[QE-users] [QE users] pseudopotential hardness and transferability

[Nicola Marzari]

Ciao Aldo,


a meaningful/useful test would be this - take a simple case of a CN
system - it can be as simple as an infinite 2-atom primitive cell of a 
graphene-like sheet (or a 2-atom primitive FCC cell (like GaAs)). 
Doesn't matter if it's not that physical. You also do not need fine 
k-point sampling. Shake one of the 2 atoms quite a bit - 10-15% of their 
typical bond lengths.

For this system, plot the total force on 1 atom, and the trace of the 
stress tensor, as a function of cutoff on the wavefunctions (30, 40, 50, 
60, 70, 80, 100 Ry) for a couple of duals (i.e. ecutrho six or 8 times 
larges than ecutwfc). If you feel good about doing it, look also at the 
frequencies of a phonon - maybe not Gamma, but at a zone boundary.

This will tell you what cutoffs you need for a structural relaxation.

Let me also repeat that the convergence on the total energy is not a 
necessary condition - you can have wonderful and accurate pseudos whose 
total energy is still changing at a cutoff that is good for forces and 
stresses.

			nicola


On 03/04/2020 10:43, Aldo Ugolotti wrote:
> Dear Nicola,
> 
> I share with you the perplexity about the unusual values for the 
> cutoffs, that's why I wished to understand the origin of it and whether 
> it could be improved or fixed (other than just setting such cutoff and 
> keep going). About checking the pseudo versus the total energy, it might 
> be not so relevant but indeed it is a necessary condition.
> 
> In my tests I just run scf calculations with different cutoffs/k-mesh 
> with fixed geometry and extract the total energy (i.e. fixing the 
> precision rather than the accuracy, in the language of Science, 351, 
> 2016, paper). I consider the minimal cutoff the one at which the total 
> energy is within ~ 1mRy from the value of the calculation with the 
> highest cutoff. I can also extract the corresponding total force or stress.
> 
> Finally, I definitely agree with you about the "hardness" of 
> pseudo-generation; the results I was reporting are from the checks on 
> the SSSP pseudos (or pslibrary v 0.1<->1.0.0 ones, in general, which 
> yield similar numbers), the only difference lies in taking into account 
> a system more complex that a single atom. In my particular case, I did 
> consider a small patch of graphene-like C,N compound.
> 
> Hence I wondered if there was any modification, given the possible 
> differences between the isolated atom and the small patch, to improve 
> the effectiveness of the pseudo (hence, not just hardness, but a proper 
> size of the dual). My first guess was to check the transferability of 
> the pseudo, which maybe was not enough to cover the electronic 
> configuration of the patch. That's what I could not test with ld1.x
> 
> Bests,
> 


-- 
----------------------------------------------------------------------
Prof Nicola Marzari, Chair of Theory and Simulation of Materials, EPFL
Director, National Centre for Competence in Research NCCR MARVEL, EPFL
http://theossrv1.epfl.ch/Main/Contact http://nccr-marvel.ch/en/project