►
From YouTube: SimPEG Jan 10, 2017
Description
SimPEG meeting on January 10, 2017. Rowan introduces the meshing package discretize and updates us on pymatsolver changes
A
So,
like
I,
don't
know
how
much
there's
really
to
talk
about
things,
and
so
basically
I
want
you
guys
to
interrupt
me
and
ask
questions
so
like
what
what
we've
decided
to
call
the
package
coming
out
of
this
impact.
Sort
of
this
is.
This
is
encompass
all
of
the
finite
volume
techniques
that
were
in
the
sim
peg
package.
So
this
is
modular
with
respect
to
the
spatial
discretization,
so
doing,
1d
2d,
3d
doing
tensor
meshes
regular
mass
meshes
logically
rectangular
non-orthogonal
meshes
on
n
and
cylindrical
as
well
as
tree
meshes.
A
A
Rectangular
but
non-orthogonal
and
then
the
tree
mesh
as
well
so
like
this
is
again
all
all
pretty
standard
stuff.
So
not
a
whole
lot
has
changed
since
pulling
it
out,
except
some
cleanup
of
utilities
and
namespace,
and
things
like
that
and
actually
pulling
it
out,
has
shown
a
couple
interdependencies
between
the
solver
classes
and
meshing
classes
and
making
sure
that
those
are
also
separated
and
so
I'll
talk
about
some
of
the
solver
stuff,
the
future.
A
Yeah,
so
basically,
what
I've
done
inside
sim
peg
is
and
again
this
is
not
a
polar
custom.
I,
don't
think
that
should
go
on
on
the
current
the
lease
to
master,
but
for
the
next
one,
I'm,
basically
pulling
out
this
mesh
stuff
pulling
out
some
of
the
solver
stuff
and
leaving
from
from
sim
peg
import
mesh,
which
will
just
be
basically
assembling
to
discretize
package.
C
A
Couple
of
other
things,
I
actually
made
like
some
more
debt,
making
the
gallery
look
a
little
bit
nicer,
using
some
open
source
tools
yeah.
Hopefully
we
can
add
a
couple
other
examples
here:
Lindsey
and
I
have
been
working
on
a
bit
of
explanation
around
this
so
and
I
can
share
this
when
it's
in
the
other
state,
but
basically
going
through
this
talking
about
the
sort
of
anatomy
of
the
various
cells
for
the
tensor
mesh
for
this
logically
rectangular
non-orthogonal
mesh.
A
A
Especially
in
weak
form,
so
I'll
share
this
when
it's
in,
like
a
slightly
better
states
on
the
synthetic
drive
and
then
hopefully
we'll
publish
it
somewhere
as
well.
If
not,
it
will
be
in
my
thesis
yeah.
So
another
part
of
this,
when
I
was
pulling
things
out,
is
that
for
the
tests
I
needed
some
access
to
the
solver
utilities,
but
I
didn't
want
to
make
discretize
dependent
on
sim
peg
I
want
it
to
sort
of
heat.
The
other
way
around
so
I
also
spent
some
time
working
up.
A
The
pie,
Matt
solder,
which
just
making
it
sort
of
it,
was
split
between
sim
peg
and
hi
Matt
solder,
so
like
sort
of
complicated
solvers
wherein
I
met
solver
and
the
simple
ones
were
in
synthetic,
so
I
just
sort
of
put
all
of
them
into
I'm
at
solver
again.
This
is
not
actually
live
inside
sandbags,
so
simpang
is
still
stand
alone,
but
now
all
of
that
code
based
ink,
I
meant
solvers.
So
it's
actually
sort
of
lu
solves,
like
triangular
solves
they
wraps.
Some
of
those
CyHi,
sparse,
matrices,
I.
A
Think
someone
who
wrote
to
Jacoby,
solver
or
preconditioner
in
there
does
Paradiso
with
MK
I'll
help
again
thanks
to
Dave
and
Joe
for
that,
and
then
someone
on
the
internet
bugs
me
about
keeping
mumps
in
there.
So
that's
still
in
there
as
well.
Although
this
the
mumps
one,
does
not
work
on
pip
install,
it
should
work
on
like
into
the
main
repo
and
just
say,
make
mumps
or
make
mumps
Matt
and
I've
left
some
notes
into
reading.
B
A
I,
don't
maybe
Joe
you
can
talk
to
that
about.
So
one
of
one
of
the
things
that
we
were
trying
to
do
is
make
the
install,
especially
the
pip,
install
a
little
bit
easier
and
more
tested
on
the
both
Windows
and
Linux
and
Mac
machines.
So
I
got
the
tests
up
and
running
on
App
there,
which
is
a
test
testing,
continuous
integration,
testing
platform
for
windows
as
as
well
as
on
Travis,
and
so
those
are
now
working
after
learning
about
a
day
and
a
half
of
my
time.
A
D
C
A
A
Mum's
you're
gonna
have
a
lot
more
trouble,
but
paradiso
should
actually
work.
So
what
made
that
really
much
easier
is
that
Conda
actually
has
I
think
Intel
like
open
sourced,
ish,
MKL
and
so
Conda
wrote
they'd
rather
for
that
and
as
a
sort
of
included
it
in
the
base,
Python
distribution
that
they
provide
and
so
Dave
roads
a
wrapper
for
MTL,
which
job
fixed
and
I
wrapped
with
her
parents.
