On Sat, 19 Feb 2005 20:49:01 -0800 (PST), Leonid Solovyov wrote: >On Sat, 19 Feb Joerg Bergmann wrote: > >> We cannot follow this suggestion. We have checked about 10 >> examples of (partially complicated, 10 phases) quant analysis. >> Of course, the R values are larger, for example: >> Rietveld: Rexp=6% Rwp 8% >> DDM: Rexp=25% Rwp 30% > >I already wrote that DDM-R values can not be DIRECTLY compared with >Rietveld-R values. Moreover, the absolute value of R-factor can not be >used as a self-sufficient criterion of the quality of results for both >Rietveld and DDM methods. >What formulae did you use when calculating DDM-R-values? Did you use >equation 7 from the DDM-article (J. Appl. Cryst. (2004) 743) or another >variant? Did you calculate Rietveld-R values after the background >subtraction or not?
BGMN serves 4 different R values: R, Rp, Rpb, Rwp plus Rexp. The DDM variant does not contain background, so only R, Rp, Rwp plus Rexp are calulated. Further see bottom. >Even disregarding the above questions, you have got the ratio >DDM-Rwp/Rexp=1.2 smaller than the Rietveld-Rwp/Rexp=1.33. > >> The errors following statistical error propagation are somewhat >> larger than for normal Rietveld. But, trying standard >> mixtures with known content, the differences between known >> and measured weight contents are at least twice larger. > >I did not understand - with which method did you get twice larger than >known contents: with Rietveld or with DDM? with DDM the deviations between known and estimated content are twice larger compared to Rietveld. > >> Maybe, DDM is of some advantage in structure refinement. >> In quant analysis, we cannot recomment DDM, at least >> with BGMNs fundamental parameter profiles. > >I do not see any principal difference between the structural and >quant-analysis in terms of profile-fitting. see bottom. > >> Maybe, it has some advantages for older Rietveld programs with >> classical UVW profile shapes. As we have observed, >> the DDM method is insensitive for wrong "tails" on the >> peaks, they become background. These tails are a problem >> with common UVW etc. profile shapes in Rietveld analysis. > >With this I can partly agree, except that the programs using UVW are >old. > >One more important point is that DDM results depend on the choice of >the convolution intervals for calculating derivatives. As far as I >heard from Joerg Bergmann, in the current version of BGMN the DDM >intervals are fixed to the minimal value of 5 data-points. This is, >definitely, not the best choice - it may work well only in rare cases. >In the DDM-article (J. Appl. Cryst. (2004) 743) I recommended flexible >intervals chosen on the basis of the counting statistics. This variant >is used in the DDM program and it gives most optimal results for >various datasets. I worry about "filtering out" some below-1%-phases while using DDM, so I decided to use only 5 data points. In structure refinement, you have always lots of intense peaks. Having in mind quant analysis, I have decided to use 5 data points only (an old rule says "at lest 5 data points on the peak width at half maximum", which corresponds to L. Solovjov's article). Using the adaptive method as described by L. Solovjov, minor phases (with partially low symmetry) may be "filtered out". They could be estimated wrong: Large ranges with lots of minor peaks may be detected as "linear" or "quadratic" and covered with only one single large range with a lot of data points. Dealing with a constant m=2 (only 5 data points), I do _not_ use equation 7, in such cases the Rwp values would change. Regards, Joerg Bergmann *** This EComStation system uptime is 1 days 9 hours ***
