like
that
reshuffle
thing,
just
kind
of
like
copies,
a
part
of
it
and
moves
it
copy
it,
like
kind
of,
does,
goes
like
this
down
through
that
array
yeah.
So
it's
like
it
just
kind
of
shuffles
it
by
one
each
time
as
it
turns
that
into
a
full
matrix
of
solutions.
Are
you
like
the
maybe
that's
what
it
looked
like?
It
was
doing
from
what
I
saw,
but
that
was
still
just
kind
of
like
the
angle.
C
B
B
C
Well,
I
yeah
could
be,
I
think,
but
the
only
thing
I
looked
at
sorky
as
what
was
a
loops
cop
horizontal
loop.
So
there
is
no
angle
factor
because
it
doesn't
matter
right.
C
B
B
B
Okay,
I
guess
the
other
thing
is
that
when
I
was
playing
with
the
tiled
or
with
the
thing
with
the
large
convolution
part
of
it,
I
wanted
to
be
able
to
like
give
it
a
list
of
have
it
like,
I
guess,
have
it
repeat
over
frequencies.
C
B
D
What
what
was
the
current
implementation,
joe
like,
were
we
using?
I
thought
we
were
just
using
the
regular
angle
fht.
Is
that
what
we
were
using
yeah
you're.
B
C
C
D
C
But
then
you
have
to
be
careful,
because
if
you
do,
if
you
come
from
the
time
domain,
you
shouldn't
do
the
convolutional
in
the
frequency,
because
then
you
have
to
loop
over
offsets
anyway,
and
it's
going
to
be
more
costly
than
if
you
do
the
standard
ankle
right.
So
this
is
only
worth
it
if
you
want
to
have
frequency
domain
data,
not
if
you
want
to
have
time
domain
data.
D
Yo,
what's
sort
of
the
like
you're
concerned,
I
think
it's
in
general,
the
current
structure
is
being
is
like
for
a
single
for
a
single
transmitter.
You
got
a
single
datum,
well,
not
multiple
time
channels,
but
still
it's
a
single
receiver
in
that
sense
that
we
could
do.
I
think
devin
expanded
to
have
xyz
yeah,
so
yeah,
if
you
can
like
compute
everything
together.
That
would
be
better,
but
is
that
what
you
were
concerned
like
a
kind
of
in
in
efficiency,
issue
of
kind
of.
B
B
D
Yeah
that
will
that'll
be
nice
because
yeah
you're
right,
if
the
like
the
height,
doesn't
change.
It's
basically
same
so.
B
If
it
did,
even
if
the
height
did
change,
I
think
we
could
use
the
lag
convolution
operator
we
just
have
to.
We
would
have
to
take
out
part
of
the
you
would
have
to
kind
of
modify
some
of
the
internal
code
or
just
use
the
parts
that
we
needed,
because
the
only
thing
that
the
the
large
convolution
does
right.
It
takes
that
the
output
and
then
tiles
it
for
each
input
or
for
each.
B
It
takes
the
frequency
or
the
wavenumber
solution
right
and
it
tiles
it
for
each
observation
point
shifted
by
one
so
that
after
it
does
that
tiling
shift
by
one,
then
we
could
apply
the.
C
But
the
kernel
is
the
more
expensive.
You
know
if
you
change
the
higher.
C
D
I
probably
double
check,
but
I
guess
your
that
reflection
coefficient
changes.
I
guess,
depending
upon
it.
E
B
D
Great
joe,
I'm
glad
that
you're
actively
working
on
this
and
you
guys
probably
don't
recognize
guys
actually
joe.
We
wrote
that
fortune
code
in
c
and
generalize
it
that
it
can
handle
like
a
derivative
of
chi
derivative,
each
height
like
a
thicknesses
yeah,
so
yeah,
I'm
very
happy
and
it's
actually
quite
faster
but
as
joe
said
that
so
yeah
it's
great.
E
I
guess
if,
if
I
understand
the
sort
of
status
of
this
package
right
now,
we're
still
we're
still
sort
of
actively
optimizing
it,
making
it
faster.
There's
some
pretty
significant
improvements
we
want
to
make
before
we
actually
bring
it
into
simpeg.
B
Obviously
I
would
say
that
yeah,
okay,
that's
what
I've
been
actively
trying
to
figure
out
here.
Just
this
yeah.
I
really
like
to
get
the
lagged
convolution
thing
and
just
like
I
said,
there's
the
the
two
issues
that
I
kind
of
saw
was
that
it
wouldn't
the
whole
tiling
operation
that
is
done
within
that
dlf.
C
B
C
B
C
D
Joe
at
the
end,
what
you're
trying
to
do,
because
suppose
you
got
10
000
sound
invocation
and
then
that's
where
you
want
to
benefit
from
the
legit
convolution
or
thinking
about
basically
single
sounding
and
you
got
potentially
got
multiple
receiver
locations
for
a
single
like.
D
D
So
one
interesting
use
case
is
like
it's
a
it's
a
large
loop
but
multiple
receiver.
D
A
With
all
of
this
work
too,
do
you
think
I
mean
it
sounds
like
there
could
be
a
really
nice
paper,
just
specifically
about
sort
of
simple
m1ds
talking
a
bit
about
the
the
code
optimizations.
All
of
that
and
then
maybe
a
couple
nice
use
cases
I
mean
it
would
be
great
if
we
had.
You
know
a
compelling
synthetic
and
a
compelling
field
data
set.
Is
that
something
you
kind
of
have
in
mind,
as
is
a
bit
of
a
goal,
or
is
that
something
we
can
put
on
the
radar?
A
Because
I
mean
this
is
a
ton
of
work,
so
it
would
be
good
to
make
sure
you've
got
some
some
mileage
out
of
it
beyond
just
it
being
experienced.
E
Well,
I
think
the
synthetics
are
already
done
so
I've
got
about
eight
to
ten
tutorials
already
for
everything.