E
A
Insert
interrelation
so
I
have
tested
psych
on
actually
the
compilation
on
an
app
or
on
a
player
which
so
it
works
on
a
Windows
machine.
I
think
the
problems
that
we
were,
having
mostly
related
to
the
set
up
high
final,
which
I
did
another
few
iterations
on
after
reading
the
monologues
on
the
internet,
I,
don't
know
Joe,
you
might
have
more
to
say
on
the
structure
of
that,
but
I've
simplified
it
a
lot.
So
it
should
actually
be
at
least
transparent.
What?
What
I'm
trying
to
do?
Yeah.
D
The
structure
that
is
on
my
cell
is
that,
if
I
can
follow
is
not
set
up
as
well.
Sigh
fights
I
have
like
yeah
our
sub
packages
of
and
leave
everything
until
a
later
package,
and
it's
just
kind
of
taking
care
of
on
it
to
me,
it's
less
in
one
giant
set
up
that
pie
at
the
beginning
and
it's
like
I
said
just
passed
on
to
individual
ones
inside
each
of
the
sub
packages.
Okay,.
C
A
Yes,
yes,
so
actually
inside
sim
back
now,
there
isn't
any
siphons
and
inside
I
met
solver.
That
should
install
it's
actually
pure
python.
The
dependency
is
in
on
high
or
it's
on
MKL,
which
means
that
you
have
to
use
content
which
people
should
sort
of
be
doing
anyways,
so
yeah
I
think
those
things
have
been
simple,
pretty
radically,
and
basically
the
only
thing
that
will
not
work
is
the
interpolation.
If
the.
E
A
Yeah
yeah,
so
so
it's
this
icon
set
up.
That
pie
is
tested
on
a
Windows
machine
so
that
that
should
work
and
yeah
I
mean,
like.
Obviously,
people
have
got
it
working
I,
don't
I.
Just
don't
know,
I,
don't
know
too
much
about
packaging
things
up
in
the
setup
that
hide
stuff.
So
if
anyone
else
has
thoughts
there
or
Joe,
if
you
have
thoughts,
do
you
have
a
windows,
machine,
Joe.
A
D
A
C
B
E
E
D
Just
got
a
kind
of
Forge,
but
I
have
my
wheel
files,
which
are
the
new
version
of
whatever
eggs
that
python
used
to
use.
So
when
you
see
death
like
installing
wheels,
I
know
what
exercise
they
do,
that
it's
just
everything's
pre-built.
So
it
just
comes
right
to
the
machine.
So
they
don't
through
my
strong
with
any
compilers
which
we're
I
mean
you
guys
always
have
to
have
those
issues
on
Windows
and
installing
the
right
c
compiler.
Everything
yeah.
C
A
C
C
B
I,
don't
actually
think
there's
too
much
code
that
touches
it
it'll,
hopefully
trim
this
down.
I
would
perhaps
like
to
see
us
doing
holes
from
deb
to
master
a
little
more
often
because
this
is
12,000
close
to
thirteen
thousand
lines
of
code,
which
is
same
like
I
am
not
capable
of
doing
a
decent
review
on
that.
But
that
says
we
need
to
get
in
and
I
think
we
need
to
get
it
in
student
and
why
that's
important,
especially
now,
is
the
this.
C
B
Coming
up
and
we
are
going
to
be
distributing
simpang
mode
and
like
pointing
people
to
to
make
so
reach
out
to
be
so
I
guess.
The
thing
that
I
would
ask
is
that
just
like
take
a
take
a
skim
through
if
their
stuff
that
you
touched
that
is
in
that
part,
just
document
it,
because
I
think
the
thing
to
do
is
basically
get
this
in,
like
I
think
we
should
change
the
depreciation,
get
rid
of
that
I'd
appreciate,
y'all,
find
them
and
then
once
that's
gone,
let's.
B
E
C
E
B
E
B
C
E
C
A
A
B
A
E
A
B
A
Joe
another
thing
that
I've
been
sort
of
pushing
on
is
a
big
refactor
in
on
the
Richards
equational
stuff,
and
so
that's
actually
a
lot
cleaner
and,
in
terms
of
you,
can
now
invert
for
any
physical
property,
and
it
actually
makes
sense
before
I
was
sort
of
like
making
in
concepts
of
log
on
the
activity
and
that's
been
fixed
and
actually
soggy.
You
might
be
interested
in
the
way
that
I'm
functional
models
for
like
non-linearity
or
like
dependence
in
your
tasting
IP
on
frequency
mine
is
dependent
on
the
field.
A
A
A
C
E
E
C
E
A
I
think
maybe
order
of
operations
there's
some.
I
think
the
composable
objective
function.
Work
that
lindsey
is
doing
is
probably
a
first
step
to
that
in
that
we've
been
sort
of
renaming
and
restructuring,
howard
accessing
derivatives
and
second
derivatives,
and
that's
one
of
the
big
things
in
the
optimization
code
that
it's
sort
of
the
way
that
you
interact
with
it
gets
all
three
of
the
function:
evaluation,
the
gradient
and
the
hessian
at
once,
which
is
a
little
bit
different
than
most
optimization
packages,
especially
side
high.
Is
they
actually
say
that
you
should
suffer?