►
From YouTube: SimPEG Meeting July 3th 2018
Description
Seogi Kang presents on using sparse norms in EM inversions
A
A
C
A
C
B
A
B
Ge
doesn't
change.
So
it's
fixed,
that's
kind
of
simple,
but
for
like
a
nominative
inverse
problem
like
a
am
problem.
First,
ec
inversion
or
yen
inversion.
So
now,
you're
for
function
is
little,
not
it's
nonlinear.
So
your
sensitivity
function
now
it's
not
fixed,
but
it's
updating
in
the
inversion
so
like.
If
you
got
10
iteration,
each
iteration
have
different
sensitivity
function,
that's
a
main
difference.
B
B
We
usually
have
a
smooth
linear
sense
for
smallness
and
smoothness,
but
here
it's
kind
of
like
a
lump
sum
of
both
of
them
and
what's
really
important
for
en
problem,
is
actually
setting
a
right
and
rest
because
if
you
think
about
what
we're
seeing
okay,
we
want
to
have
most
of
most
barometer
zero.
That's
what
we're
asking
when
we're
using
sports
inversion
so
for
magnetic
inversion,
it's
kind
of
makes
sense!
You
don't
really
hear
em.
Ref
is
usually
zero
or
four
gravity,
I.
Think
as
well.
B
I
guess,
your
M
ref
is
usually
zero,
but
for
conductivity
inversion
like
this
conversion,
your
M
ref
is
not
usually
zero.
So
you
have
some
background
conductivity
model,
so
you
need
to
set
that
MRAP
really
right
to
kind
of
like
make
really
sparse.
I.
Think
that's
that
that's
a
very
important
point.
So
your
initial
and
Rath
is
very
important.
B
Well,
I
got
if
you
guys
have
any
question
just
let
me
know
like
it,
not
very:
it's
not
very
prepared
talking
casual,
so
here's
an
example,
but
this
is
an
example
that
I
used
for
a
ground
water
problem
in
California.
So
our
background
is
about
our
background
is
not
kind
of
constant.
That
first
layer
is
kind
of
it
could
be
an
aquifer
about
the
kill
meter.
This
is
the
resistivity.
We
got
a
clay
layer
here,
but
in
clay
layer
having
I,
don't
500
meter
and
then
we've
got
another
aquifer.
B
It's
a
little
bit
more
conductive
I.
Think
the
distinction
here
is
water
level.
I!
Think
if
you
have
a
water
level
somewhere
here,
then
you're
gonna
get
more
conductive
aquifer
compared
to
other
parts,
and
then
you
can
have
a
very
conductive
batteries,
for
instance,
and
this
is
a
Skycam
data
like
I'm,
not
sure
anyone
knows
despite
em
yeah,
but
it's
a
dual
moment.
So
you
got
a
two
waveform
and
so
you're
gonna
get
to
decay
curves
as
a
data.
So
if
I
have
one
sounding
at
this
one
be
later
I'm
gonna
get.
B
That
is
two
sounding,
but
the
orange
one
is
low
moment
and
blue
one
is
high
moment.
They're
usually
like
this
is
normalized
data,
but
if
you
actually
plot
the
real
amplitude
high
moment
will
be
much
larger
compared
to,
and
the
idea
here
is
high
moment
seemed
very
much
later
time.
Channel
little
moment
can
much
color.
B
The
Caillou
1b
inversion
with
that
this
is
what
you
get
for:
L
you're,
seeing
the
clay
layer,
okay
and
both
aquifer.
You
may
see
something
reasonable,
but
the
bottom
part
kind
of
see
something.
But
it's
hard
to
say
whether
you
see
or
not.
Okay,
then,
if
you
do
a
sports
inversion,
that's
what
you
get
you're,
actually
cleaning
up
all
these
structures
and
then
you're
gonna
get
very
sparse
here.
B
C
B
C
B
C
Okay:
okay:
go
ahead!
Well,
any
comment
done.