From
like
the
stitched
stitch,
stuff,
fem
tem
ford
simulation
inverse
the
whole
suite
of
everything
is
actually
in
tutorials
already,
but
a
an
actual
field.
Example,
I
think,
would
be
the
missing
piece
for
that.
A
Yeah
yeah,
but
I
think
it
also
takes
a
bit
of
thought
to
sort
of
think
through,
like
what
would
you
want
to
include
in
the
paper,
because
I
don't
think
that
you'd
want
to
include
like
everything
that
we
would
do
in
sort
of
tutorials
and
documentation.
That
needs
to
be
there
by
the
time
the
paper's
ready
completely.
A
D
Yeah
so
joe
one
interesting
example
that
I
was
thinking
was
boo
pronoun.
So
what
what
we
have
currently
is
the
resolve
and
the
skytem
acquired
at
multiple
times.
D
So
there's
two
year
interval
actually
decline,
and
one
of
our
like
case
story
sort
of
handles
that,
but
I
think
there's
an
interesting
case
where
you
can
invert
them
together
and
see
how
much
changes
have
made
and
that
the
current
that
won
the
stitches
structure
is
kind
of
nice,
putting
dumping
everything
into
a
single
inversion
and
then
all
of
them
together
see
if
there's
a
changes.
D
Anyways
it's
just
like
this.
Is
it's
because,
like
what
we
have
told,
it's
actually
done
right.
Other
people
already
done
it,
but
I
think
that
there
are
a
few
ways
we
can
sort
of
make
it
a
good
example
and
also
talk
about
the
technical
aspect
of
how
we
optimize
that
I
think,
we've
probably
to
weave
it
together.
Well,
because
yeah
we're
just
talking
about
the
code,
it's
not
very
exciting
because
I
mean
it
has
been
a
while
right.
If
you
leave
things
together
right,
I
think
there
could
be
an
interesting
paper.
C
Yeah
yeah,
it
would
definitely
help
if,
if
you
can
line
it
out
a
bit,
so
I
understand,
but
because
I
think
it's
mainly
an
interface,
how
we
talk
to
each
other
now
how
we
implemented
the
dlf
in
m1d,
because
I
think
you
can
do
most
stuff
at
once.
You
can
do
arbitrarily
a
number
of
offsets,
but
yeah
offsets
and
so
lambdas
are
the
number
of
frequencies
and
arbitrary
numbers
of
layers
in
one
go.
C
D
B
F
We
were
just
not
acting
on
it.
Is
that
the
we're
just
taking
a
step
on
it.
F
B
It's
important
just
to
kind
of
give
the
idea,
like
you
know,
as
I
was
testing
this,
the
kernel,
the
the
forward
operation
is
essentially
the
same
speed.
B
D
Wow,
I'm
I'm
thrilled
to
test
this
so
usually
like
what
sort
of
the
size
of
the
inversion
that
I'm
running
is
about.
D
20
000,
20
000
soundings
and
I'm
using
usually
16
core
takes
about
three
to
four
hours
to
run.
I
don't
know
20
iterations
or
30
iterations
of
the
one
year
to
yeah.
So
if
the
sensitivity
is
the
most
expensive
part,
so
if
you
can
reduce
that
to,
I
don't
know
10
times
faster,
it'll
it'll
be
really.
F
The
since
these
are
not
recalculating
between
at
each
cg
et
terror,
right,
it's
it's
stored
and
just
applied
as
a
as
a
matrix.
All
right!
So
that's
correct.
D
A
That's
really
exciting
joe.
Do
you
have
any
other
updates
or
things
you'd
like
to
chat
about.
B
I've,
I
I'm
pretty
sure
I've
gotten
the
discretized
to
what
I
would
like
it
to
be
again.
It's
still
if,
when
we
push
it,
it's
going
to
break
what's
in
scenic
right
now,
because
there's
a
few
things
in
there
that
are
relying
on
those
shaped
like
opera
shape,
like
properties
to
be
numpy
arrays,
where
this
will
be
something
that
we
want
them
to
be
tuples
or
I
want
them
to
be
temples,
at
least
so
they're,
not
modifiable
and
yeah.
So
they
will.
B
That
does
break,
as
I
run
through
the
tests
on
some
peg,
but
that's
the
only
part
that
breaks
so
we
might
need
to
put
a
we'll,
probably
have
to
put
a
depend,
a
maximum
discretize
dependency
on
synthetic
right
now
and
then
do
a
quick
passover
to
make
to
go
catch.
Those
parts
where
it's
using
it
like
that
yeah
I
I've
gotten.
A
B
I
mean
I'm
pretty
happy
with
where
it's
at,
like
I
said
when
I've
been
using
with
simpeg.
The
only
errors
I
found
was
those
kind
of
errors
like
it
all
imports
just
fine.
It
all
starts
up
just
fine
just
when
you
start
using
it.
It
breaks
because
of
that
assumption.
That's
the
only
thing
that
goes.
Okay
purposely
wanted
to
do
that,
and
I
tried
to
put
a
warning
there
that
this
will
cause
code
to
break.
This
is
intended
right
so.
B
That
part
of
it's
a
little
bit
scary,
which
is
like
so
well
as
soon
as
I
push
that,
like
the
only
kind
of
the
only
reason
I
was
have
been
a
little
bit
hesitant
to
merge.
It
is
because
of
I
still
have
no
idea
what's
causing
the
that
sim
peg
error
on
testing
and
the
segmentation
fault
error.
It's
a
segmentation
fault
and
I'd
like
to,
and
I
want
that
passing
so
that
when
we
updated
it
it'll
disseminate
on
through
everything
still
like.
B
B
Yeah
I
kept
trying
to
like
get
the
docker
image
that
travis
uses
to
test
the
environment,
and
I
I
got
something
close
to
it
and
that
still
didn't
even
reproduce
the
image
or
pre-produced
the
sick.
