occupancy close to zero

[pascal22p]

When the occupancy get close to zero the refinement is extremely unstable

[richardicooper]

Possible solutions:

1. Spot problem, report and reset back to 'good' L5.
2. Stabilize matrix inversion (risks hiding the problem).

[pascal22p]

can't do 2. I think the inversion is stable but the answer is non physical.

It should probably be in a bigger scheme on diverging shifts, when this happens the adps go wild. Maybe something similar to what was there before.

• check xyz shift so that they don't move more than 20% of the unit cell
• check new adps Ueq so that is no bigger than something (Ueq could be too expensive)
• check occupancy, twin fractions and other parameters in a similar way

Revert to previous step when an issue is detected

[richardicooper]

The derivatives for all the atomic parameters tend to zero as its occupancy tends to zero, so whole rows and columns of the normal matrix will be approx zero. This should show up in the condition number, and eigenvalues of the matrix.

Would be good to have some sensible checks on shifts as they are applied. We did try code that limited shifts as a % of the value, but in some cases (i.e. poor starting model) a big shift is what you want., esp. for Uij and scale parameters.