GW unoccupied states have strange anomalies

[engelund]

The 'check bounds' flag did not help, unfortunately.
"
B=252 Eo= 6.57 E= 6.79 E-Eo= 0.22 Re(Z)=0.95 Im(Z)=-0.137750E-2 nlXC= -1.10386 lXC= -2.82827 So= -1.48863
B=253 Eo= 6.58 E= 6.79 E-Eo= 0.21 Re(Z)=0.94 Im(Z)=-0.216946E-2 nlXC= -1.38619 lXC= -3.45002 So= -1.83773
B=254 Eo= 6.60 E= 6.84 E-Eo= 0.24 Re(Z)=0.95 Im(Z)= 0.737372E-3 nlXC= -1.03810 lXC= -2.73317 So= -1.43820
B=255 Eo= 6.63 E= -4.53 E-Eo=-11.15 Re(Z)=0.72 Im(Z)=-0.168397E-3 nlXC= -30.7016 lXC= -19.3797 So= -4.16704
B=256 Eo= 6.66 E= -4.42 E-Eo=-11.08 Re(Z)=0.72 Im(Z)=-0.303378E-3 nlXC= -30.5368 lXC= -19.3115 So= -4.15021
B=257 Eo= 6.67 E= -4.43 E-Eo=-11.10 Re(Z)=0.72 Im(Z)=-0.631261E-3 nlXC= -30.5498 lXC= -19.3008 So= -4.14023
"
The jump occurs in the same place 254-255.

[Daniele Varsano]

Dear Mads,
indeed what you get it is very strange and not easy to rationalize.
In order to have a deep inspection, I suggest you do a simple and fast calculation calculating HF contribution *only* for the two bands where the jump occurs and inspect the HF and local matrix elements: nlXC and lXC.

Code: Select all

%QPkrange                        # [GW] QP generalized Kpoint/Band indices
 1|1|254|255|                                    
%

In this way, we can check if the problem is related to some index mismatch (even if I do not think so).
If the problem persists, so the problem should be traced to the wavefunctions itself and you can try to recalculate such unoccupied state with a stricter convergence threshold.

Best,
Daniele

[engelund]

Hi Daniele,
Thank you for pointing out a way to reproduce the error at a lot lower cost:)
<254|nlXC|254> = -1.041347 0.27219E-09 <254|lXC|254> = -2.733156 0.000000
<255|nlXC|255> = -29.39285 0.499217E-7 <255|lXC|255> = -19.37973 0.000000

When you talk about an improved convergence threshold, do you mean in Yambo or in QE? Could you give an example?

[Daniele Varsano]

Dear Mads,

as there's a large jump both in the Fock term and <Vxc> matrix elements, I suspect that wfs play the essential role.
My suggestion are:
1) Check convergence in the wfs in then nscf calculation, eventually increasing the threshold
2) Can you recalculate the ground state using a different pseudopotential, e.g. sg15 (http://www.quantum-simulation.org/potentials/sg15_oncv/) and
see if the problem persists.

Best,
Daniele