[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: FEFF question?

Rich Reeder has inquired about the differences between FEFF7 (e.g.
feff702) and FEFF8 (e.g., FEFF8.10).
 
> > We've been using FEFF7 but have just started to use FEFF8.1, so far just
> > for basic EXAFS fitting.  We happened to notice that the phases and to a
> > lessr extent the amplitudes calculated by FEFF8.1 are noticeably different
> > than those calculated by FEFF7 (using the same input file and with the same
> > "keywords").  In other words, when we compare the different phase and
> > amplitude columns (basically cols 2, 3, and 4) in the FEFFnnnn.dat file for
> > a given path, the results are different depending on whether we use FEFF8.1
> > or 7.  In doing this comparison, we specifically did not select SCF or FMS,
> > so the calculations should be comparable (we think).
> >
> > BUT, when we then use the paths to do the fitting, we find basically no
> > difference in the fit results for paths from 8.1 or 7.  So, we're puzzled
> > about two separate things.  First, why the difference in the phases and
> > amplitudes (much smaller difference in amplitudes)?  And second, why is
> > there NOT a difference in the fit results if the phases and amplitudes in
> > the FEFFnnnn.dat files are different?  (Actually, there is a systematic
> > difference in the fit results in E0, but nothing else).  I assume we're not
> > the first to see this.  Have we overlooked something very simple?

 The differences between FEFF6, 7, 8.00 and 8.10 are fairly subtle. The biggest
change was between FEFF6 and 7 in the atom code. The other major change is
in how the average interstitial parameters are calculated: 

FEFF6 uses Hartree-Fock-Slater atom and average interstitial parameters
FEFF7 uses Dirac-Fock atom and average interstitial parameters
FEFF8 uses Dirac-Fock atom and improved interstitial parameters 

  Also going from FEFF7 to FEFF8, we adopted a new default muffin-tin overlap
factor of 1.15 which seemed to give good results in a wide range of systems.
The default in FEFF7 was 1.0.  As a result of these changes, one will notice
small changes in E0 and hence the phase shifts. Along the way lots of
seemingly innocuous changes were also made to make the FEFF8 compile and
run without complaints on any operating system.

> For the second question: I'm not sure why there's no difference in the
> fit results.  Is it possible that some of the amplitude is just moved
> from one column to another??  For feffnnnn.dat, the amplitude used is
> [col 3 * col 5] (mag[feff] * red factor), and the total phase shift
> used is [col 2 + col 4] (real[2*phc] + phase[feff]).

  This must happen, since the numbers printed out in feffnnnn.dat are
all renormalized so k=0 at E=E0. [For example, the factor exp(i2pR)=exp(i2kR)*
exp(2i(p-k)R) and the exp(2i(p-k)R) is absorbed in the other factors.]

> I think the general consensus has been that Feff8 doesn't improve
> Extended XAFS over Feff7 (or even over Feff6 for non-polarized
> calculations).  I'm still using Feff7 myself.

  We would be interested in seeing your (or anyone else's) numbers on this.

  Our benchmarks using Matt Newville's feff7test suite, on Cu, Au, and Ag
show that non-scf FEFF8.1 is usually better than FEFF7 in the sense of a smaller
chisqr for EXAFS. FEFF81 also does better for complex, heterogeneous systems,
e.g., systems with H atoms.  However FEFF6 also does well by comparison on
the test suite; this may not be entirely surprising since FEFF6 was a very
well tested and robust code and the physics in the EXAFS regime is not
very sensitive to chemistry.  In any case the EXAFS numbers are not significantly
different for any of these codes. But the XANES definitely are, which is
where FEFF800 and 810 have their biggest advantage.

Results from the f7tests on Ag, Au and Cu data of M. Newville
Ag:   used  reff = 2.892AA      T  =  80K
                feff6      feff7          feff800     feff810
red chi-square =  343.       78.              553.        324.

 
Au:   used  reff = 2.885AA      T  = 300K
                feff6       feff7         feff800     feff810
red chi-square =  377.       817.             252.        371.


Cu:   used reff = 2.553AA   T  = 150K
                 feff6      feff7          feff800     feff810
red chi-square =  18.6       27.9             19.1        23.9

  Sincerely
  John Rehr
  PI FEFF Project