A
Exactly
cool
tier,
would
you
like
to
go
next.
C
Not
not
much
for
me.
What
is
that
is
the
misfit
that
it's
gonna
be
implemented
in
emg
3d
in
the
future.
C
It's
going
to
be
very
similar
to
the
standard
one
as
well
as
the
same
as
in
yeah
in
image.
3D,
it's
going
to
be
very
similar
to
the
one
in
simpek.
We
first
had
a
misfit
from
the
paper
from
the
plastic
small,
the
paper
with
the
joint
state
gradient,
but
yeah.
That
was
a
misfit
that
had,
at
the
same
time
a
offset
rating,
a
frequency
rating.
C
C
So
we
go
back
to
one
over
standard
deviation,
yeah,
which
will
make
it
easier
to
compare
it
to
simpek
and
maybe
even
if,
if
he
wants
to
get
jtvx
implemented,
that
it
would
be
the
same
yeah
other
than
that,
it's
mostly
the
the
3d
paper
that
I
just
want
to
get
away.
C
So
I'm
not
spending
any
more
time,
but
that
I
think
shaped
up
nicely
excited
to
see
what
what
the
reviewers
gonna
say
and
yeah.
I
seem
peg
wise,
not
much.
D
I
just
wanted
to
add
what
theater
was
mentioning
about
that
the
sensitivity
kernel
like
what
from
emg
3d.
It's
currently
considered
the
data
as
a
complex
value,
whereas
what
we're
doing
is
dissecting
that
into
real
and
imaginary,
but
I
think
theater
what
we're
doing
is
basically
still
kind
of
doing
a
cup
like
considering
a
complex
value
when
we're
solving
at
the
very
end,
we're
sort
of
breaking
that
apart
to
be
on
an
imaginary,
so
yeah.
We
can
still
do
the
exactly
same
thing
in
in
your
code.
C
Exactly
yeah,
I
think
that's
the
so
it's
going
to
be
the
same.
If,
because
I
I
I
don't
have
to
give
it
in
as
a
complex
source
that
the
residual
to
back
propagate
and
then
it
comes
back
as
a
gradient.
And
then
you
just
have
to
be
careful
that
we
give
the
same
weighting
so
that
the
real
and
imaginary
that
is
stored
in
simpac
together
gives
the
same
as
the
complex.
In
my
case,
sort
of
yeah.
F
I
can
so
I
spent
a
fair
amount
of
time
this
weekend,
looking
into
the
clients
of
a
desk
right.
F
So
apparently
the
way
we're
doing
it
is
just
a
direct
compute,
which
implies
that
it
only
works
for
a
single
machine,
and
I
found
some
pretty
important
efficiency
gain
if
we
start
using
clients,
because
mostly
it
has
like
a
scatter
function
to
be
able
to
to
maintain
data
on
the
threads
and
all
that
and
we
had
a
major
bottleneck
when
you
started
using
the
tiling
on
the
cg
cg
solves,
so
I
basically
rewrote
the
cg
as
a
desk
version
so
that
we're
basically
doing
all
the
iterations
as
a
in
a
single
graph.
F
Instead
of
like
going
down
the
iterations
and
it
looks
like
it
makes
a
huge
difference.
John,
you
probably
remember
right,
seeing
the
cycles
of
cpus
going
up
and
down
during
the
cg
stuff,
and
that's
just
just
because
of
that
right.
We
had
like
highly
paralyzed
like
derivatives
there,
but
then
we're
going
after
this
into
a
single
threaded
part
and
then
kicking
back
again
so
by
basically
mapping
the
entire
cg
as
one
task.
F
We
now
have
like
full
efficiency
on
the
on
all
the
processors
and
then
using
clients
also
opens
up
the
possibility
of
being
able
to
spin
computations
on
multiple
nodes
right
because
technically
we
just
need
to
submit
a
client
sorry
workers
ip
addresses.
So
I
think
that's
the
way
to
go.
I'm
going
to
keep
testing
it.
I'm
running
like
a
a
series
of
inversions
right
now
using
the
old
method
and
the
the
new
one.
F
It
looks
right
now.
At
least
I
have
a
2x
faster
and
memory
management
is
a
lot
a
lot
nicer
too.
So
that's
what
I
spent
time
this
weekend
and
then
reviewing
devons.
You
submitted
some
changes,
good
catch
on
the
on
the
error
of
the
potential
fields.
I
think
we
still.
We
should
still
get
try
to
not
do
logical
indexing
yeah.
You
can
go
away
with
arguments.
You
know
just
thresholding
the
arguments
that
will
be
they'll
be
better,
but
that's
that's
it
yeah
yeah.
E
I
guess
I
I
don't
want
to
arbitrarily
add
a
small
tolerance,
because
I
mean
that's
the
action
of
just
like
shifting
all
of
the
data
over
by
a
little
bit.
So
I'd
like
to
do
something
that
actually
recognizes
if
the
locations
are
over
a
a
node
or
over
an
edge.
So
if
you've
got
a
more
clever
way
to
do
it,
that
was
just
the
easiest,
but
I
mean.
E
It
ends
up
being
like
really
bad,
because
you
have
some
terms
that
basically
you're
going
to
be
dividing
by
almost
zero.
No,
but
that's
that's
the
point
of
the
that's
the
point
of
a
small,
constant
yeah,
but
if
you
start
shoving
them
in
there
I
mean
the
the
errors
over.
The
nodes
were
like
15
percent.
F
No,
I
I
know
with
the
old
implementation
that
it
sucks,
but
if
you
could
try
to
just
add
the
epsilon
on
all
the
arguments,
instead
of
doing
the
logical
indexing,
because
logical
analysis
is
gonna,
it's
gonna
be
costly
right,
then
yeah
and
it's
like
a
numpy
operation.
It's
not
sure
how
that's
gonna
it's
gonna,
handle
and
all
that.
