►
From YouTube: SimPEG Meeting December 8th 2021
Description
SimPEG Weekly Meeting December 8th 2021
A
It's
better
for
it'll,
be
better
for
repeated
searches
on
meshes
like
using
the
closest
point
index,
and
also
it
just
there's
a
small
bug
originally
with
the
first
implementation
that
we
pulled
in
that
didn't
work
for
1d,
and
then
I
fixed
that
as
well,
but
that
new
patch
releases
all
out
and
then
I
finally,
we
were
able
to
pull
in
the
cross
gradient
pull
request
from
over
a
year
ago.
So
it's
all
in
it's
pulled
in
the
sim
peg,
now
tested
it
works
covered.
B
A
A
A
A
A
D
A
C
E
B
F
I
was
actually
revisiting
that
year,
one
decade
and
now
I'm
actually
motivated
to
rework,
because
I
need
to
use
this
so
yeah,
I'm
making
actually
quite
a
bit
of
progress.
So
yeah
joe
did
a
lot
of
work,
but
the
what
he
worked
is
not
really
transferred
to
the
stitching
problem,
so
I'm
kind
of
working
on
transferring
his
optimization
into
the
stitching
problem
so
like
which
is
sort
of
the
major
use.
On
my
end,
I
don't
really
use
much
that
individual
1d
problem,
but
basically,
like
many
many
1d
soundings
to
invert
yeah.
F
I
think
I'm
getting
pretty
close
to
finish
up,
so
I
made
a
pull
request
and
I
think
most
of
the
tests
are
pressing
except
some
of
the
tests
are
not
related
to
me,
which
I'm
not
sure
I
guess
like.
Where
are
you
on
joe?
Are
you
kind
of
like?
Have
you
fixed
most
of
the
tests,
or
is
that
something
that
we
need
to
take
a
look.
F
F
Even
though
you
got
many
receivers,
if
the
height
is
fixed,
that
the
coefficient
doesn't
really
change,
so
you
can
actually
in
a
way
you
can
actually
optimize
that
process.
Also,
if
you
at
low
moment
and
high
moment,
for
instance,
if
you,
although
they
are
two
different
data,
but
all
you
need
to
compute
is
just
a
step
response,
so
you
can
actually
share
quite
a
bit
of
part.
So
I
think
what
joe
did
basically
kind
of
figuring
out
all
of
those
shared
part
and
then
just
computing
once
and
yeah.
F
That
actually
makes
quite
a
bit
of
progress
in
the
code,
but
that's
that
means
that
kind
of
that
idea
needs
to
be
transferred
to
this
digit
code,
which
actually
calls
each
individual
1d
problem.
So
what
I
was
working
on,
basically
sending
that
that
the
stored
coefficients
to
the
main
whenever
I
like,
to
just
store
on
the
whatever
parent
node
and
whenever
I
compute
on
the
child,
just
transfer
that
like
so
I
just
kind
of
back
and
forth
that
communication.
That's
what
I
was
working
on.
F
Right
and
also
joe
has
written
the
gradient,
the
sensitivity
for
mu
and
the
height
of
each
cell,
each
layer
and
right
so
that
actually
I
haven't
implemented
yet,
but
that
all
the
mechanics
are
sort
of
ready,
so
basically
kind
of
implementing
that
into
stitching
code.
So
dumb.
I
think
this
is
potentially
might
be
an
interest
of
you
because
you
may
want
to
invert
it
with
mu
and
conductivity,
and
that
actually
implementation
now
is
like
a
relatively
straightforward.
F
F
What
joe
did
actually
that
kind
of
make
that
as
a
kind
of
linear
operation,
so
you
can
actually
format
a
matrix
and
then
once
you've
got
the
step
response
you
can
just
compute
multiply
that
and
what's
kind
of
cool
it
can't
the
same
idea
can
be
applied
for
the
3d
code.
So
because
all
this
idea
of
setting
up
the
waveform
and
then
oh,
what
is
the
time?
Discretization
that's
kind
of
cumbersome
and
also
it's
really
depend
upon
like
whether
you're
an
expertise
of
this.
