I have never seen a refinement "converge" in the mathematical sense.
The optimization algorithms we use are not expected to converge until
you have run a number of cycles greater than some multiple of the number
of parameters in the model. No one is ever going to do this.
The mistake I have seen, and what is being recommended against, is to
just run some standard number of cycles and quit without checking to see
if anything is still happening. Some programs have a standard of three
"big" rounds of refinement but the R values and over all geometry might
still be changing at the end of that, particularly early in a project.
You certainly shouldn't stop if the major stats are in flux (either
going up OR down).
More subtly individual atoms may still be moving while the overall
stats look stable. You should check that the largest shifts in your
final rounds are "small". This is a fuzzy definition, but if you are
interested in distances to a precision of 0.1 A you certainly shouldn't
be stopping if you still have atoms moving in steps approaching that size.
Refinement is fast these days. I tend to beat my refinements to
death just to be sure. If the standard practice is three "big" rounds
I'll run ten.
There is no such thing as "over refinement" only "over fitting".
Dale Tronrud
On 8/5/2016 3:22 PM, Keller, Jacob wrote:
>>>- make sure refinement is done till convergence.
>
>
>
> I have never understood and still do not understand what is meant by
> convergence—in my experience, refinements are always incremental, with
> various improvements here and there, but there’s never been any magic
> moment of convergence. Is there a rigorous meaning to this term? Is it
> just when the refinement program stops improving things?
>
>
>
> Jacob Keller
>
>
>
|