E
A
E
B
Is
like
no
I'm,
not
saying
it's
bad
I
think
it's
not
sports,
so
to
make
your
buy.
My
point
was
to
make
a
really
sports.
If
that's
what
you're
looking
for
I
think
you're
setting
an
MRAP
is
very
important.
So
when
your
M
ref
is
little
wrong,
you
could
do
an
okay
job,
but
I
mean
like
yeah.
That's
what
I'm
saying
it's
like
an
executive
like
a
proper
and
rapid.
You
got
a
really
small.
B
E
B
B
When
it's
running
I
just
got
a
quick
comment
about
what
them
asked
his
number
here
that
nonlinear
version,
your
sensitivity
is
changing.
It
updates
every
iteration,
but
what
I
was
thinking?
It's
just
a
little
bit
of
cheat
because
let's
say
you
were
when
you
reach
to
L
to
solution
now,
you're
got
a
model
that
fits
the
well
so
you're,
pretty
close
to
solution
so
and
then
I
think
that
there
I'm
fixing
like
for
this,
like
for
the
person
version,
that's
a
starting
point,
your
L
to
a
smooth
model,
and
then
that's
your
starting
point.
B
Point
force
portion
version,
they're,
pretty
close
to
the
solution.
Well,
I
was
assuming.
Oh,
my
I.
Don't
really
need
to
update
the
sensitivity,
so
I'm
gonna
fix
the
sensitivity
matrix
whoa
when
I'm
or
mulling
and
I'm
using
the
real
model.
That's
what
the
masks
I'm
I
mean
that
way.
You
could
be
a
little
bit
more
efficient
and
I
found
like
that's
actually
fine
or
your
inversion.
Rdc.
B
E
B
Or
like
a
first
question
like
a
sparse
of
people,
the
after
l2
inversion,
I'm,
not
updating
this
if
I
use
the
exchange
so
here
this-
and
this
is
an
option.
So
if
X
3
matrix
that's
true
and
I'm,
not
updating
so
well,
we
could
be
a
little
bit
smarter,
like
at
the
end.
We
could
kind
of
do
a
one
more
step.
We've
been
using
up
like
updating.
B
B
C
Mean
another
another
trajectory
through
here:
gum
I'm,
not
sure
whether
you
ever
tried
this,
but
if
you're
going
from
the
L
to
then
to
it
to
a
different
normal
assay,
l0
I
mean
you
could
actually
step
in
successive
iterations
go
to
like
L
to
you
know:
1.5
1.2.
You
know,
and
you
know,
sort
of
decrease
the
L
as
you're
converging
to
whatever
final
value
of
P.
E
So
just
thing
to
just
march
down
the
the
p-value
slowly
to
where
we
want
to
go
and
when
I
was
saying.
Is
that
because
we're
cooling
epsilon
it's
it's
a
little
bit
similar
to
to
Doug's
idea
where
we're
going
from
we're
trying
to
transition
something
that
looks
like
an
altitude,
something
that
looks
like
an
LP.
That's
all.
B
Right
yeah,
so
those
sort
of
my
experience
and
one
thing
kind
of
cool
like
this
image
is
so
the
question
like.
If
somebody
would
question
okay,
can
we
see
really
big
conductor
here?
I
think,
like
this
two
images
saying
it's
kind
of
hard
to
see,
I
guess
we
probably
need
more
I
got
other
information
to
actually
kind
of
resolve.
This
conductor,
I'm.
E
C
B
B
Yeah
so
I
think
good.
There's
a
value
here
like
that
kind
of
what
dummy
you
suggested
in
an
excess
or
Jie
move
like
you're,
using
different
alarms
and
kind
of
like
having
some
sort
of
statistics
or
or
or
model
credibility
where
I
wouldn't
go
to
a
modal
uncertainty.
I
think
there's
a
probably
some
value.
B
B
And
I
hear
I
think
also
setting
a
reference
was
actually
pretty
important.