E
F
To
change
on
my
side,
devon
and
if
you
make
sure
that
you
apply
it
on
all
the
args,
you
know
the
way
that
the
way
that
john
rewrote
it
a
bunch
of
args
right
so
just
add
epsilon
on
this
alone,
and
then
that
seems
to
solve
it
on
my
side.
So
try
it
on
your
side
too.
E
E
Oh,
I
might
be
able
to
oh
yeah
so
when,
when
you're
doing
the
mag
forward
problem,
if
the
observation
location
is
over
an
edge
or
over
a
node,
you
have
some
terms
in
it
that
are
like
zero
divided
by
zero
and
analytically.
Those
terms
would
cancel
out,
but
numerically
it
gets
angry.
E
And
so,
when
I
compare
the
ubc
code
with
the
simpeg
code,
the
simpeg
added
stability
by
adding
tolerances
to
make
sure
you
would
you
were
never
dividing
by
zero,
but
it
propagated
error
through,
whereas
the
ubc
code
being
written
in
fortran
they
actually
just
they
looked
and
found.
Oh
I'm
over
a
node
or
I'm
over
an
edge,
I'm
not
going
to
use
this
term.
So
it
is
a
hundred
percent,
analytically,
accurate.
E
So
what
I've
been
doing
now
is
playing
around
with
a
way
to
keep
the
sim
peg
code
quick,
but
also
make
it
stable
and
right
now
the
solution
has
been
to
look
for
the
locations
that
are
over
nodes
and
edges
and
then
add
some
kind
of
shift
them
over
a
little
bit.
E
F
F
G
E
E
I
don't
know
if
that
answers
your
your
question.
I
bet
I
mean
in
in
practice.
I
think
it's
pretty
unlikely
that
that
you're
going
to
be
having
a
large
amount
of
your
locations
actually
be
above
edges
or
nodes.
This
is
something
that's
you
know.
If
somebody
does
a
synthetic
they're
going
to
potentially
do
it's
actually
pretty
likely
that
they're
going
to
be
doing
it
unless
they
decided
to
put
them
at
sell,
centers
or
oversell
centers.
E
Oh,
no,
no!
No!
It's
it's!
It's
in
simpek
yeah,
because
it's
it's
not
like
it's
not
like
a
finite
volume,
it's
an
integral
solution
and
and
it's
getting
constructed
within
simpeg
and
so
we're
doing
a
parallelization.
I
think,
to
each
each
row
can
be
constructed
independently.
I
think
that's
how
the
parallelization
works.
A
Thanks
appreciate
the
overview
and
dom
I
don't
know
if
maybe
in
a
couple
of
weeks
or
whenever
it
makes
sense,
if
there's
a
bit
of
a
demo
or
or
something
like
that,
a
bit
of
a
walk
through
that
you'd
be
willing
to
give
just
to
sort
of
get
us
up
to
speed
on
how
the
client
implementation
with
das
works.
I'm.
I
certainly
would
appreciate
that,
and
I'm
sure
others
might
as
well.
F
Yeah
I
was
hoping
to
be
able
to
show
you
show
you
some.
You
know
some
graphs,
but
my
computer
is
still
running
it
so
next
week
I
can,
I
can
show
you
yeah,
basically
showing
the
difference
between
all
the
implementation
over
problem
size
and
over
a
tile
number
of
tiles,
or
so,
if
we
increase
the
tiles,
ideally
we'd
like
to
be
able
to
decrease
the
computational
time,
but
that's
not
what
we're
seeing
with
the
current
implementation.
F
D
F
And
jt:
well,
if,
if
your
jvec
gt
vex
are
not
bask
operations,
you're
not
going
to
see
the
the
effect,
because
dusk
will
stop
at
every
every
jt
vac
can,
like
you
know,
do
the
separation
in
line.
But
this
is
really
it's
going
to
be
a
gain.
If,
if,
for
instance,
yeah
the
action
is
a
stack
of
desk
operations,
and
then
this
is,
you
know
a
series
of
those
into
the
cg.
F
This
is
where
it's
gonna
really
gonna,
going
to
shine
right,
because
now
the
whole
thing
is
just
one
big
graph
that
the
desk
can
chunk
through.
So.
D
H
I'm
currently
kind
of
headed
towards
their
soggy,
with
the
width.
F
F
We
just
need
to
figure
out
how
this
is
done
right
do
either.
We,
you
know,
submit
a
misfit
function
to
a
client
or
I'm
not
too
sure
at
that
point,
but
it
opens
up
the
potential
for
that.
F
Yeah,
do
you
you
work
with
multiple
ip
addresses?
How
do
you
yeah.
H
Yeah
we
can.
We
have
it
set
up
for
that
right
now,
I'm
not
really
on
to
that,
but
yeah.
It's
already
set
up-
and
I
was
just
gonna-
try
and
start
doing
that
yeah.
We
have
about
five
computers
that
we
can
just
kind
of
shoot,
give
it
the
ip
address
and
just
spawn
off
a
problem.
H
H
A
That's
really
exciting
tom,
it's
very,
very
cool
john.
Do
you
want
to
go
next.
H
Yeah
yeah
I've
been
actually
fighting
with
the
mt
a
little
bit
more
here
yeah.
It
was
going
good
for
a
while
with
the
impedances
and
then
this
last
project
they're
looking
to
do
like
a
really
low
frequency,
like
0.2
hertz,
to
about
a
thousand
hertz,
and
they
want
to
invert
it
over
all
of
them
and
the
impedance
is
just
it's
just
failing
it
or
it
just
can't.
H
I
have
it
almost
implemented
just
pretty
much
all
on
the
receiver
side,
and
I
just
I
had
added
another
receiver
as
a
complex
resistivity
for
the
resistivity
in
phase
and
since
we're
pretty
much
like.
We
have
all
that
stuff.
All
the
calculations
done
in
the
impedance.
