►
From YouTube: SimPEG meeting Aug 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
A
Can
you
see
that
Joe
very
vague,
okay,
so
I
posted
this
issue
just
a
couple
days
ago,
and
one
of
the
things
that
I
was
noticing
is
that
okay,
so
the
way
that
we
construct
parametric
maps,
you
need
to
have
some
way
to
approximate
a
step
function,
and
so
what
we've
done
is
we're
using
arctan
functions
right.
That's
I,.
A
Can
use
our
tenant
so
let's
go
with
that
and
so
to
create
different
structures
and
things
like
that,
initially
what
I
had
implemented
it
is
that
we
just
multiply
these
things
together
right.
So
if
you
multiply
a
block,
we
have,
you
know
two
going
in
the
horizontal
directions
and
then
two
in
the
other
directions,
but
the
thing
that
I
was
encountering
is
that,
like
you,
can
see
that
that
actually
looks
pretty
funky.
A
If
you
soften
the
slopes
right
and
why
you
want
to
soften
the
slopes,
is
it
helps
very
yes,
yeah
helps
your
gradients,
and
so
this
obviously
is
like
just
not
really
physical
and
not
what
we
want
to
be
inverting
for
it.
So
there's
a
couple
options:
I
mean
one
like
an
ellipsoid
is
the
easy
thing
to
parameterize,
but
it
is
nice
to
also
have
a
plate,
and
so
what
I've
been
looking
at
is
actually
changing.
A
Instead
of
multiplying
these
things
together,
defining
it
through,
like
you
think,
of
the
the
equation
for
a
sphere,
and
so
it's
1
minus
and
then
you
have
all
day
to
the
R
squared
terms
right
and
so
a
square
is
actually
just
an
approximation
of
the
Infinity
norms.
So
you
replace
the
two
norms
with
Infinity
norms
and
you
get
a
square,
and
so
what
I've
been
working
on
is
sorry
I'm,
so
lost.
I
love
this
up.
The.
A
A
Is
we
have
now?
Let
me
just
start
with
this
okay.
So
if
we
have
a
large
P
and
a
larger
slope,
we
get
square
back
or
rectangle
back.
What's
kind
of
interesting
is
the
P
the
norm
value?
Actually,
controls
like
held
around
your
corners
are
so
that's
often
the
corners
a
bit,
and
if
we
go
to
the
one
you
get
an
ellipsoid,
it's
just
kind
of
fun.
A
A
It's
still
thinking
r10,
so
what
we're
doing
is
now
we're
just
not
multiplying
them.
What
we
are
doing
is
so
here
is
our
expression
for
a
block
when
this
guy
is
below
zero,
we're
outside
the
block
and
when
it's
above
zero
we're
inside
the,
and
so
all
we
do
is
pop
that
inside
of
this
guy,
which
is
our
approximation
of
and
our
tan
function
and
so
anywhere
that
it's
outside
we're
going
to
go
to
the
value
the
background
and
anywhere
that's
inside.
A
We
go
to
the
value
of
of
the
that
makes
sense,
and
so
the
slope
there
is
this
game
and
there's
two
ways:
I've
tried
to
set
it
up
for
the
user
to
actually
interact
with
this
look
one
is
you
can
provide
the
attack
slope,
the
other
one
is
you
can
provide
just
a
scale
factor,
and
then
it
looks
it
scales
it
by
the
smallest
cell
size
in
your
mesh,
because
what
it
might
show
this
in
his
thesis
and
I've
been
playing
around
and
I'm.
Seeing
the
same
thing
is.
A
A
So
then
we
can
go
ahead
and
derivative
test
that
and
make
sure
that
that's
all
all
working
in
a
simple
case.
So
even
if
I
like
crank
this
up,
the
derivative
is
still
I
mean
you
can
see
that
it's
it's
less
stable
right
and
so
that
sort
of
does
indicate
that
it's
it's
gonna
be
harder
to
search
so
I
actually
have
a
DC
example
where
we
can
look
at
that,
because
it's
on
its
own,
the
parametric
like
this,
isn't
too
crazy.
A
A
So
what
this
example
is
is
we're
doing
a
cylindrical
asymmetric
simulation.