Yet
a
sports
model
when
I
was
flying
now,
I
also
fixed
a
day,
so
I
thought,
like
yeah
I'm,
not
sure
like
what's
the
best
way,
but
I
think
for
me
kind
of
having
an
okay
day
that
okay
inversion
reserved,
quick
fixing
J
seems
like
an
okay
good
option.
E
B
E
B
And
I
think
there's
a
lot
of
that.
What
I'm
excited
about
this
so
like
it
using
actually
in
3d
inversion
for
like,
for
instance,
DC
and
also
everyone
I
am
and
the
types
of
like
like
2d
and
3d
inversion,
I
think
it
has
a
lot
of
value
and
then
it
seems,
like
other
people
are
just
doing
on
the
smoothness
term.
That's
my.
E
B
So
yes,
I,
think,
there's
two
two
things
going
on
right,
this
so
smallness
a
smoothness,
so
you
can
kind
of
like
play
with
that
here:
yeah,
it's
actually
really
nice!
Yes,
I
really
want
to
kind
of
use
it
which
I
haven't
really
tried.
I'm,
really
like
these
see
any
MP
in
inversion.
So
that's
from
my
next
step.
C
B
During
the
beta
search,
okay,
good,
as
far
as
I
know,
what
he's
doing
is
like
he's,
comparing
the
and
then.
B
I'm
not
exactly
sure
but
I
think
he's
kind
of
like
changing
the
beta
value
rather
than
like
keep
cooling
it,
and
there
are
certain
conditions
like
that
kind
of
flame,
that
that
didn't
really
work
for
me.
So
I
think
I
had
to
always
in
a
fix
debate
and
not
dumb.
Can
you
help
me
I,
think
I'm
not
sure
where
I'm
going
so
what
I
used?
What
I
had
to
do
was.
B
E
E
B
B
B
F
B
A
B
B
B
Here
what
we
did,
what
I
did
was
actually
using
kind
of
three
different
data
set.
One
is
frequency
domain
result
and
the
other
Tam
I'm
Jimmy
p.m.
and
the
final
one
is
groan.
Brown
TM
data
called
nano
chem,
and
so
this
is
a
model.
So
what
I'm
trying
to
do
here?
Okay,
I
got
three
different
data
set
I
want
to
invert
jointly.
That
was
what
yeah.
B
This
is
the
structure
conductivity
we
got
a
new
surface,
resistor
up
top
and
then
big
conductor,
so
I
can
get
all
of
these
codes
actually
running
each
inversion,
separate
inversion
and
comparing
a
glass
I
think
this
is
the
movie
inversion
result.
So,
with
all
this
frequency
domain
system,
it's
very
sensitive
to
the
new
surface
compared
to
kinda
me
you're,
seeing
at
the
surface
resistor
well!
B
Well,
not
you
know,
you
can
see
the
existence
and
then
Uttam
under
a
system
is
a
deeper,
but
the
near
surface
resolution
is
and
of
course,
compared
to
results
so
they're,
seeing
the
deeper
part
well,
but
not
that
out
part
and
then
a
note.
M
is
the
ground
system.
Sorry,
it
has
the
strongest
sensitivity
feel
in
the
surface
part.
So
you
can
see
that
inertia.
Please
resist
it
pretty!
Well,.
A
B
B
Yeah
I
think
this
is
actually
important.
No
I
see
that
the
name
because
domain
and
ground
this
is
basically
what
you
need
to
do
or
joint
conversion.
So
you
just
need
to
generate
three
different
problem,
generate
three
different
misfit
and
then
sum
them
over.
So
here
we're
using
Lyndsey's
combo
trick
to
function
which
is
actually
pretty
cool,
and
then
you
can
do
the
same
thing
for
regularization.
B
Right
so
here
we
don't
need
to
do
anything
for
regularization
right,
so
we
got
just
one
conductivity,
so
we
don't
need
to
you.
Don't
need
to
use
the
combo.
Take
your
function.