We
just
have
to
use
all
that
information
to
kind
of
construct,
the
resistivity
in
the
phase,
but
yeah
I
was,
I
was
kind
of
stumbling
at
the
eye
joint
part.
H
So
when
we
use
jpec,
I
don't
know,
I
I
think
I've
got
something
star
or
something
going
here,
but
I
was
wondering
if
there's
anyone
I
can
maybe
bounce
the
math
off
or
someone
who
can
or
how
do.
I
test
this
kind
of
thing
to
make
sure
that
it's
okay,
john.
D
H
Yeah
exactly
a
lot
of
people
that
I've
been
like
a
lot
of
the
mt
processors
and
stuff
from
other
yeah
that
we've
just
kind
of
hired
they've
said
that
this
problem
is
usually
solved
by
going
like.
If
you
can't
do
it
with
impedance
the
range
you
can
solve
over
larger
ranges
with
the
resistivity
in
phase.
So
that's
what's
kind
of
leading
me
to
it.
D
But
what's
what's
exactly
the
problem,
like
I'm
kind
of
not
sure
why,
let's
say
having
like
a
big
spectrum
of
impedance
is
a
problem.
H
H
F
And
yeah.
H
Most
of
the
time
that
was
the
solution
like
just
having
the
xx
and
the
yy
that
was
kind
of
tough
by
having
all
the
components
in
there.
So
just
breaking
it
down
to
the
off
seemed
to
be
working
great
on
the
smaller
frequency
ranges
like
the
ones
that
I
ran
before
were
about.
They
were
either
like
one
hertz
to
30
hertz,
or
something
like
that.
So
now
that
I'm
going
into
this
wider
couple
decades
worth
yeah.
H
So
this
one
or
recently
I
I
well
I'll
go
by
frequency
and
I
go
by
what
else
have
I
done?
Oh,
I
guess
it
was
by
component
before,
but
I'm
not
doing
that
too
much
anymore.
So
now
I'm
just
doing
like
a
floor
plus
percentage
of
the
data.
E
E
H
Yeah
but
again
I've
I've
tried
to
do
low
as
well
like
I
I've
pretty
much
thrown
everything
at
it
and
it's
just
not
not
functioning
right
now,
yeah,
but
no
that
totally
makes
sense,
because
that's
pretty
much
what
I
was
seeing
is,
if
I,
you
know,
crank
up
that
that
floor,
then
it's
just
it's
going
to
fit
that
resistance.
D
Yeah,
I'm
I'm
more
than
happy
to
help,
but
I
think
I
actually
kind
of
struggled
at
one
point
because
we
wrote
this
1d
code
and
initial
goal
was
actually
having
both
real
and
imaginary
and
the
amplitude
in
phase
yeah.
I
think
it's
actually
discontinuous
that
phase,
so
I
had
a
hard
time
actually
taking
the
derivative
because
if
you
think
about
there's
a
phase
wrap,
so
it
is
a
discontinuous
function.
D
So
you
need
to
be
careful
when
taking
the
derivative
of
the
complex
like
a
phase
of
the
complex
value,
but
I'm
pretty
sure
there
is
a
solution.
I'm
just
not
100
sure
of
what
you
know.
D
And
joe
you're
right,
because
I
think
it
doesn't
matter
because
if
that's
constant,
when
we're
taking
the
derivative,
it's
it's
gone,
so
I
guess
it
should
be
fine.
But
anyway
I
had
a
actually.
I
got
tried
a
couple
of
things
that
I
didn't
work
as
I
kind
of
didn't.
You
know.
H
B
I
think
the
issues
that
I
had
like
when
it's
I'm
starting
like
trying
to
invert
phase
or
something
if
I
ever
had
to
do
a
I
mean
it's
never
happened,
but
I
had.
If
I
I
was
also
inverting
for
like
angles
and
stuff,
if
you're
inverting
for
angles,
then
it
kind
of
gets
weird
if
you
have
like,
if
it
kind
of
goes
over
the
that
jump
there,
that
cycle
yeah.
B
It's
weird
like
a
spatial
spatial
smoothing
on
that
is
weird,
but
I
don't
think-
and
I
don't
think
there
should
be
too
many
issues
with
taking
the
derivative
of
phase
as
if,
as
far
as
a
data.
D
H
A
I'd
be
super
interested
in
taking
a
look
too
so
feel
free
to
ping
a
few
of
us,
because
I
think
this
is
an
interesting
right.
D
D
D
H
Cool
all
right,
well
I'll,
do
that
here
next
and
yeah
actually
dieter's
giving
me
a
hand
and
hooked
me
up
with
some
other
empty
code
and
yeah.
So
I
can
do
some
comparisons
and
we're.
H
And
yeah
I
still
want
to
get
on
that
tiled
vc
problem,
but
things
just
keep
piling
up
so
yeah.
That's
all
I
got.
E
Sure
I
have
my
fingers
and
many
pies,
so
I
think
we've
kind
of
gone
over
the
the
potential
field
stuff
and
I
think
yeah-
maybe
dom
you-
and
I
can
have
a
conversation
at
some
point
and
finalize
this.
E
The
other
thing
as
I've.
I've
gotten
through
the
doc
strings
on
discretize
utilities,
so
yeah,
maybe
I'll
even
just
like
screen-
share
it
for
fun.
Just
for
a.
C
E
Yeah,
so
I
mean
where
this
is
kind
of
how
it's
rendering
now
and
so
I've
I've
gone
through.
I've
followed
the
numpy
way
of
doing
it
and
for
all
these
utilities,
I've
I've
added
examples
whenever
possible.
E
E
E
Yeah,
so
I've
been
doing
some
of
these
validations
so
so
comparison,
comparing
the
the
ubc
codes
to
the
simpeg
codes
and
I've
done
it
for
the
dcip
octree,
and
this
motivated
I
mean
the
validation,
but
also
I
wanted
to.
I
decided
to
come
up
with
an
interesting
way
to
plot
in
3d
arbitrarily.
