►
From YouTube: SimPEG meeting Feb 19
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).
C
D
C
A
little
bit
to
CSM
guys
who
worked
on
the
IP
so
yeah,
so
yeah
like
I,
implemented
the
complex
conductivity
model
in
the
coaster
domain
code
anyway,
so
it
he
he
ran,
ran
it
and
that
matches
with
analytics.
So
I
guess
he's
good
to
go.
He
said,
I
think
Andre
Andre
is
underage
students,
I
guess
so
anyway,.
C
That
was
good
and
well
not
like
some
of
the
Stanford
crews
are
working
on
time-lapse
inversion.
So
we
got
that
to
airborne
surveys
flown
in
two
years:
separation,
particularly
for
the
sea,
otter
intrusion,
so
yeah
we're
yeah
we're
ready
to
run
the
time-lapse
inversion,
I
guess
so.
It's
gonna,
it's
kind
of
run.
A
C
So
I
think
they
got
both
dcpip
and
data
and
seems
like
both
of
them
are
showing
the
IP
effects
so
yeah.
My
portion
is
like
revisiting
what
John
did
and
invert
that
EC
and
IP
and
see
if
there's
a
correlation
between
the
other
one
I
guess
so
yeah.
It's
like
try
to
manage
things,
but
my
plates
getting
getting
full.
B
If
I
can
add
a
point
on
the
urban
IP
I,
don't
know,
you've
done
already
made
that
one.
So
like
late
January
there
was
the
roundup
confirm
scene
in
Vancouver
and
we
went
to
see
the
talk
from
Andrea's
visually.
It
was
really
cool,
yeah,
yeah
and
like
for
me
the
like
the
really
the
message
we
had
like
it
was.
He
had
a
lot
of
message,
but
one
that
really
strike
me
was
that
his
personal
estimation
was
that
at
least
50
percent
of
urban
TM
soundings
were
influenced
by
IP.
B
C
E
Yeah,
quite
a
bit
of
headway
on
the
Mt
I
saw
what
you
guys
were
doing
with
storing
the
factorizations
fields
on
the
frequency
domain
side.
I
think
so.
I've
applied
that
and
then
danced
the
field
as
well
for
the
MT,
when
I've
got
Jim
back
in
JT,
Beck
I
was
playing
around
with
that
to
get
it
parallelized
and
using
the
test.
The
test
models
I
sped
up
the
Jamie
back
from
90
seconds
to
about
45
seconds,
so
JT
Beck
is
a
little
slow
and
I.
Don't
know,
I!
E
Think
it's
because
when
I
run
the
fields
I
JT
Beck
needs
to
store
a
in
or
a
transpose
inverse,
so
I'm
storing
both
of
them.
So
they
they
cost
a
lot
quicker.
What
they
cost
quite
a
bit
and
I
was
thinking
about.
Maybe
writing
them.
Twos
are
but
I,
don't
know
by
the
time
you're
reading
and
writing
them
over
again.
C
E
D
E
A
D
E
D
E
D
E
D
D
D
So
to
take
it,
we
could
take
advantage
of
things
like
so
it
with
Pardy.
So
you
can
add,
you
can
do
a
step
that
just
redoes
the
numerical
side
of
it
and
miracle
factorization
like
doesn't
have
to
recompute
this,
the
symbolic
factorization
every
time,
which
was
something
we
should
be
taking
into
account
like
most
of
these
times
right,
because
the
matrices
are
not
hard
depending
on
the
mesh.
We're
not
really
changing
the
mesh
right,
yeah.
A
D
We
should
probably
fight
an
interface
to
the
PI
Matt
solver,
that
does
that
kind
of
thing,
and
that
would
save
a
little
bit
of
time.
I
mean
the
numerical
factorization
still
does
take
a
decent
amount
of
time,
but
it
cuts
a
little
bit
of
one
step
out.
At
least
it
gets
repeated
every
single
time.
Okay,
just
just
thoughts,
I
mean
these
are
things
I
think
about
when
I
saw
people
doing
that
and
I
just
kind
of
yeah.
E
A
D
You
start
running
to
the
thing,
so
it's
like
okay,
well,
I
produce
the
solver.
Okay,
that
one
works
on
a
single
node
right,
it
doesn't
pretty
so
isn't,
doesn't
solve
on
MPI
systems.
You
can't
take
advantage
of
MPI
stuff,
so
you're
always
gonna
be
limited.
To
keep
my
memory.
You
have
on
one
node
when
you're
using
perdy,
so.
D
We're
modeling
project,
so
it's
been
using
MTE
IP
thing,
IQ,
modeling,
just
kind
of
going
through
and
double
checking
that
I
understand
how
to
use
everything
and
making
sure
I
know
her
when
I
correctly
and
make
sure
everything's
going
on
and
the
optimizing
finding
a
little
worse
to
optimize
and
attention
to
certain
parts
like
what
I'm
talking
about
like
the
emails
good,
really,
oh,
it's
actually
can
get
to
be
a
pretty
big
system.