Yep
three
inversion
was
inverting
for
a
single
conductivity,
now
I'm
going
back
to
joint.
That
was
actually
a
joint
inversion,
which
is
pretty
cool,
so
you're
kind
of
like
getting
all
the
good
things
into
if
you're
using
joint-
and
this
is
actually
a
l0
research,
so
you're
getting
sparse
model
for
resolve
and
time
and
mountain,
and
this
is
the
drinking
dirty
little
spots.
B
Think
it's
all
the
functionality
we're
getting
so
again.
We
can
actually
doing
conversion
or
converting
for
a
single
conductivity,
which
is
pretty
exciting,
and
this
is
a
simple
example
for
1d,
but
at
least
we
got
the
structure,
so
we
could
actually
run
through
Peter
yeah.
We
haven't
got
the
actual
capability
to
do
it.
I.
E
B
Yeah
so
I
kind
of
carefully
designed
the
survey
so
I
think
I
chose
I
chose
a
system
that
they
can
help.
That's
the
only
thing
that
I
did
thanks
in
each
survey,
have
some
pros
and
cons
and
then
by
summing
all
three
they
kind
of
like
a
compensate
each
other.
That
was
actually
what
I
did,
but
actually
I
got
cool
wow.
It's
a
fairly
well
fast.
B
A
B
D
B
E
B
Yeah
I
kind
of
want
to
learn
this
as
soon
as
possible
because
you
want
to
make
all
the
notebooks
that
we
use
for
our
one
course
available.
So
I
think
this
is
one
of
the
poster
that
I
can
use
for
that.
Yes,
yes,
so
filling
your
son,
yeah,
so
I
think
in
some
gravity
and
maggin
version
I
think
it
since,
like
you're
testing
the
sparseness,
so
I
think
that's
where
the
fit
like
tests
are
failing.
Yeah.
D
B
E
B
E
B
E
B
A
Alright,
there's
a
couple
of
things
to
share.
I'll
share
my
screen
here
in
just
a
second
there's,
a
couple
updates
to
simply
just
with
the
most
recent
releases
that
I'll
share
with
you,
and
then
there
was
a
call
one
of
Rosemary's
students
means
Meredith.
She
is
interested
in
doing
do
is
for
airborne
iam,
and
we
just
don't
have
any
examples
right
now
and
so
I
was
wondering
if
somebody
would
be
game
to
create
one.
A
If
there's
something
that's
relevant
to
what
you're
doing
it
doesn't
necessarily
have
to
be
Ian,
but
if
it
would
be
helpful
for
us
to
have
in
the
docs
anyways
it
thing
so
if
you're
interested
in
doing
eeehm,
that
would
be
great.
But
it's
not
something
else
relevant
to
what
you're
doing,
and
you
want
to
take
that
on
that'd
be
awesome,
so
she
was
gonna
write
up
an
issue.
I
haven't
seen
it
yet
so
I'll
ping
her
today
and
we'll
see.
B
C
You
know
how
much
currents
are
going
down
or
how
much
change
you're
getting
when
you
just
change
the
the
background,
conductivity
or
or
or
any
of
these
these
things
and
then
not
put
it
forward
like
okay.
This
is
how
you
calculate
the
DOI,
but
these
are
things
that
you
know.
If
you're
going
to
calculate
a
DUI,
then
these
things
might
might
be
useful
and
you
know
if
you
want
to
set
up,
you
know
on
threshold
parameter
and
what
you
can
see
or
not,
then
that's
fine.
C
A
A
So
if
you
actually
want
to
see
if
there
is
a
chain
here,
if
you
guys
are
running
on
different
versions
and
that's
what's
causing
the
problem,
this
is
like
a
very
transparent,
quick
summary
of
what
your
environment
is.
That
depends
on
for
that
simp
independent,
so
they
Keys
got
a
nice
version
that
prints
nicely
in
HTML
in
the
notebook,
but
it
also
runs
and
will
print
in
a
Python
script.
