[Ifeffit] a question about background refinement

[Matt Newville]

Hi Stanislav,

Yes, that makes me a little bit uneasy too.  If we choose to be
optimistic, we could be reassured that the amount of Ru-O is fairly
independent of how you do (or re-do) the background refinement.

For the most part, I don't find re-doing the background subtraction
a problem.  One can view autobk/spline() as the crude approach and
including the background in the fit as refining this crude
background.  And, we _know_ that the normal use of autobk/spline()
(that is, without a 'standard') is imperfect: the idea that chi(R)
should be as small as possible below Rbkg is not strictly correct,
as there should be non-zero contributions 'leaking' to low-R from
the first shell (especially when it's oxygen, which is a large
fraction of the EXAFS on interest!).  The repeated background
subtractions do not by themselves make me uneasy.

The slightly negative sigma2 doesn't bother me too much either, as
that's close enough to zero to be influenced by non-ideal
normalization, energy resolution, S02, and how the Ru-metal shell
fit is done.  It seems that you're finding that the first shell
sigma2 is correlated with the background parameters, so that setting
sigma2 to 0.010 and re-determining the background works simply by
altering the background function.  That seems like a reasonable
constraint on the background function.  I've usually found E0 to be
most highly correlated with the background parameters, but the first
shell sigma2 and R are also correlated with these.  (if they weren't
somewhat correlated, it would make no difference how you did the
background subtraction).

The thing that makes me the most uneasy is the relatively large
values of Rbkg/Rmin (that is rbkg in spline() and rmin in
feffit(do_bkg=1,....) which also refines the background) and how
they change around in your analysis. Normally, you'd want to keep
Rbkg/Rmin below the 'first shell peak', which is sometimes difficult
to define for the asymmetric peak common with short metal-oxygen
distances, but if the peak is around 1.3Ang, the Rbkg/Rmin of 1.5Ang
seems like it might be too high.  Can you get decent results with
a lower Rbkg/Rmin?  

Just to be clear, spline() with Rbkg=1.5 *tries* to remove the
component below 1.5Ang, but it's always incomplete and not a shart
cut-off.  In feffit(), setting do_bkg=1 makes the fit really use an
R-range of [0,Rmax], and adds variables for the refined spline
according to Rmin, but the correlation between spline and structural
parameters gets messy around Rmin -- which is why you do the
refinement of the background.

I'm not sure that's very clear, but I hope it helps,

--Matt

On Tue, 13 Apr 2004, Stanislav Stoupin wrote:

> Hello all,
> 
> Recently, I was able to improve fitting results in one of my projects
> using background refinement/subtraction in some "strange" way.
> I have got reasonable results but I was not totally sure
> if that was a valid approach. So, any feedback is appreciated.
> 
> Here is the story:
> 
> The system I am working on is Ru-Pt fcc alloy in a form of nanoparticles
> which may be oxidized. There is a reason to suspect presence
> of Ru-O bonds in the system.
> For the moment I am doing only the first shell analysis and the model
> I am using includes a part that models alloy first shell structure and
> also Ru-O path.
> 
> If I don't use the Ru-O path in my model all the alloy parameters
> that I get from the fit are reasonable but the fit works bad in low R region.
> In that region there is a broad peak around 1.3A.
> The rmin parameter which is responsible for background refinement
> is set to 1A ( the same value was used when I did initial background
> subtraction in Athena).
> 
> What I found a long time ago is that the fit could be significantly
> improved  if I add the Ru-O path in my model
> (all alloy parameters were even better in that case).
> However, the DWF of Ru-O bond always went negative (about -0.005).
> At the same time I got reasonable value for the Ru-O amplitude
> which was about 0.3.
> 
> I attempted to get rid of the negative DWF value many times.
> For example, if I fixed or restrained the DWF to a reasonable value
> (0.01 is what I expect) it didn't help.
> 
> Finally, I decided to play with background refinement and what I have
> done can be described as follows:
> 
> 1. Set rmin to about 1.5 A
> 2. Fit with my model which included Ru-O path with fixed
>    delr = 0 and DWF = 0.01 ( the Ru-O amplitude was allowed to vary).
>    Backgound fitting was turned on.
> 3. Subtracted the background from the data
> 4. Moved rmin back to 1A and allowed Ru-O delr to vary (the DWF
> remained fixed to 0.01)
> 5. Fit again with no background fitting
> 
> A very decent looking fit came out as a result. (Chi-square was
> improved) The Ru-O amplitude was about 0.3 (as expected)
> and I had the DWF set to 0.01.
> 
> Thanks for you patience if you read it all the way through.
> My own impression of what I have done is that it looks very artificial
> and the fact that I use AUTOBK two times bothers me a lot.
> The second time it is supposed to remove all structural
> information below 1.5A (rmin value). However, with this approach
> I hope to get rid of the background that dominates the fine structure
> in the low R region (Ru-O scattering) using some preliminary
> knowledge about the compound.
> 
> I will be happy to hear any point of view. Thank you in advance.
> If you are interested and want more details please let me know.
> 
> 
> Stanislav