GW unoccupied states have strange anomalies

[engelund]

Concerning using the box cut-off, nothing changed - the jump even occurs in the same place 254/255. I also tried it without cut-off entirely with the same result.

I looked at the wave-functions before and after the shift, but I cannot really see anything problematic (perhaps more trained eyes can so I attached the images).

[Davide Sangalli]

Ok. Then I suspect something strange is happening, for example some index going out of bounds, or similar.

If you are using gfortran, try recompiling running the configure with the following options

Code: Select all

FCFLAGS="-O3  -mtune=native -fcheck=bounds"

"-fcheck=bounds" is the relevant one.

you can just run the HF part of the simulation

Code: Select all

yambo -x

I can also try to reproduce the issue, but I need the pseudo-potentials. Can you provide them?

Best,
D.

[engelund]

Hi Davide,

Here are the exact pseudos I used if you wish to do some testing. It is just the "stringent" versions of the dojo pseudos (which for these elements are the same as the standard ones I think).

Testing the alternative compilation may take a while for me since I need to go through IT to get an as similar as possible compilation.

[Daniele Varsano]

Dear Mads,

Pseudodojo pseudopotential do have non-linear core correction. Can you try to repeat your calculation (just yambo -x) to look at the <Vxc> value after adding in your input file the keyword:

Code: Select all

UseNLCC

In this way the NLCC contribution is included in the Vxc term which otherwise discarded.

Best,
Daniele

[engelund]

Hi Davide,
Answering to an earlier request, we use the "ifort" compiler for Yambo, is there an equivalent compilation flag to try for this compiler?

[engelund]

Hi Daniele,
I have attached the test that you suggested. I don't see any Vxc in the output, though. Where would I find that?
Anyway, the jump in values still occurs.
I expected that since changing the pseudos completely did not resolve the issue.

[Daniele Varsano]

Dear Mads,

the <vxc> values are in the report and indicated as <nlXC|n>, i.e. local xc. In the same lines, you find also <n|nlXC|n> which is the non-local part, i.e. Fock's elements. And yes, the problem persists.

Best,

Daniele

[engelund]

Dear Davide,
I think you missed my question: How should I modify the compilation for Yambo when using the ifort compiler?

I will make a test with the gfortran compiler as well, but if there is a way to avoid such a large departure I would prefer it so that I can more easily zero in on the problem.

[engelund]

It has been a while and I have not made much progress.

Did anyone try this calculation?
Can anyone tell me how I should modify the compilation to potentially fix the "out-of-bounds" problem with an ifort compiler?

[Daniele Varsano]

Dear Mads,

the flag to check bounds in intel should be:

Code: Select all

-check bounds 

to be added in the FCFLAGS variable.

Best,
Daniele