E
But
now
you
can
basically
feed
it
in
three
points
that
define
a
plane
and
it
will
just
go
and
find
the
locations
that
are
a
certain
distance
from
the
plane
that
you've
and
defined
and
plotted
up.
So
that
was
actually
just
kind
of
handy.
Instead
of
having
to
take
this
3d
survey,
and
then
you
know,
you
would
need
a
you'd
need
a
line
id
vector
to
basically
decompose
it
into
a
bunch
of
2d
data
sets
and
then
plot
those
pseudosections.
E
G
G
Oh,
I
think
I
might
be
using
cell
centers
yeah.
That's
the
problem,
so
try.
D
D
A
Yeah,
the
cell
centers
we'll
put
the
source
at
the
cell
center,
whereas
the
the
nodal
it
interpolates
right,
soggy.
D
Yes,
and
also
that
cell
center
code
in
tensor,
it
uses
the
mixed
boundary
condition,
but
that's
not
implemented
in
the
octet
code,
so
accuracy
like.
So
what
what's
actually
doing
that
is
it's
basically
putting
a
dirichlet
boundary
condition,
so
you're
going
to
get
pretty
bad
accuracy
on
your
surface.
I
think
so
error
interface.
What
you're
doing
is
saying.
Oh,
it's
dirichlet,
potentially
zero.
That's
the
boundary
condition
imposed
yeah.
E
Yeah,
I
think
I
tried
it
with
nodes,
and
I
was
I
got
something
that
I
was
also
really
not
happy
with.
So
I'm,
okay
yeah
I'll,
continue
to
play
around
with
it
but
yeah.
I
think
ideally,
I'd
want
to
use
the
the
the
nodal
formulation
but
yeah
I
I
do.
I
still
do
feel
that
the
the
way
that
ubc
code
is
getting
your
right
hand,
side
is
still
different
than
what
simpeg
is
doing.
E
B
E
To
topography
yeah,
I
haven't
I'm
not
implementing
any
of
that
kind
of
stuff
on
purpose
and
right
now,
I'm
doing
an
example
that
doesn't
have
topography.
Okay,.
B
Because
I
know
there's
like
the
code
that
dom
had
written
to
shift
the
oh
drake
togo
or
something
yeah-
that's
great
like
it
should
it
does
not
only
it.
It
moves
the
electrodes
to
cell
centers.
B
It
doesn't
it
doesn't
just
like
decrease
the
height
it
like
actually
moves
into
cell
centers,
and
the
other
thing
is
like
yeah.
If,
if,
if
you're
doing
that,
cell-centered
one
it'll
always
put
it
to
the
nearest
cell,
because
there's
not
a
great
way
to
interpolate
cell
centers
on
a
octree
mesh
right
now:
okay,
and
it's
just
as
near
as
neighbor
as
far
as
interpolation,
like
that.
That's
why
it'll
always
end
up
at
the
center
of
a
cell.
E
Yeah,
well,
I'm
not
using
that
functionality
yet,
but
I
guess
these
are
all
the
things
that
we're
figuring
out,
as
as
we
do
these
validations
nice,
but
yeah
I'll,
give
it
a
I'll
give
it
another
shot
with
the
the
nodal
formulation
and
see
if
there's
any
discrepancy
there
but
yeah.
I
also
think
that
it
would.
It
should
be
better.
F
It
would
be
nice
too,
if
you
could
test
dev
to
just
you
know,
run
a
run,
the
cr3
vcip
actually
by
putting
all
the
all
electrons
in
cell
center
and
then
compare
with
the
synthetic
you
know
so
that
we're
we
know
that
the
interpolation
is
exactly
the
same.
We're
not
like
fudging
it
different
way.
You
know
just
so
that
we
can
test
the
raw
accuracy
of
the
year
at
the
solves.
E
D
Can
you
also
do
just
a
half
space
because
we
don't
know
which
was
right
at
this
point?
We
probably
put
more
confidence
on
ubc.
I
guess
but
it'll
be
nice
to
actually
have
some
analytics
and
tests.
E
E
F
And
dev,
if
you
want
to
play
around
with
the
plotting,
I
recommend
the
user
scatter
plot
of
plot
lee
is
actually
much
nicer
than
than
using
the
the
matplotlib.
So
you
can
actually
spin
the
spin
the
view
and
select
the
points
and
see
the
values
in
3d
space.
E
Yeah,
I
didn't
want
to
spend
too
much
time
on
it.
I
just
wanted
a
quick
and
dirty
way
of
being
able
to
plot
up
points
if
you
had
an
arbitrary
3d
survey,
because
I
know
that
yeah
pi
plot
isn't
really
a
3d
plotting
package,
so
yeah
pretty.
E
Yeah,
it
was
just
something
that
was
quick
and
dirty,
but
I
think
it's
pretty
useful.
You
can
just
select
planes
and
and
see
these
things
see
the
pseudo
sections
kind
of
pop
out,
but
is
sorry,
are
you
saying
plotly.
F
Yeah
yeah,
probably
I
can.
I
can
give
you
a
little
snip
up
if
you
wanna
yeah,
sure
might
as
well
the
scatters
yeah.
E
And
I
mean
just
to
wrap
up
for
the
last
thing
that
I
was
working
on.
I
started
working
on
the
the
self
potential
spontaneous
potential
stuff,
so
yeah
soggy.
I
was
able
to
go,
take
a
a
single
point
charge
and
match
that
analytic
solution
for
the
either
using
the
q,
the
the
source
term
or
there's
actually
an
analytic
solution
for
the
current
density
that
results
from
that,
and
so
I
input
that
and
also
get
it.
E
So
those
three
things
match
analytic
and
then
using
jsource
or
q
source,
so
we're
in
business.
I.
E
Okay,
so
yeah
I
mean
probably
try
and
do
the
test
for
the
sensitivities,
all
that
fanciness
and
maybe
try
and
solve
some
some
synthetic
inverse
problem.