The
size
of
the
model
domain.
D
We're
working
on
is
about
25
clouded
by
30
kilometers
and
just
breaking
that
up
into
something,
and
the
model
is
pretty
variable
across
the
whole
domain.
I'm
just
breaking
that
up
into
things
that
are
like
even
like
a
200-meter
discretization,
which
there's
certainly
things
that
are
that's
getting
a
lot
about
big
for
the
size
of
the
anomalies
that
are
in
this
model,
even
that
it's
like
okay,
it's
100
meters,
120,.
A
D
D
A
Rick,
we
don't
need
like
so
for
DC
or
IP.
We
don't
need
the
connectivity
between
between
the
two
locations
between
your
two
electrode
locations
so
effectively
like
the
cell
centered,
with
the
cell
centre
discretization
we're
just
treating
treating
resources
like
a
charge
at
the
cell
center
with
the
notable
we
then
just
like
average
to
that
yeah.
D
C
It'll
cancel
out
Joe,
so
if
you
define
the
current
source
as
a
current,
which
is
a
vector,
then
you
take
the
divergence,
then
you're
left
with,
like
plus
or
minus,
or
a
at
the
end
point.
So
in
between
it's
zero.
So
you
could
like
you.
Could
you
could
actually
use
the
line
current
source
and
what
we're
using
in
the
e/m
code
by
just
taking
the
divergence
you
can
actually
effectively
get
that
at
the
point
source
as
well.
What?
But
that
not
not
not
not
needed,
but
no.
A
C
D
D
E
C
A
C
C
Yeah
yeah
yeah
yeah,
or
you
can't
like
that
you
could
solve
that
this
equation
for
the
right
hand
side
it's
a
boy,
oh,
that
oh,
this
is
now
a
Linux
solution.
You
can
solve
for
the
right
hand
side
before
yeah,
so
I
mean
there
are
a
couple
of
tricks.
I
don't
know,
I
mean,
like
depends
like.
If
you
really
need
a
really
good
accuracy,
and
maybe
that's
typical
or
5%
DC
data,
it's
pretty
fine.
With
that
all
we
have
now
that's
my
conclusion
was.
D
So,
okay,
for
that
stuff,
I've
been
going
through
like
something
all
the
votes
of
the
pull
requests
and
just
trying
to
like
see
where
things
are
at
and
it
feels
like
there's
some
basically
dead,
pull
requests
that
are
either
just
kind
of
like
okay.
Well,
somebody
did
this
in
another
one.
The
same
way.
Somebody
did
like
open
this
pull
request
and
then
did
nothing
with
it.
D
Yeah
oh
I
mean
I've
been
work,
trying
to
been
working
through
the
other
stuff.
I
mean
it's
little
bits
and
pieces
here
and
there.
Basically,
what
I
really
need
to
get
done
is
I
want
to
get
to
the
real
figured
out
as
far
as
the
simulation
stuff,
not
many
people
I
mean
on
that
issue
that
I
posted
on
simple
sorry.
D
A
D
A
A
A
E
Think
with
the
percent
air
thing,
it
might
be
just
easier
just
to
have
a
check
if
someone
puts
it
in
as
a
point
or
as
in
like
a
say,
50
percent
or
something
like
that.
It
kind
of
seems
hard
to
what
wouldn't
you
name
it
I,
don't
know
the
name
is
just
you
can
call
it
a
purse
and
you're
gonna
get
two
different
thoughts
on
it.
Of
course,
I
think
it
just
might
be
easier
just
to
do
a
check
within
the
color
to
see.
A
D
D
A
A
This
yeah,
okay,
yeah
I,
am
liking
relative
error
or
a
noise
ratio.
I'll
be
curious
to
hear
Doug's
thoughts
on
this
I
know
like
with,
with
his
experience,
I'm
sure
this.
Through
some
of
these
things,
you
have
a
strong
gut
reaction
to
any
of
those.
A
A
A
A
A
I
mean
they
sent
me
some
field
data,
so
I
will
be
using
some
of
that,
but
for
now
what
I'm
doing
is
generating
and
using
their
for
modeling
code.
It
suggests
it's.
You
know
it's
an
approximate
physics
model
where
we
treat
an
ordinance
object
with
three
different
axes
of
basically
like
they're,
three
different,
three
orthogonal
dipoles,
each
with
a
characteristic
decay.
A
C
A
B
D
A
D
D
A
Some
of
the
data
that
you're
looking
at
for
just
a
single
transmitter,
so
this
is
for
a
Twitter
over
forget,
that's
basically
centered
one
meter,
ish
and
so
with
what
we
have
is
basically
a
hundred
and
sixty-five
input
channels,
each
location,
approximately
each
location
within
a
given
like
pretty
small,
because
it
moves
along
like
we're
not
sitting
and
firing
a
transmitter.