In
plain
text
or
I
think
ipython
kernel,
so
I
can
do
without
the
HTML
tag.
A
It
just
prints
in
plain
text
and
then,
if
there's
other
stuff,
that
you've
imported
that's
important
for
your
analysis,
like
maybe
you're
using
psych
Hitler,
or
something
like
that,
and
that's
really
important
for
what
you're
doing
there
are
there's
Doc's
on
how
to
actually
look
at
I'm
not
to
the
listing.
So
all
it
does
is
imports
it
in
checks.
The
version
in
principle
how.
E
A
A
There's
a
way
to
do
it
and
you
know
first
thing:
I
can
just
glance
it.
What
you
did
and
I
can
always
hopefully
go
back
to
the
email
or
it's
like
your
shirt,
yeah
yeah,
but
I
recommend
for
anything
that
you're
sending
out
for
publication
or
stuff
that
you're
pointing
people
to
in
repositories.
Just
add
this
spending
your
notebooks
yeah.
B
A
A
The
3d
SIL
stuff
I'm
working
on
trying
to
get
in
I
ran
into
some
issues,
just
with
versions
of
properties.
Franklin's
gone
in
and
looked
at
that
and
I
just
need
to
figure
out
what
he
figured
out
and
that's
what
we
should
be
on
that
on
the
same
page,
and
so
once
once
this.
These
two
floor
of
us
are
ready
to
go.
A
Then
we're
actually
in
a
place
to
merge
in
the
3d
SIL
code
in
disparate
eyes
and
so
there's
some
changes
I
made
in
discretize
that
require
that
we
make
updates
to
pay
the
the
biggest
thing
just
being
that
the
cell
grout
stencils
our
properties
rather
than
methods.
They
should
have
been
properties
in
the
first
place,
but
we
missed
that
and
so
everywhere
in
sin
think
there's
brackets
right
now
and
with
that
floater
question
I
just
put
in
that
fix
to
shoot
it
on
the
step
release
button.
G
F
F
A
Think
so
are
you
looking
at
multi,
processing
or
yeah
yeah
I
mean
I?
Think
that's
that's
worth
starting
to
bring
in
as
long
as
it's
like
clear
Oh,
especially
Windows
users,
can
use
so
I
mean
multi.
Processing
is
core
Python,
so
I,
don't
think,
there's
a
problem
starting,
oh
yeah,
so,
let's
just
maybe
before
it
goes
like
too
deep.
A
Then
the
only
other
thing
we're
gonna
create
some
issues
on
the
apps
side
of
things.
So
the
Houston,
of
course,
that's
coming
up
in
a
couple
weeks,
we're
hoping
to
make
heavy
use
of
the
apps
so
way
more
so
than
the
disk
like
the
disk,
was
sort
of
showed
people
and
download
a
couple.
This
we
actually
want
to
have
people
using
it
like
using
them
everyday
and
hopefully
will
sort
of
get
almost
all
of
them
used.
A
So
there's
a
couple
of
small
adaptations
that
we'll
be
looking
for
and
potentially
the
addition
of
some
apps,
but
the
biggest
thing
that
I
think
will
ask
for
help
on
is
just
extending
the
current
ones,
so
we'll
get
some
B's
written
up
and
then,
if
people
are
willing
to
jump
in
and
help,
hopefully
none
of
them
are
too
big,
so
it
completion
it.
Just
like
you
know
an
easy
thing
to
knock
off
in
an
hour
or
two.
A
A
C
C
Just
being
able
to
you
know,
have
the
apps
use
them,
connect
up
with
material
in
EMG
aasaiya
at
different
levels
and
all
of
the
case,
history
and
physical
surveys
just
to
be
able
to
kind
of
somehow
to
walk
through
that
seamlessly
and
have
the
you
to
become
familiar
with
all
parts
of
that.
If
we
could
accomplish
even
a
good
step
towards
that
final
goal,
I
think
we'll
really
have
something
that's
going
to
be
useful.