We've
got
sorry.
Let
me
just
see
if
I
have
a
simple
example
of
the
model
yeah,
so
we've
got.
This
is
a
casing
style
example.
So
cylindrically
symmetric
we've
got
a
casing
in
the
middle
and
we've
got
a
conductive
target
at
depth
and
it
extends
to
a
radius
of
50
meters.
So,
if
I
make
this
simple,
you
can
see.
So
that's
the
connected
target
that
was
can.
C
A
So
this
is
a
bit
of
a
subtlety
with
the
cylindrical
symmetric
mesh,
so
this
is
a
symmetric
mesh.
Actually
all
we
have
is
this
right
hand,
side,
Brandon,
Rose,
yeah,
but
like
all
we
ever
store
is
just
the
right
hand.
Side
it
doesn't
need
we
don't
need
to
store
to
the
other
side
and
so
under
the
hood
I'm
playing
a
trick.
A
I'm
just
asking
it
to
hear
the
data
and
I
just
didn't:
ask
it
to
hear
the
grid.
That's
all
so
then,
here
what
we're
gonna
do
is
I've
got
receivers
on
the
surface
so
from
25
meters
to
a
thousand
meters,
and
then
we
have
sources,
we're
gonna,
do
10
sources
between
900
meters
depth
and
a
thousand
meters
depth.
So
just
like
focused
right
around
for
this
case,
set
up
a
DC
simulation
run.
A
A
We
know
that
there's
an
improvement
that
gave
to
the
imaging.
So
in
this
case
I'm
gonna
push
the
data
pretty
hard,
because
it
is
a
small
target,
we're
just
your
standard,
l2
and
virgin
right
now,
Tekin
off
regularization
and
start
with
the
background.
And
so
it's
a
pretty
quick
conversion.
A
A
B
B
A
Interesting
is
like
what
I
think
would
be
kind
of
cool
is
to
also
allow
us
to
invert
for
the
epsilon
value
in
the
norm,
because
that
actually
gives
you,
the
cows
smeared
be
the
conductivity
is,
and
so
here
I'll
show
you
what
I
mean
by
that.
So
here
we
have
the
P.
We've
got
a
strong
slope,
but
if
we
have
I.
A
So
that
gives
you
have
smeared.
It
is
so
that
could
be
an
interesting
thing
to
think
about
inverting
Ford
not
implemented.
It
might
not
actually
be
a
good
idea,
but.
C
A
C
A
Yeah,
it
sort
of
it
softens
it
softens
how
yeah
yeah
I
mean
it's
similar,
so
that
we
could
sort
of
do
it
would
almost
be
the
same
as
softening
the
slope,
but
not
quite
because,
as
you
see
here,
when
I
softened
or
when
I
increase
epsilon,
we
didn't
actually
change
the
like.
Really
the
balance
since
this
guy.
A
So
one
thing
what
I
found
is
that
okay,
so
let's
jump
back
to
this
and
I'll
come
back
to
your
questions?
Auggie,
okay!
So
we
running
out
to
get
a
smeared
out
log
at
the
right
depth,
so
that's
good
and
so
soggy.
To
answer
your
question:
what
I
found
is
that
the
best
way
to
tune
the
parameters
is
to
actually
construct
your
mapping
and
then
test
it
for
a
model
that
is
about
appropriate
or
like
ballpark,
what
you're
expecting
and
tune
the
parameters
till
you
get
second-order
convergence.
A
So
here,
if
we
start
with
a
moderate
slope
factor
and
moderate,
like
softening
of
the
corners
that
should
give
us
second-order
if
I
cranked
up
the
slope.
So
let's
make
it
sorry
the
P
there
was
one
there's
miss
pH
5!
No!
No!
But
what
was
is
that
this
menorah?
Basically
so
that
that's
going
to
control
how
round
the
corners
are?
It's
still
like
so
here
if
I
show
you
this
guy.
So
if
we
do,
this
is
about
the
same
scale
as
the
target
that
I'm
looking
at
and
the
other
example.