E
So
yeah
lots
of
different
things.
That's
pretty
much
it
for
me.
A
Thanks
dev
cool
soggy:
do
you
have
any
quick
updates?
Deb
you're
still
sharing
your
screen
just
up
right?
Oh
I'm.
Sorry.
D
That's
nothing
really,
I
think
I'm
not
working.
I
guess
I
went
to
scg
and
it
was
actually
interesting
to
sit
on
machine
learning,
so
I
had
a
chance
to
listen
to
boss.
Peter's
talk
about
actually
talking
about
what
lds
and
lars
is
doing,
and
so
I
actually
was
motivated.
Okay.
How
like
do
I
understand
what
they
are
actually
doing,
so
I
spent
some
time
on
reading
their
papers
and
it
seems
actually
pretty
cool
the
the
architecture
they
developed.
D
D
Oh
can
we
have
syntax
ml
and
basically
inheriting
sort
of
what
they're
doing
just
a
thought,
but
the
because
actually
the
structure
is
quite
similar
and
I
think
that
we
can
basically
use
our
optimization
package
to
solve
training
problem
anyway,
just
just
a
thought
and
that's
sort
of
where
I'm
moving
on
next
phase
here
in
stanford,
like
data
integration
using
machine
learning,
but
I
think
it
seems
like
they're
going
in
the
right
direction,
because
the
data
is
sparse.
D
What
they're
dealing
with
and
then
you
want
to
learn
deep,
so
their
architecture
seems
actually
nice
and
kind
of
like
imitates
the
physics.
So
anyway,
just
like
a
different
direction,
and
I
I
thought,
like
in
general,
it's
kind
of
interesting.
A
One
good
question
on
that
would
folks
be
interested
in
actually
starting
just
like
a
simpeg
ml
slack
channel,
because
it
would
be
interesting,
I
mean
I
would
enjoy
chatting
further
on
some
of
these
papers
and
things
like
that
could
just
give
us
an
avenue
to
discuss
some
of
those
things
cool.
I
see
a
thumbs
up
from
tebow.
Do
you
want
to
start
them.
D
And
also
I
was
going
to
have
a
chat
with
like
a
professor
in
our
house,
so
he's
a
geoscience
like
a
geo
kind
of
geostatistician
and
he's
kind
of
developing
a
lot
of
geostatistical
algorithm
and
doing
a
monte
carlo
inversion,
but
he
basically
needs
some
sort
of
geophysical
simulator,
so
yeah
it
would
be
an
interesting
collaboration,
basically
feeding
him,
here's
a
simulator
that
can
use
and
do
whatever
you
want
your
simulation
and
get
some
output
of
it.
I
think
there's
like
that
kind
of
just
providing
simulation
in
these
days.
I
Yeah,
we'll
just
be
brief,
so
I've
I've
started
over
doing
the
merge
between
pgi
and
simulation,
and
I'm
happy
that
finally
figure
out
a
way
to
do
it
efficiently,
and
I
got
like
the
first.
My
first
example
running
yesterday.
I
So
that's
the
first
step,
so
I'm
gonna
be
working
on
that,
but
then
yeah
after
that,
I
guess
it's
like
I'm.
Just
gonna
try
the
other
example.
But
after
that
I
guess
there
is.
There
will
be
some.
I
I
will
create
a
pull
request,
but
there
will
be
some
work
on
for
me
on
documenting
my
code
and
maybe
after
that
also
like
make
it
more
clear,
maybe
more.
If
there
are
some
ways
like
I'm
sure,
can
you
make
more
efficient
and
also
I'm
pretty
sure
that
the
name
convention
is
not
right
at
the
moment.
So
but
at
least
at
least
I
got
the
first
example
working.
So
that
was
like
that's
the
first
step.
I
A
That's
excellent
I'll
be
quick
here
because
I
know
we're
we're
running
a
bit
over
but
yeah
so
sdg.
Well,
I
really.
I
really
enjoyed
this
session
that
we
had
on
sim
peg,
so
a
huge
shout
out
to
dom
and
john
for
awesome,
presentations
and
organizing
us,
especially
tom
thanks
and
we're
also
getting
up
super
early
starting
at
6
a.m.
In
vancouver.
A
Yeah,
so
for
those
of
you
who
weren't
able
to
tune
in
it
was
a
really
I
really
enjoyed
it
and
so
dom
and
john
feel
free
to
jump
in,
but
it
was
really
cool
to
see
like
such
a
spectrum.
I
think
it
this
session
really
brought
together,
like
everything
from
sort
of
research.
A
Dequan
gave
a
tutorial
on
using
the
3d
time
domain
code,
which
was
awesome.
It
was
a
great
tutorial
and
really
cool
to
see
that
you
know,
we've
put
enough
documentation
out
there
and
been
supportive
enough
that
you
know
somebody
else
who
hasn't
been
deeply
involved
in
developing
the
code
felt
confident
to
give
the
tutorial.
It
was
a
great
tutorial
and
then.
F
Sorry,
lindsey,
and
also
he
gave
us,
we
should
definitely
follow
up
right.
He
gave
us
a
snippet
how
to
run
the
code
on
hpc
yeah.
I
think
that's
a
great
example
that
we
should.
We
should
just
connect
to
absolutely
yeah.
A
I
will
make
a
note
of
that
yeah,
which
was
super
cool
and
then
both
seeing
john
and
dom's
talks
connecting
much
more
to
sort
of
industry.
User-Facing
ways
to
interact
with
simpeg,
so
john
was
showing
the
ds.
What's
it's
caravel
is
the
system.
A
Which
is
super
cool
to
see
simpek
like
wrapped
in
in
a
gui
and
you're
launching
huge
jobs
on
your
servers,
which
is
it's
awesome
and
I
appreciated
the
overview
of
solvers
and
all
the
work
you've
been
doing
digging
into
that
I
mean,
I
think,
that's
a
huge
benefit
to
all
of
us,
and
so
knowing
you're,
making
some
headway.
