Diffpy - CMI used for RMC

[Sarah Hoback]

Hi,

I have been reading the CMI GitHub files, and I am still confused about what type of regressions are being used.I am unsure if you are able to do an RMC with an amorphous material using the CMI, or if there is only least squared fitting available. I am also unsure of how the iterations are being calculated, during the regression. As they seem to change regardless of if any modifications were made to my script. 

Thank you,

Sarah Hoback

[Simon Billinge]

Hi Sarah,

In principle we can plug in many different regression algorithms.  In practice we don't have many examples where we have done that.

When you say you want to do RMC we should be clear what you want to do.  I view RMC as a methodology that was named after a regression method.  Are you interested in the RMC methodology, or the simulated annealing regression method?  In principle CMI is flexible enough to be able to write an RMC methodology using it, though there are good RMC programs out there that you may want to just use.  On the other hand, CMI gives you huge flexibility to kind of build your own refinement program, so you could come up with your own structural representation and mathematical model and plug in a simulated annealing or parallel tempering algorithm to do the regression if you like.  So if there is something more adventurous like that that you are after doing, then I would say that CMI is a good choice.  But as I said, we don't have examples of that that you could work from.

If you do want to do something like that and want to reach out to us (off diffpy-users) then do so, and we can try and help you.

S

--
You received this message because you are subscribed to the Google Groups "diffpy-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to diffpy-users...@googlegroups.com.
To post to this group, send email to diffpy...@googlegroups.com.
Visit this group at https://groups.google.com/group/diffpy-users.
For more options, visit https://groups.google.com/d/optout.