F
A
J
J
E
A
D
A
Yeah,
okay,
but
if
yeah
so
one
thing
I've
found
just
playing
around,
is
that
if
you
make
the
P
so
here
in
this
case,
when
I
mean
P
five
like
it's,
basically
a
rectangle,
if
you
crank
it
up
too
high,
it
does
again
have
trouble
with
derivatives.
So
so
I
get
to
answer
your
question,
so
this
guy
was
second-order
if
we
crank
up
the
slope.
A
B
A
See
that
this
guy's
overflowing
and
so
like
it
is
it's
just
less
stables,
hey
John,
yeah,
okay,
so
this
one
actually
worked
out,
but
it's
still
what
I've
found
is.
If
you
make
these
parameters
too
harsh
it
line,
searches
like
crazy.
So
do
you
want
to
find
something
where
you're
getting
sort
of
stable,
second-order,
whatever
it.
C
A
A
E
A
Yeah,
absolutely
and
so
like
we
can
run
this
inversion
and
right
now.
It's
a
lot
slower
so
like
just
taking
a
single
tip,
is
a
lot
slower
than
taking
an
l2
step.
I
think
some
of
that
has
to
do
potentially
with
just
the
mappings
and
the
mapping.
Derivatives
I
think
there's
probably
room
to
make
those
a
bit
more
efficient.
I
was
a
time.
C
C
C
A
A
So
you
can
see
that
if
you
provide
this
load
factor,
it
just
scales
it
by
the
match
so
that
you're
getting
something
in
the
right
range,
then
all
that
the
parametric
really
has
is
just
a
couple
things
to
Center
you.
So
it's,
oh
sorry,
just
to
deal
with
the
active
indices,
because
right
now,
if
I've
got
a
mesh
and
I'm
inverting
I'm
doing
an
inversion,
so
I
have
ourselves
I,
don't
want
necessarily
the
air
cells
considered
in
the
parametric
map.
So
these
are
just
to
help
you
not
have
to
worry
about
dealing
with
air
cells.
J
J
A
A
Don't
know
if
it
makes
sense
to
provide
it
it
different
cell
by
cell,
because
what
would
happen
is,
is
you
would
like?
So
if
you
think
about
an
octree
and
if
you're
going
from
a
very
fine
region
to
your
coarser
region,
you're,
gonna,
sort
of
spirit
yeah,
you
can
increase
the
length
of
your
yeah
yeah.
That's
something
to
think
about.
I,
don't
entirely
know.
But
for
now.
J
A
A
So
all
that
we
need
to
have
like
all
of
this
guy
has
it's
basically
coordinates
and
then
some
step
function
and
it's
derivative
and
so
what's
kind
of
fun.
Is
that
eventually
we
can
actually
think
about
adding
different
types
of
step
functions
like
if
that
is
something
that
people
want
to
play
with.
I
know
in
my
experiment:
engined:
there
are
no
differences
in
convergence
or
whatever
this
pad
like.
So
that's
something
that
we
can
implement
at
this
level
and
then
you
could
actually
just
have
a
keyword
or
a
property
that
that
says:
I
want.
A
Okay,
so
that's
the
parametric
block,
we've
got
it
for
2d
and
for
3d.
I
have
some
thoughts
about
like
because
right
now
the
order
is
sort
of
arbitrary.
Eventually
I
would
like
to
actually
allow
the
user
to
input
something
like
a
dictionary
or
a
wire
map
so
that
you
can
have
it
in
whatever
order,
but
right
now
you
can't
which
is
okay,
so
so
here's
the
epsilon
value
and
the
p
value.
I
think
this
is
actually
too
high
of
a
default
having
played
around
with
it,
at
least
for
the
casing,
matches
probably
too
high.
A
One
of
the
things
that
I've
been
working
with-
and
this
is
sort
of
looking
forward
for
allowing
users
to
input
dictionaries
or
something
like
that-
is
a
dictionary
that
just
grabs
the
right
parameter
given
a
a
keyword,
because
when
I'm
figuring
out,
for
example,
the
derivative
of
the
block,
I
just
want
to
say
you
know
it's
not
up
the
block
rather
than
saying
this
is
model
parameter,
2
or
4,
or
whatever
so
soggy.
Here's
the
PDF
on
that
piece
and
it's
derivative
and
then
to
get
a
2-d
block.
