[QE-users] vc-relax for orthorhombic systems

[Davide Ceresoli]

> to everybody:
> 3 days in a row I was testing everything that my imagination allowed to and came 
> up to a conclusion that with soft materials, with orthorhombic cells of > 20 
> atoms there are TOO many degrees of freedom to expect vc-relax to give good 
> numbers for cell parameters.
> Within the same third decimal digit in total energy (-xxxx.xxX Ry) one might 
> came to slightly(?!) different cells depending on starting points and general 
> logic of relaxation (say, you start from orthorhombic space group doing relax 
> and for a corresponding ibrav continue with vc-relax Or you start from the 
> closest tetragonal space group and continue with xyz vc-relax).
> I will also do a few variable-cell NEB (coupled with QE for ab initio part) 
> between the orthorhombic and tetragonal systems, maybe it will help me more to 
> better understand the problem.
> 
> PS I have a full right to be wrong I am learning and will be thankful for any 
> feedback ;-)
Dear Alex,
     we made this experiment: we created 10 initial configuration by randomizing
the lattice spacings by 5% and we performed very tight vc-relax-ations.
Then we took the average lattice spacings and angles with their standard
deviations. For a typical orthorhombic perovskite with 20 atoms, we got
+/-0.007 angstrom accuracy on the lattice, +/-0.05° on the angles and
+/-0.0003 Ry on the energy.

For a given XC functional, this is the accuracy due to the numerics.
We are doing worse than PXRD experiments! If your system is softer, I
would expect even larger standard deviations.

HTH.

Best,
     Davide