F
Even
if
you
are
expertise,
that's
a
pretty
challenging
process
to
set
the
right
time
in
a
role,
but
I
think
it
kind
of
makes
the
problem
much
easier.
So
what
you
need
to
do
you
just
make
sure
have
a
time
step
that
can
handle
the
step
response,
which
is
somewhat
easier.
I
think
compared
to
with
the
waveform,
so
so
that's
that's
what
he
did
and
but
what's
actually
nice
once
you
form
that
if
you
got
all
of
sources
sharing
the
same
waveform,
you
just
need
to
form.
F
Once
now
you
can
just
multiply
that
for
all
those
responses
that
can
be
quite
fast,
because
the
and
also
in
3d,
you
can
actually
remove
a
lot
of
time
channels,
time
steps
that
are
needed
to
calculate
the
complicated
waveforms
so
yeah,
it's
kind
of
that's.
Where
am
I
and
kind
of
speaking
of
that?
The
other
thing
that
I
need
to
need
for
now
for
the
like.
The
project
that
I'm
working
on
is
the
the
tiled
3d
aem
code.
F
I
wrote
it
before,
but
it's
kind
of
like
a
outdated,
so
I'm
going
to
rework
on
that
and
the
idea
is
similar
like,
as
I
said,
I'm
just
going
to
evaluate
the
step
response
for
the
local
problem
and
then
one
once
I
got
the
step
response
for
all
of
the
sources.
Then
I'm
just
going
to
multiply
this
projection
matrix
once
and
the
parent
node.
F
So
in
such
a
way
we
can
kind
of
like
quickly
run
the
3d
m
simulation,
so
that's
sort
of
what
I'm
you
know,
I'm
kind
of
working
on
now
and
yeah,
because
I
I
this
point.
I
need
a
quick,
fast
3d
simulation
code,
not
the
inversion
code
yet,
but
at
the
point
when
I
need,
I
probably
revisit
that
so
anyway.
That's
that!
That's
where
I
went.
B
F
I
think
I
think
the
what
I
have
found
the
best
is
actually
changing
the
time
step
each
time
because,
like
if
you
kind
of
logarithmically
increasing
the
time
step
and
slowly,
you'll
you'll
need,
like
a
multiple
time,
steps,
that's
sort
of
the
best,
but
I
think
for
most
of
the
am
case
we
can
even
fix
the
time
step.
That's
well
designed
for
the
aem,
for
instance,
and
I
just
use
it
or
most
of
the
am
case.
F
I
that's
what
I
was
doing
and
that
this
kind
of
works
in
most
cases,
but
that
most
cases
I'm
not
sure
it's
enough
sampling.
But
that's
still
a
question.
I
I
don't
have
a
good
answer,
because
it
obviously
depends
upon
the
conductivity
structures
and
but
I
think,
if
you,
if
you
kind
of,
have
put
enough
sort
of
time
steps
then
so
I
think
that
my
idea
is
like
you
make
this
problem
very
small,
then
number
of
time
step
doesn't
really
matter
too
much
like
we
can
just
have
many
time
steps
as
needed.
F
D
D
D
F
D
F
F
F
So
if
you
guys
can
come,
it'll
be
an
it
gets
an
open
source
session
and
leo
the
sort
of
like
the
major
developer
of
veriando
is
gonna
talk
about
his
sort
of
practices
and
also
the
other
guy
who's
going
to
talk
is
from
what's
called
pythia
project.
It's
an
educational
project
focusing
on
like
a
more
climate
science
but
the
thing,
but
on
the
education
side.
F
So
he's
going
to
talk
about
like
how
kind
of
he's
learning
from
the
pangea
project
and
how
to
deal
with
the
big
data
and
how
to
transfer
the
knowledge
into
kind
of
educational
resources,
and
there
are
quite
a
few
other
interesting
posters
like
from
usgs.
They
are
developing
airborne
em
database
how
to
like
actually
how
to
transfer
the
data
that
they
have
into
some
sort
of
database,
so
that
kind
of
exporting
that
data
into
or
sharing
the
data
into
the
other
people
or
other
scientists
can
be
easier.