[sdpd] Re: F1 and Monte Carlo

[Armel Le Bail <alb... @cristal.org>]

Andrzej Grzechnik wrote:

>I used ENDEAVOUR for some structures that I already knew to see how it works
>and whether it finds the structure. It actually does the thing. Then I used
>it for some unknown cases for which I had the total number of atoms in the
>unit cell, lattice parameters, and bad data (peak positions + intensities).
>No indexing. No structure factors. The most "complicated" case was with 20
>atoms, tetragonal lattice parameters, and 46 reflexions. I got a  structure
>(four independent atoms) that I refined with GSAS. 

I had not seen that ENDEAVOUR could work from only peak positions +
intensities + composition, no indexing, no structure factor ?? I thought that
this was the domain of structure prediction by packing considerations,
knowing the molecular formula, for instance.

But for the tetragonal most "complicated" case, you had the cell, right ? 
Then I guess that the 20 atoms correspond to a description in P 1 space 
group ? In fact this is a 4 independent atoms structure, so that 46 reflections
appear sufficient for a refinement with GSAS. With ESPOIR, you may 
expect to solve problems as complex as 15-30 independent atoms : 4 times
more complex. Of course, you will need approximately 10 time more 
reflections than atoms (150-300 reflections). And they should be very good 
data... They may be obtained by structure factor extraction methods (Pawley
and so on), provided you select those reflections not too much biased
by overlapping.

In fact, the next ESPOIR version, which is already working on my
desktop computer, will be able to use the whole set of extracted
reflections, whatever they are overlapping or not. This is not a miracle.
The test for accepting or rejecting an event in the Reverse Monte
Carlo process will not be made on the "|Fobs|", but on a pseudo
powder pattern reconstituted from those extracted "|Fobs|". You
may say, why not using the true powder pattern ? For saving
time, which is the main problem of Monte Carlo... By using this
pseudo pattern, you will not have to take account of background,
assymetry, Lorentz polarization, zeropoint, multiplicity (...).
Compared to methods of molecules location based on simulated
annealing (...), the number of DoF - Degrees of Freedom -
attains 45 to 90 in ESPOIR, against less than 20 (positions +
torsion angles) in, for instance, the school cases treated by the
Powder Solve software.

>Anyways, the idea of not using any potentials is a good one, providing that
>the minimum distances between atoms can be changed depending on the
>conditions of data collection. In my calculations for high pressure
>structures, I usually set them up to values 10% to 20% lower than the ones
>observed at ambient conditions. 

In Espoir, the minimum distances and constraint coordinations
used in the version 0.9 were removed in version 1... All
appeared as if using minimum distances or not using them gave
the same result, with a very slight advantage to not using them !

I do not see why Endeavour could not succeed in problems
as complex as those solved by Espoir, since the Espoir code
is under GNU license (which does not exclude commercial
aspects, but force to let the source code open - but here may
be the problem ;-).

Coming back on Earth, you should realize that the number
of cases is small where Espoir or Endeavour will have a real 
advantage on the classical direct or Patterson methods. These
are the last chance methods, to try when everything else failed.
So that they also will have large chances to fail in those
specially difficult cases...

Best


Armel Le Bail - Universite du Maine, Laboratoire des Fluorures,
CNRS ESA 6010, Av. O. Messiaen, 72085 Le Mans Cedex 9, France
http://www.cristal.org/