Basically,
you
get
like
five
transmitters
times
eleven
receivers
times.
A
C
A
That
would
be
a
next
step
right
now.
It's
basically
just
like.
If
we
have
an
image,
can
we
just
identify
it
like
this?
This
anomaly
is
due
to
this
type
of
object
and
then
from
there.
If
you
already
know
what
type
of
object
it
is,
then
that
can
basically
be
fixed
in
the
inversion
if
you
want
or
like
you're,
starting
from
basically
the
right
you're
starting
from
a
place.
That's
much
closer.
A
A
A
E
E
B
I
was
gonna,
say:
I
played
a
bit
with
both
and
I
like
fight
or
two
better,
because
also
tensorflow
iPhone
was
like
very
like
it
does
well
what
it's
supposed
to
do,
but
if
you
want
to
implement
something
new
or
like
yeah
like
just
play
around
with,
it,
is
much
harder
to
modify
the
wall.
Part
is
much
more
flexible.
A
B
A
E
C
E
A
little
small,
but
so
we
got
things
like
two
lyrics
happening:
contact
contacts,
just
bad
data,
that's
disconnected,
say
a
current
electrode,
that's
drying
up
and
it's
burying
the
VP's
yeah.
But
this
using
the
whole
time
series
as
that.
What
am
I
implementing
here
I
just
have
two
layers,
just
like
softmax
and
then
just.
E
Yeah
I
was
just
playing
around
an
example,
so
just
I
think
this
one
only
has
like
a
hundred
one
hundred
and
sixty-three
samples
and
I
get
about
an
accuracy
of
73
percent
by
the
end,
and
it
does
actually
do
pretty
good
like
right
here.
I
give
it
that
time
series
and
it
sees
that
oh
yeah,
that's
a
timing
issue.
A
E
D
E
E
Painstaking,
but
it's
starting
to
pay
off
a
bit
so
I'm
stoked
on
that.
The
only
thing
I
find
is
the
also
using
different
generator
set
different
transmitters
like
the
G
Dee
Dee's
got
like
a
spike,
usually
when
it
tends
to
turn
on
so
when
I'm
trying
to
train
it
at
the
spot.
Iam
spikes,
it's
sometimes
mistakes,
a
good
time
series
because
of
the
transmitter
where.
D
E
E
C
C
C
E
C
E
C
A
E
C
What
I
usually
do
is
use
the
DC
data
like
it's
like
a
like
a
just
like
a
1
millivolt
like
you,
multiply,
1
minus
3
to
the
DC
data
and
I
used
that
so
effectively
like
it's
like
an
apparent
chargeability.
So
what
you're
looking
is
an
apparent
chargeability.
You
convert
that
into
the
voltage
data,
but
what
you
care
is
apparent,
chargeability,
so
I
don't
really
care,
like
small
apparent
chargeability,
so
anyway,
yeah
just
curious.
D
C
C
C
D
E
A
D
D
C
D
C
D
Work
for
tree
maybe
work
for
tree
much.
It's
at
least
find
the
locations
of
them.
We
just
need
to
go
through
and
look
at
some
of
the
spots
where
we're
making
assumptions
that
it
like,
even
if
we
don't
realize
it,
but
there
are
a
few
spots
in
it,
we're
just
making
assumption
and
that
it's
like.
Oh,
it
is
gonna,
it's
gonna,
be
a
you
know
some
sort
of
structured
mesh
thing
just
little
things
I
think
yeah
right
yeah.
We
should
definitely
do
it
in
a
way.
That's
a
little
bit
more.
Oh.
A
C
Submitted
that
well
paper
thanks
Joe,
so
I
used
the
octree
code
and
though
I
think
was
dumb,
wrote
another
rapper
and
both
like
Joe.
That
was
very
helpful,
so,
like
I
discretized
that
power
lines
and
solve
like
M
problem
and
yeah,
that
was
actually
we
use
that
for
a
surf
like
a.m.
survey
design
so
anyway,
says
yeah.
D
B
B
B
No
I
never
done
that
before.
So
probably
we
can
like
site
chatons
like
choose
to
I
won
out
like
the
few
details
that
might
be,
but
yeah
otherwise
too,
like
that
means
there
is
a
repose
to
you
now
or
so
is
in
the
research
like
I
didn't
set
up
a
binder,
because
the
example
were
celibate.
Okay,
Zeus
example
date
back
before
the
tree
mesh,
so
I
was
teasing
and
peeked
and
Somesh,
but
and
I
just
kept
it.
But
at
least
the
scripts
and
notebooks
are
online,
just
not
on
binders.