In general, the thermal factors are the last parameters you vary. It is best to start with background coeffcients and zero error or displacement and scale factor. Then introduce lattice parameters and crystallite size (peak shape parameters). Finally atomic coordinates (check if there are atoms in special positions) and thermal factors, but not together, first the coordinates then the thermal factors. Otherwise there are so many correlations between these parameters that no reliable values are obtained. You can also vary the occupancies, it depends on the crystal structure of your material. I suggest to use a tutorial, certainly present in the software, to be sure to proceed in a correct way

It is not so strange to find negative values for thermal factors. They can reflect problems of others parameters of the refinement. In general to find reliable values the angular range of the pattern should be sufficiently wide (up to 100-120°); also the occupancies of the sites can influence the U values. Do you have cationic substitutions? You can also try to use in the first instance the global thermal factor and fix the single thermal factors to small positive values.

The most likely culprit of your negative thermal parameters is the absorption correction. Make sure you have properly taken into account this correction based on sample size, geometry, x-ray energy, etc... The absorption correction and thermal parameters are highly coupled. As Marcella mentioned, it is best to fix the thermal parameters first. Once you have converged (from different directions) to the best fit of the other parameters such as atomic positions then you can fix these values and begin to fit the thermal parameters. If you don't have the proper absorption correction, though, fitting the thermal parameters will be an exercise in futility.

Thanks to @Marcella Bini and @Lawrence Margulies. I don't have any cationic substitution. I am at the starting stage of using GSAS. I kept Uiso = 0.025 .Can you please tell me how to fix the global thermal factor and single thermal factor? Is there any step by step way to do X-ray refinement using GSAS-EXPGUI?

In general, the thermal factors are the last parameters you vary. It is best to start with background coeffcients and zero error or displacement and scale factor. Then introduce lattice parameters and crystallite size (peak shape parameters). Finally atomic coordinates (check if there are atoms in special positions) and thermal factors, but not together, first the coordinates then the thermal factors. Otherwise there are so many correlations between these parameters that no reliable values are obtained. You can also vary the occupancies, it depends on the crystal structure of your material. I suggest to use a tutorial, certainly present in the software, to be sure to proceed in a correct way

Thank you very much @Marcella Bini

By my opinion, you must get only positive value for Uiso, but you can get negative value for U(anisotropic).

How can you can a negative thermal parameter? It doesn't matter whether it is isotropic or anisotropic, it still makes no physical sense. If you get a negative anisotropic thermal parameter then you have found a false local minimum.

Endorse the statements of the Professor Lawrence, a negative anisotropic thermal parameter is physically impossible. The problem has already been discussed here a few times and even by Professor Lawrence. The article attachement showed a critical review about crystal structure determination. 

Best regards.

Excellent article, Alan. Anyone doing x-ray crystallography should be aware of those points. There was an interesting round robin not too long ago in which a standard powder sample was sent to a number of different x-ray labs. The results were disturbing and point to exactly what that article discusses. The standard deviation of all the results on the lattice parameters were larger by a factor of 10 then the average error reported for each individual measurement! Once again a great reminder that just because you have a low chi squared or R factor or whatever parameter of choice you use to judge goodness of fit does not mean your fit has produced the correct results. Refining a structure using GSAS or any of the various refinement programs is not something that can be done correctly unless you actually understand the first principles of everything you are fitting. There is no algorithm for doing a refinement. If your sole criteria for determining the correctness of your results is the fitting error then you should not be using these programs.

Thank you by explanation Lawrence. Rietveld Refinement is ia very powerful tool in solving crystal structures, but subject to many interpretations. Make choices and correct judgments requires a little study of the theory behind the method. And I believe that here is a place where we can help resolve these questions and direct our colleagues to make improvements choices.

What type of Absorption function I can put for XRD refinement in EXPGUI?

Please see the website link below. There are several tutorials on the website, explains step by step how to do it.  

https://community.dur.ac.uk/john.evans/topas_workshop/pcg_workshop_menu.htm

I hope those tutorials help you to get a start.