[Ifeffit] weights data processing

[Kelly, Shelly D.]

> -----Original Message-----
> From: Stefano Ciurli [mailto:stefano.ciurli at unibo.it] 
> Sent: Friday, July 23, 2004 6:31 AM
> To: ravel at phys.washington.edu; XAFS Analysis using Ifeffit
> Subject: [Ifeffit] weights data processing
> 
> 
> Hi folks,
> 
> 1) when removing background using athena, there is an option of using 
> a k-weight different than 1 (I am NOT talking about k-weight for FT). 
> I noticed that changing this parameter the resulting chi(k) does not 
> change significantly. Actually, going from 1 to 4 as a value for 
> k-weight for Bkgd removal, the only effect I see is a significant 
> reduction of the peak below the first coordination shell in chi(R) 
> while the other features of the spectrum are essentially the same in 
> terms of absolute and relative intensity. A drastic change in both 
> chi(k) and chi(R) is seen if I use k-weight = 5 and above. What is 
> the exact meaning of it? I know what the meaning is in the FT 
> process, but I do not see the reason why I should use a weight 
> different than 1 (let's say 3 for my noisy data) for background 
> removal. How should I behave? Should I take the largest value of 
> k-weight so that the essential features of the chi(k) and chi(R) do 
> not change?

Hi Stefano,

I have found a procedure for removing the background that "almost"
always works well.  It goes something like this and uses different
k-weightings in the background removal process.

1) Check out the curvature of the XANES region and how quickly the data
goes from zero to 1, call this DExanes.
There are in my small opinion two categories for each of these different
parameters.  Curvature of Xanes can be step like or it can have a large
peak.  DExanes can be small ~10eV for k-edges of Cu or wide ~30eV for
L3-edge of U.   

2) Choosing Ezero for the first time:  If you have a wide DExanes, then
choose Ezero up near the top of the edge.  If you have a small DExanes,
then choose Ezero around the middle of the edge.

3)Set kweight equal to one or zero (and no theory) and remove the
background.  This will emphasize the low k-region around the edge, but
often get the background for the higher k-region not as
well...especially for noisy data.

4)Check out the shape of the background function, the chi(k) data and
the FT.  Make sure that rbkg=1.0 is a reasonable thing to do.  Make sure
that you look at the chik data with all three k-weights.  Often people
say that there is no change in the chik data by changing the kwieght but
they are looking at the chik*k3.  The change will be greatest at low k,
which you don't really see with chik*k3.
  
5) If you have a large peak in the XANES region, Athena may need help
getting the bkg to go through that peak.  You have two options.  A)
increase rbkg so that the spline has more freedom to curve in this
region. or B)  increase kmin of the spline range, essentially removing
the large peak from the data.

6) Fit the first shell of your data, use a well padded chik data
range...don't include at least one full oscillation at the beginning.
i.e. use kmin=3 or greater. Monitor the value for ezero.  If it is
greater than 10eV you need to start over. (I have some more details on
how to create a theory to be used for bkg removal on my web site
http://www.mesg.anl.gov/Skelly.html from the 2004 NSLS EXAFS workshop
presentation)

7) Read the theory back into Athena.  Now Athena knows what the low
k-range should look like, and you can increase the k-weight of the
background, use kw=2 or 3, to get the high k-range part of the data
correct as well.  

8) Fit the data again and use as much of the k-range as possible.


> 
> 2) In artemis, is it a bug or a wanted feature to have the 
> possibility to activate more than one button for k-weight for the fit 
> and FT? I suppose it is a wanted feature, but then I wonder what is 
> the meaning of using more than one weight for the fit and FT. I am 
> sure some of you can unveil this enigma.
> 

Another couple of reasons to use more than 1 k-weight in the FT of the
data:
I have found that the uncertainties in the fitted values are about half
as big by using all three k-weights, as opposed to using just one
k=weight.  I have also found that using all three k-weights will force
you to model the data much more correctly.  When you look at the fits,
you need to look at the FT and the chi(k) data using all three
k-weights.

HTH
Shelly