suggestion: Include lower and upper boundaries for fitting

[Rogier Besselink]

Hi I recently started working on pdf analysis and
I like pdfgui since it is a relative straigthforward program to work with.

However, I do have a suggestion to include optional lower and upper boundaries.

I try to fit a structure of iron sulfide, which seems to be some kind distorded nanosized version of mackinawite or greigite. However, during fitting some parameters tend drift away to ridiculous values, such as negative scaling factors, atoms outside unit cell, occupation factors > 1.
Therefore it would be nice to restraint such parameters within an acceptable range, i.e. by introducing lower and upper boundaries.

kind regards,

Rogier Besselink

[Pavol Juhas]

On Fri, Jul 10, 2015 at 03:37:41AM -0700, Rogier Besselink wrote:
...

> I try to fit a structure of iron sulfide, which seems to be some
> kind distorded nanosized version of mackinawite or greigite.
> However, during fitting some parameters tend drift away to
> ridiculous values, such as negative scaling factors, atoms outside
> unit cell, occupation factors > 1.
> Therefore it would be nice to restraint such parameters within an
> acceptable range, i.e. by introducing lower and upper boundaries.

Hi Rogier,

It is already possible in PDFgui to force simulation variables within
the desired limits by using suitable constraint equations, for example

Positive scale factor:
@1 * @1 (or "sqr(@1)" where "sqr" is the square function)

Occupation less or equal 1:
sqr(sin(@10))

Fractional coordinate within the [0.5, 0.7] bounds:
0.6 + 0.1 * sin(@50)


That said, if variables refine to non-physical values it is usually
a sign of some problem with the model. For example, there could
be too much of a difference between the measured PDF and simulation
or perhaps some of the phases in a multi-phase refinement is not
present in the sample. Another option is that refined variables have
either too little effect on the simulation (ie, are not reliably
expressed in the data) or are strongly correlated. A way to deal with
this is to start with all displacement parameters set to small
isotropic values and then refine only the cell parameters and scale
factor(s) with all other parameters fixed. If this gives fit with
a reasonably matching peak positions, you can start to refine more
parameters, while making sure the fit remains stable and there are
clear improvements in the fit residuum Rw for each extra parameter.

Hope this helps,

Pavol

--
Dr. Pavol Juhas
Condensed Matter Physics and Materials Science Department
Brookhaven National Laboratory
P.O. Box 5000
Upton, NY 11973-5000
tel: +1-631-344-3594
fax: +1-631-344-2739