Both CrysFML and GSASII calculators failed to work with minimization

EasyScience / EasyDiffractionApp

Diffraction data analysis application

https://easydiffraction.org

BSD 3-Clause "New" or "Revised" License

4 stars 1 forks source link

Both CrysFML and GSASII calculators failed to work with minimization #120

Closed rozyczko closed 2 years ago

rozyczko commented 2 years ago

Issue 1:

AttributeError: 'Site' object has no attribute 'Uiso'

Issue 2 Additionally, changing a parameter value when CrysFML is chosen gives

'NoneType' object has no attribute 'strip'

rozyczko commented 2 years ago

Both of these issues are now addressed by workarounds: 1: https://github.com/easyScience/easyDiffractionLib/commit/19b97b866781e3564c87ff4b9e03150b99804a35 2: https://github.com/easyScience/easyCore/commit/f7ffdb85d68a5758180ca5a43116738989a680ad

However, these need to be properly understood and fixed.

AndrewSazonov commented 2 years ago

There are no more errors, but the minimisation with CrysFML or GSASII still doesn't work. And it remains broken even after switching back to CrysPy.

Choose a project example 3
Go to the Analysis page
Deselect parameters 28 and 29
hange parameter 1 from 11.8992 to e.g. 11.8
Press start fitting button -> the value will return to 11.8992
Select CrysFML
Change parameter 1 from 11.8992 to e.g. 11.8
Press start fitting button -> value remains 11.8
Select CrysPy
Press start fitting -> value remains 11.8

rozyczko commented 2 years ago

This has been fixed by implementing a larger numerical differentiation step for CrysFML and GSASII and changing the default method to least_square. This method allows the step to be adjusted. The problem is with the fact that CrysFML uses a 32 bit float to store cell parameters and it runs out of precision when applied the default, very small step by lmfit.

The correct fix is to have CrysFML use double precision values for cell parameters, but we should also be prepared to support calculators which can't be modified.