There
is
super
cool.
A
And
then
for
folks
who
haven't
seen
the
geo
app
that
dom's
been
working
on,
I
think
that's
some
of
the
coolest
widget
work
that
I
have
seen
and
like
piecing
together,
a
like
gui
interface
in
a
notebook
for
inversions.
So
if
you
are
looking
for
some
inspiration
on
really
cool
apps,
you
should
absolutely
check
that
out.
A
The
other
thing
is
geo,
app
yeah,
it's
on
the
mirror,
geosciences
github.
We
can
have
don,
maybe
drop
a
link
into
the
into
the
slack
yeah.
It's
super
cool
and
I
also
really
enjoyed
so.
A
There
were
a
number
of
folks
who
came
in
who
are
using
simpeg
and
so
in
particular,
jaja's
group
is
actually
using
simple,
quite
a
bit
as
well
as
declines
group,
and
so
I
suggested
to
jaja,
and
we
should
do
the
same
for
dequan,
maybe
in
a
few
weeks
time
once
we're
sort
of
through
you
know
catching
up.
A
After
being,
you
know
at
scg
last
week,
just
sitting
down
with
them
and
getting
a
better
sense
of
what
they're
up
to
how
we
can
potentially
be
a
bit
more
supportive,
because,
like
I
was,
I
was
amazed
to
see
some
of
the
stuff
that
is
being
done
with
simpeg,
but
you
know
they
only
reached
out
sort
of
minimally
for
help,
and
so,
if
there's
stuff,
we
can
be
doing
to
bring
in
their
work
and
and
be
supporting
them
more.
That
would
be.
That
would
be
cool.
I'm.
F
A
That's
cool
that
also
brought
up.
I
think
the
important
point
of
we
need
to
get
a
page
on
the
documentation
about
publications
using
simpeg
cool
examples
with
simpeg
that
may
or
may
not
have
a
paper
like
geo
apps.
We
need
to
get
that
up
and
running
because
there's
like
some
really
cool
stuff
happening,
that
if
we
just
had
a
space
where
people
can
drop
a
link,
we
would
know
about
it
so
yeah
we
can
try.
E
And
give
that
a
go,
I
mean
in
some
ways
I
almost
thought
that
was
a
bit
of
a
purpose
of
the
gallery
we
have
right
is
that
we
we
wanted
to
get
people
showing
us
some
cool
stuff
that
we
did
and
then
putting
it
in
the
gallery
and
especially
published.
We
had
us,
we
had
that
space
for
published
works
using
simpeg.
A
I
think
it's
a
great
place
to
put
things,
but
I
think
we
also
want
something:
that's
a
lower
bar
to
entry,
where
somebody
can
basically
just
copy
paste
a
doi
and
say
I
wrote
a
paper,
it
used
simpeg
and
be
done.
It's
because
I
mean
it
would
be
great
if
they
then
want
to
go
the
next
step
of
you
know,
making
an
example,
but
I
think
at
a
minimum
we
just
want
a
place
where
we
can
catch
what
people
are
up
to
that
are
using
zenpeg.
A
Cool
okay,
so
that's
the
high
levels
for
those
things
I
don't
know
dom
or
john.
Was
there
anything
else
you
wanted
to
bring
up
from
that
event,.
F
D
Were
there
like
any
kind
of
discussion
on
road
ahead?
What's
the
future
like
you
know
where
something
goes,
and
that
kind
of
questions
or
discussion.
F
There
is
mostly
with
the
judges
group
right,
I
think,
they're
still
wondering
how
to
how
to
contribute.
You
know
they're
they're,
happy
users
and
they're
tweaking
the
code,
but
they're
not
feeding
anything
anything
back
and
I
think
there's
a
little
bit
of
a
I
don't
know
nervousness
and
in
showing
and
showing
their
code
or
something.
But
I
think
the
role
ahead
for,
in
my
mind,
is,
is
to
encourage
more
collaboration.
H
H
F
Yeah,
because
both
joe
and
what's
his
name
at
laurentian.
F
A
And
with
that-
and
this
is
a
longer
discussion-
so
I'll
just
throw
the
idea
out
and
then
we
can
table
that,
but
I
think
also
sort
of
with
the
publications
sort
of
the
conversations
around
governance.
A
It
would
be
great,
I
think,
to
actually
also
have
sort
of
like
industry
partners
listed
somewhere
publicly,
so
that
it's
kind
of
clear
that,
like
the
mirror
involvement,
the
ds
involvement
are
adding
value
and
it's
it's
a
productive
partnership,
and
so
just
throwing
that
out
there
and
we
can
maybe
have
a
chat
of
what
what
you
folks
think
would
be
sort
of
a
productive,
beneficial.
Both
ways
way
to
highlight
your
contributions
and
involvement
in
senpai.
A
The
last
thing
I'll
say
is
I'm
giving
a
presentation
tomorrow
for
the
sage
group
and
so
sage.
They
mostly
it's
mostly
like
global
earth-
whole
earth,
folks
in
geophysics
with
this
group.
But
there
is
a
fair
bit
of
interest,
I
believe
in
mt,
and
so
I'm
going
to
try
and
have
a
slide
or
two
on
that.
So
I
might
ping
later
today,
if
folks
are
willing
to
just
take
a
quick
look
and
maybe
add
a
few
bullet
points
and
things
like
that.
It's
a
short
talk.
A
It's
like
a
15
minute
talk
plus
conversations,
but
I
want
to
give
sort
of
a
bit
of
a
perspective
on
what
we're
up
to,
but
then
also
where
we're
going
that
might
connect
with
this
community,
so
I'll
ping
with
some
slides
later
later
today,
cool
well,
I
know
we're
a
bit
over.
So
thanks
for
sticking
around.
Does
anyone
else
have
last-minute
burning
items.