[Ifeffit] Possible bug in atoms files generation by Artemis

Matt Newville newville at cars.uchicago.edu
Wed Mar 27 21:32:23 CDT 2013


Hi Matthew, Bruce,


On Wed, Mar 27, 2013 at 10:33 AM, Matthew Marcus <mamarcus at lbl.gov> wrote:
> Some users do have FEFFx (x>6l) on their own, so it would be useful to
> prepare Artemis/Demeter/Larch... for them and provide
> methods for using higher versions if an executable is present.

I completely agree that this would be valuable, and that it's not
quite moot because many people do have Feff8 or Feff9 available to
them.

But, I also want to be clear that I totally agree with Bruce's point
on this as well.  It's really hard to support code that is actually
poorly defined and cannot be changed.  The Feff input file format is a
complete mess, and it's not Bruce's (or my) fault.  In fact, this *is*
what the original question was about.  Feff changed form one awful
form to another -- why is this Bruce's problem to solve?     There is
no API or library provided for reading feff.inp.  We are forbidden
from changing it or redistributing a changed version of Feff8.   It
is, to be very clear,  the choice of the Feff project to break these
input files.    If we had a "Feff8 for EXAFS" that could be relied
upon and redistributed, it would be a totally different story.   We
don't.

We've lobbied (begged) for changes in the i/o and freer access to the
code for many, many years.  I can't explain the restrictions in any
rational terms, fathom why this restriction is a priority for John or
anyone else, or understand how anyone who writes scientific software
would use anything except an open-source license.  I've given up on
pestering them about it.  I was told last summer that a version of
Feff8 for EXAFS that we can distribute would be made, but haven't
heard a thing about it since.  It could happen soon .... that would be
great.  The license they use is their choice, and I respect that even
if I don't actually like their choice, and actually have real
reservations about the choice being solely theirs to make.
Ultimately, science will demand that all versions of Feff will be made
free or be forgotten.  It is completely believable that Feff6L will be
what is used twenty years from now.  I expect that Feff8 for XANES
calculations codes will never be made free and will be forgotten in
time.

I actually don't have a copy of Feff9.  Kevin J did a great job with
JFEFF, and emulating or including this approach of using a remote
cluster in the analysis codes would be interesting to think about.
>From a practical point of view,  it would be easier to reproduce a
similar system than to have to argue about licensing.  Then again,
without the calculation engine being freely available, running it on a
cluster actually seems like a problem... wouldn't you need a license
key or something?  Right, sorry, the Feff license actually makes no
sense -- it's better to not think about this.

> What does the multipole self-energy do?  Is that the thing that requires the
> dielectric response?  As you point out, the purpose of
> the exercise is to analyze unknowns, so by definition one doesn't have the
> dielectric response.  We can't expect a program that runs on
> a PC to do a proper, all-electrons, excited-state calculation.

Yes, the multipole self-energy uses a better model for the self-energy
based on the dielectric response of the system in the low (electron)
energy regime -- in short, how a "free" photo-electron will travel
through the multi-electron system.  So, yes, in principle one needs to
know the dielectric function.  That's conceivable for Cu metal, and
maybe an ion in an aqueous solution, but not so much for lots of
systems really studied.  Still, if you think that a single-pole model
works pretty well (the current normal implementation of the
self-energy), perhaps a mediocre dielectric response function (say,
treating 'iron oxides' as all more or less the same) would be an
improvement.

> One thing I do is to use experimental data from models as sources of
> amplitudes and phases.  At present, I do this using my own
> multishell fit program.  Is there an easy way to do this in your programs?
> What I think would be nice is a subsystem which allows one
> to do the filtering and extraction of amp and phase from a model .chi file
> from within the program, rather than having to create
> FEFF-path-like files.

Creating a model "Path" from experimental data is definitely possible,
but not conveniently encapsulated in Larch -- it would not be hard to
do, and is a fine idea.  It's now on the to-do list.  In years past,
Anatoly has done similar things by making "fake feff.dat files" and
using them in Feffit.  It can work, but it should be made easier.

> As long as we're talking wish-list, I'd love to have some way of defining
> atom positions using symbolic variables and have the system
> compute the path distances automatically as functions of those variables.
> That way, I could, for instance, define a dopant system in terms
> of the displacements of the near-neighbors without having to do the geometry
> by hand, which is difficult, tedious and error-prone.
>         mam

Oh yes, this is a major motivation for wanting to replace the
PathFinder with a version that was reversible (ie, which atoms in the
cluster correspond to this structure, and vice versa) and for wanting
to move the GenerateFeff_MuffinTin(Potentials, Path) into the fitting
loop.   Ultimately, we'd like to replace the sum-over-paths with a
sum-over-atoms-in-a-cluster, where the x,y,z coordinates would be the
variables.    I think it's completely possible, though a fair amount
of work.   Probably needs a person to really dedicate some serious
time to it...  Sorry to sound like a broken record, but if anyone is
interested in this, let me know.

Cheers,

--Matt



More information about the Ifeffit mailing list