I've got many questions again. I will follow the notation of Frank May:
>1. What type of computer hardware is required? Operating system?
I've developed the program using 32-bit OS/2. Until now, the program
is a native command line one. Using the Watcom Compiler (which is a cross
compiler) I'm able to bound executables for OS/2, WinNT (which are also
running under Win95) and for the dos4gw DOS extender. The latter, I may
distribute free with my programs (this is granted by the licence agreement
from Watcom). The program is also tested to be compiled with GNU, Borland
and Microsoft compilers and to run on an AIX Workstations from IBM, on latter with IBM and GNU Compilers. I recommend as a minimum a 486 CPU with 8 MB (DOS) or 16 MB (OS/2, Win95) main storage. Pentium would be better.
>2. Do you intend to sell the program, or will you give it away? (If for
>sale, how much?)
On this point of discussion, I want to thank the people from the
firm Seifert-FPM, Freiberg. Without their support, I won't have been
able developing and debugging the program.
I've joined the rietveld mailing list just at the time the Freiberg firm
starts to sell the program.
For all commercial requests, please ask the firm
Seifert-FPM
Dr. Thomas Taut
Am St. Niclas-Schacht 13
D-09596 Freiberg
Tel: +49 3731 781 271
Fax: +49 3731 781 265
e-mail: SEIFERT-FPM@t-online.de
Ask for the Rietveld-program BGMN, not for EFLECH! Latter is an older
program. By selling BGMN, the Freiberg firm is developing a manual.
This manual will (since a independent one) be better than a manual
written by me, the programs author.
For scientific discussions, ask me again.
>3. Does the program access a data base of known structures for starting
>models?
At now, no. Maybe, one will be developed in future.
>4. Does the program deconvolute peaks and use the information it finds
>for input to structure refinement/quantitative analysis?
The deconvolution program is EFLECH mentioned above. Since EFLECH and
BGMN use the same peak shape model, the deconvolution is done by
fitting the convolution of ideal shape and specimens function (e.g.
influenced by real structure).
>5. There are at least 6 functions which contribute to peak asymmetry and
>broadening. What algorithm(s) do you use for peak broadening?
The basic formula is very old: The observed pattern is a kind
of convolution
Lambda*G*P,
where Lambda is the spectra of the K-alpha-doublet, G is the influence of
all geometrics (divergencies, penetration depth...) and
P is the specimens function (caused by ideal and real structure).
(For example, this formula fails for primary monochromator.)
But, it has taken several years of trial and error between experiment and
algorithmic design, to develop a practical usable implementation of this
formula. Thousands of lines of source code deals with that question.
There are two main programs (beside BGMN/EFLECH) dealing with precomputations for fast final computations in BGMN (and EFLECH).
In addition of the "natural width" of the line, given by Lambda*G,
there is a broadening with up to 2 "real structure" parameters.
One stands for a Lorentzian (e.g. caused by grain size) and the other
for a more gaussian quadratic Lorentzian (e.g. caused by strain/stress).
>In my opinion there are several useful features which are missing from the
>commercial Rietveld programs. Perhaps yours has something I need.
I will give a short survey over some features:
-good precomputated profile shapes.
-texture compensations with 5 different models
-isotropic and/or anisotropic (in crystal coordinates) grain size
and strain/stress parameters
-isotropic and/or anisotropic Debye Waller Factors
-All anisotropies including texture are automatic corrected for symmetry of
the crystal/the Wyckoff point
-parametrized atomic substitution
-Another Kind of nonlinear constraints: nonlinear parameters.
One can define an ideal molecul or ideal structure element (e.g. a
tetrahedra), and then define parametrized shifts/rotations/torsions of
the molecul or of parts of it. Of course, also the ideal structur element
may be parametrized, e.g. with bond lengths.
On this way, there are much less parameters than by full parametrizing
and nonlinear constraints.
-possible upper and lower bounds for each parameter
-practical no restrictions: unlimited phase count, unlimited parameter
count, unlimited wyckoff count, unlimited count of measuring points.
-And, last but not least (I feel, this point should be natural; but it
doesn't for other programs, as I learned from users): The program
*always*converges*. Give 50 parameters for 5 phases, and dummy initial
values: The program converges, never blows up, and after some time it
finds a minima in R. Mostly the right in the first trial.
Best Regards
J"org Bergmann
bergmann@rcs1.urz.tu-dresden.de