convergence tests

[Stephan]

Hello,

I'm about to determine the dielectric function of a certain material with all available runlevels.
This mean
yambo -c which gives RPA (often called chi, but why?)
yambo -b which gives Bethe-Salpeter (Is this correct?)

and both of them with all available kernels IP, Hartree, ALDA, LRC, BSfxc

My question is for which parameters of the input file do I have to make convergence tests?
I know that I have to do it for NGsBlkXd in yampo -c for Hartree-kernel but not for IP.
Are there any more parameters I should check. (Besides BndsRnXd which is obvious)

Thanks and regards

Stephan

[Daniele Varsano]

Dear Stephan,
Careful:
yambo -c and yambo -b does not gives any optics, but they are meant for the treatment of coulomb interaction and static dielectric screening.
I imagine you mean:

Code: Select all

yambo -o c 
yambo -o b 

anyway they are not RPA and BSE, but they indicate the base you want to perform the calculations (g-space or transition space) as documented here.
What gives you different approximation is indeed the kernel "Hartree" for RPA and "SEX" for the BSE. BSE can be solved in transition space only.
Parameter to converge are k-points, bands, Gblk etc.. I invite you to have a look to the tutorial example. In Gspace or transition space the name of the variable is different as they enter in different way in the expression of the dielectric functions. For instance BndsRnXd in Gspace and BSEBands in transition space. Of course in BSE you need the screened interaction so the static RPA (calculated in Gspace) is also needed.
Note that BSE calculations need knowledge of quasi particle energies, so a GW calculation is needed or experimental input.

RPA (often called chi, but why?)

chi or \chi (greek letter) is the letter usually used in the literature for the response function \chi(r,r') in real space of \chi_GG' in reciprocal space.

Best,
Daniele

[Stephan]

Dear Daniele,
thank you for your help.
I have a few more questions:
1. You said a parameter to converge are the kpoints. I suppose you mean FFTGvecs. They can be reduced in order to accelerate the computations.
When I test convergence with respect to FFTGvecs in IP can I use the result also in the other runlevels like 'hartree' or do I have to repeat that
convergence-test for each runlevel?
2. I calculate the linear response for a bulk solid, so ALDA cannot give any improvements compared to hartree, right? Alternatively I can choose
yambo -o c -k lrc in order to introduce TDDFT. Do I have to look for convergence with respect to FxcGRLc in this case or is this variable just useful for ALDA?

[Stephan]

Sorry I forgot to add:

Best regards,
Stephan

[Daniele Varsano]

Dear Stephan,

1. You said a parameter to converge are the kpoints. I suppose you mean FFTGvecs.

Well, no. I really mean k points, as from the grid of k points it depends the accuracy in the Bz integrations etc (q=k-k'). In order to check it it is needed to perform different ground state calculations increasing the k-point sampling.

When I test convergence with respect to FFTGvecs in IP can I use the result also in the other runlevels like 'hartree' or do I have to repeat that
convergence-test for each runlevel?

Well, as the FFTGvecs enter essentially in the calculations of dipoles, oscillators etc. I think that the convergence can be seen already at IP level and it is not needed to repeat it in other runlevels.

2. I calculate the linear response for a bulk solid, so ALDA cannot give any improvements compared to hartree, right? Alternatively I can choose
yambo -o c -k lrc in order to introduce TDDFT. Do I have to look for convergence with respect to FxcGRLc in this case or is this variable just useful for ALDA?

For bulk systems ALDA does not give much improvements, it can be seen from the Dyson equation for X considering the limit of q->0 knowing that X goes to zero as q^2. I'm not very familiar with lrc, but as far as I know it is just \alpha/q^2. \alpha it is an external parameter so there is not FxcGRLc to test, but just an "empirical" alpha. In the case of ALDA yes, FxcGRLc it is an important parameter, but ALDA can be solved in transition space and ALL Gvec are automatically included as in that case the ALDA is evauluated in real space.

Best,
Daniele

[Stephan]

Thank you very much...