To: "L. Cranswick" [L.M.D.Cranswick@dl.ac.uk] Subject: Re: CCP14/RIET: GSAS Macro examples files for setting up restrained inorganic refinements. Sender: "J.P. Wright" [jpw22@cus.cam.ac.uk] Lachlan, You can feed expedt an input file which contains whatever keystrokes you would normally give to the program. Just type "expedt < input.txt" at a dos prompt, where input.txt contains the keystrokes. In theory it would be possible to set up a batch file for a cascading refinement, but you'd need a lot of work to make it intelligent enough to figure out which atoms to let go at what time. It would be possible to write a helper program which could figure that kind of thing out for you, but then you'd be looking at a black box, which I don't think is welcome in the powder community. Every problem usually has it's own quirks. As for convergence problems, it's worth playing around with the damping until things settle a bit and then reducing the damping factor. It depends on your data and space group, but 100 atoms in the asymetric unit, all with 3 xyz's suggests you'd need around 3000 peaks to get a 10 to 1 data to parameter ratio. That's a lot. Frequently you can get away with refining with a subcell or higher spacegroup than really applies, then look at your structure as a distortion of the one you've got converged. Convergence strongly depends on the data quality. And obviously beware that if your starting model had some pseudosymmetry then it tends to fly apart on the first refinement cycle. Best wishes, Jon |
Date: Tue, 31 Oct 2000 09:15:06 -0700
To: rietveld_l@ill.fr
From: vondreele@lanl.gov (Bob Von Dreele)
Subject: Re: CCP14/RIET: GSAS Macro examples files for setting up
restrained inorganic refinements.
[text deleted]
Lachlan wrote:
>PS: While GSAS scales very well in terms of increased problem size and
>interface control - I am having some difficulty with getting convergence
>within Genles on a "real" dataset. (this is on Le Bail fitted data where
>only the background and atomic parameters are allowed to refine - starting
>from an idealised starting model guided from TEM)
Convergence for big problems is somewhat different from little ones. You
can now adjust the convergence criterion (it's a log scale!) and invoke a
very heavy duty global damping factor (Marquardt factor) to keep things
under control. The convergence factor probably should be scaled with the
size of the problem. I use a convergence criterion of ~50 for a 1000 atom
protein (>3000 variables). Note the default is 0.01.
>Does anyone have hints and tricks for parameter release conditions for
>large inorganics (that required restraints to keep the Oxygen distances
>reasonable) (~100 atoms in the asymmetric unit - may have to be doubled
>based on reinterpretation of the TEM data). All the metal sites have
>potential dual occupancy with that can be varied. What can you get away
>with and have a good change of convergence with Genles? How far away from
>their true positions can the heavy atoms be to drop in? Estimates of
>a good numbers of cycles to use to let the atoms rattle around?
Another feature of GSAS is the availability of a band matrix approximation
to the full matrix. Assuming that the atoms can be ordered by their
connectivity - easy for a protein, hard for a complex oxide; then a band
matrix can be used. I typically use a 300 variable band for a 3000 variable
problem. That will cover 10-12 amino acid xyz's which will catch all the
near-neighbor interactions captured in the restraints. With this the
convergence is pretty clean & smooth. Some fiddling then is done to
complete the refinement - mostly fixing messed up stereochemistry and then
repeating the refinement. A 1000 atom protein will converge in about 50-100
LS cycles (including fiddling) with the starting positions 1-2A away from
the final ones.
>Are there possibilities for what I believe on the commercial(?) version
>of Shelx 76(?) was a "cascading" refinement(?) - to automatically
>release small sets of atoms - and cycle through automatically?
>(is it possible to interlink expedt and genles under macro control
>to achieve this?)
I have done this for a protein - refine the structure using a sequence of
overlapping blocks doing one cycle of LS for each block. It can be setup
using DOS bat files and GSAS macros. It works but the convergence is slower
than using the band matrix approach.
By the way I've been doing this with powder data but all these facilities
are available for single crystal data as well. I haven't tried refining a
single crystal protein data set yet but it should work very well.
|