J
A
So
then
the
full
transform
we
do
the
background
plus
and
then
the
difference
between
the
background
and
or
the
block
and
the
background
and
the
times
that
arctan
function
so
that
we
basically
we
take
an
arctan
function.
We
lift
it
up
to
the
value
of
the
background,
and
then
we
say
that
this
value
up
here
needs
to
be
equal
to
that
of
Y.
J
J
E
B
E
A
Yeah
yeah,
but
in
what
I've
found
like
with
this,
is
that
you
want
to
design
so
before,
like
when
we
were
multiplying
arctan
functions.
Basically,
what
we
were
doing
is
saying:
I
have
an
indicator
of
if
I'm
on
the
left,
the
right
whatever
and
then
we're
just
gonna
multiply
those
indicators
together.
So
in
this
case,
what
you
actually
want
to
do
is
define
one
indicator
that
says
I'm
either
inside
or
outside,
and
then
we
take.
We
smooth
the
edges
of
that.
A
So
here,
if
there's
what's
nice,
you
can
derivative
test
each
piece
independently,
so
that
if
I
see
that
it's
not
second-order,
you
can
go
and
make
sure
that
the
H
vector
that
you're
using
to
test
so
say
I,
see
it's
not
second-order
and
I
suspect
that
this
one's
wrong,
then
the
H
vector
pass,
would
be
1,
0,
0,
0,
0,
0
and
not
just
tests.
Is
this
derivative
correct?
Well
that
makes
sense,
and
then
you
can
go
and
see
the
derivative
the
background
and
go
double-check
that
that
lines.
Okay.
A
C
C
C
A
A
A
C
A
Anything
like
one
of
the
things,
that's
kind
of
fun
that
I
want
to
try
with
this,
too,
is
then
using
this.
Someone
is
using
like
in
Britt,
crisp,
smooth
background
and
then
basically
specify
here
that
I
want
the
background
to
be
0
and
then
we're
just
adding
some
from
parametric
blob
on
top
of
it.
So
that
sort.
J
J
J
B
A
A
B
A
It's
like
the
source
is
in
the
casing,
but
the
source,
actually
because
it's
in
the
casing,
it
smears
out
the
currents,
and
then
we
have
data
on
the
surface.
So
what
I've
found
is,
if,
like
you,
just
really
push
this
parametric
inversion,
it
just
tends
to
crank
the
wreath
or
the
conductivity
of
the
block
up
without
really
moving
the
radius.
Necessarily
it
can
sort
of
go
I
mean
because
we're
not
regularizing
the
conductivity
right,
so
it
can
just
like
start
going
to
the
roof
mm-hmm,
which
is
kind
of
interesting.
So.
J
J
A
J
A
F
A
C
C
So
that's
pretty
good,
okay!
This
is
this.
Is
this
guy
number
61
and
we're
increasing
the
coverage
by
a
fair
amount
and
that's
mostly
because
I
got
rid
of
every
dumped
them
coke
Virginia?
Now
before
we
had
like
a
problem
for
the
NPI
and
a
problem
for
amplitude
they're,
all
separate
and
I
just
collapsed
it
all
into
one
and
are
they're
just
cases.
C
J
C
C
C
C
G
C
It
will
consume
but
an
area.
Basically,
the
example
is
just
it's
a
sparse
and
the
I
in
in
a
tree
mesh
so
kind
of
touching
a
lot
of
things
right
and,
as
usual,
my
tests
are
kind
of
like
all
in
one.
So
it's
just
making
sure
that
that
the
starting
model
in
that
and
the
emergent
model
are
almost
able
to
go,
and
if
they
are,
then
that's
the
best.
Even.
C
Joe,
so
currently,
what
we're
kind
of
like
we're
talking
about
last
time
is
to
be
able
to
have
the
the
playing
function.
It
would
be
nice
if
we
create
you
start
a
PR
on
discretize
to
be
able
to
plot
vectors
on
the
slices
and
I
took
the
I
just
took
it
apart
and
just
slap
it
in
the
example
to
create
the
Natura,
so
the
menstruation
one
second
I'll
find
them
yeah
to
center
it
basically
just
utils
to
send
earlier.