Anyway thanks a lot for all the help, hopefully I will not bother for some time now…

]]>So surely, you can initialize new BSE calculations on top of the same GW output. For this you need only to copy EVALQP.OUT (binary) and to have the <gw> element present in input.xml. To disable running GW, one needs simply replace: taskname="skip"

Regarding the list of "know issues", we have it internally and try so solve it as soon as possible ;-)

]]>I'll try to do it by hand it is not a big deal, however I will have to think more about this how to select the best offsets so that the combined sampling of the BZ is reasonable. I may need to read something more on the double-grid technique, any pointers?

However this brings me back to a question from the third post:

Is is possible to do xc calculations on top of existing GW without redoing it?

e.g. something equivalent to do="skip" parameter for the groundstate?

This is now even more urgent if I'm about to do the multiple xc calculation on top of single GW calc.

BTW is there some list of known issues or something similar?

I looked at the github but the repo there seems to be quite outdated.

I'm glad to hear that you have no problems with GW part :-)

Regarding the BSE part… Unfortunately, as we have recently discovered, there is a problem in the implemented version of the double-grid technique.

However, you can still proceed with this technique by doing it manually.

For this, e.g., to double k-grid, you should choose 4 k-point offsets and perform 4 corresponding BSE calculations. And in the end to make the spectrum averaging. We did it for another system and I wonder to repeat for Si.

Best regards,

Dmitrii

thanks for the advice, I played little with the other parameters and there is not much change, however with all those improvements with the nbands and others combined, the direct Gamma-Gamma gap is now 3.24eV, which should be quite comparable with experiments…

Now, looking at the comparison of my BSE spectra with 0K exp data the first peak position is only about 0.1eV below the experimental one, the second peak maximum is little bit worse, bit over 0.2 eV below the experimental, however I'm quite happy with this ;-)

The remaining discrepancies could be caused by the 10x10x10 grid in the BSE (since it still seems not so well converged).

My last remaining problem is with the double-grid technique for the xs. Can the ngridksub parameter just be added into the calculations as is or are there some other changes needed? E.g. if I add ngridksub="2 2 2" to the BSE element in the example input file from the first post, the calculations seem to run fine (looking at the INFOXS the BSE calculation is repeated three time), however the resulting averaged dielectric function is wrong. It actually looks more like it should be with the RPA or IP bsetype, any ideas what I did wrong or what should I do to fix this? (even adding ngridksub="2 2 2" into the example LiF calculation produces completely bad results…)

Anyway thank a lot for all the help.

]]>Do I understand correctly, that your major concern is to reproduce the experimental band gap in Si?

There are many things to mention prior going into technical details.

What is the value you're aiming at? There are many experimental as well as the theoretical results.

You can find in the literature the hole spectrum of G0W0 band gaps depending on the input structure and implementation.

Correct me if I'm wrong, the value you have reported before is the indirect band gap between X1(0,0.3,0.3)(CBM) [in the reciprocal unitcell units] and Gamma(VBM). This quite close to the X* point. However, since you're doing optics, I'd rather recommend you to look at the convergence of the direct band gap instead.

Before playing with the basis sets and fighting for the ultimate numerical convergence, I'd still recommend you to check the parameters which remained hidden (default). What about to expand the <gw> element:

<gw

taskname="g0w0"

ngridq="6 6 6"

nempty="300"

ibgw="1" nbgw="22"

>

<mixbasis

lmaxmb="3"

epsmb="1.0d-4"

gmb="1.0d0"

></mixbasis>

<barecoul

barcevtol="0.1"

></barecoul>

<freqgrid

nomeg="32"

freqmax="1.d0"

></freqgrid>

<selfenergy

actype="pade"

><selfenergy/>

<scrcoul

scrtype="rpa"

></scrcoul>

</gw>

The grid 6x6x6 should already be quite good, especially for tests. The product basis parameter could be an issue for some materials, but not in 'simple' Si. The number of frequencies is rather important and should not be overlooked.

In the posts above, you were correctly mentioning that the total number of the available empty states is restricted by the chosen in <groundstate> rgkmax parameter (Gmax). This number you should be able to see in KPOINT.OUT file, last column. So to reach high values of nempty, one has to modify rgkmax and recalculate groundstate cycle.

Thus, in your case for rgkmax=8, 300 empty states should already exceed the maximum number of the augmented plane waves.

You have also correctly mentioned that to reach numerically converged results it's often not enough to operate with the well converged groundstate (starting-point dependence), the number of states and the size of k/q-grids. One should extra take care about the complimentary basis-set incompleteness error. Similarly to pseudopotential method, for the excited state calculations one needs to employ specially optimized LAPW+LO basis (choose another species file). However, it has been shown a few times, that the basis set incompleteness is rather small for Si. So the reason for the inconsistent results one should look somewhere else.

The my suggestion is to modify the gw element and repeat calculations. The values you should ~1.21 eV [Gamma-X] and 3.21 eV [Gamma-Gamma] (depends on the lattice parameters).

Please let me know about your findings.

With best regards,

D. Nabok

(the exciting developers team)

nempty=100 : GAP 1.2150 eV

nempty=200 : GAP 1.2441 eV

nempty=300 : GAP 1.2537 eV

nempty=500 : GAP 1.2607 eV ]]>

3) I also had a look at the article "D. Nabok, A. Gulans, and C. Draxl, "Accurate all-electron G0W0 quasiparticle energies employing the full-potential augmented plane-wave method", Phys. Rev. B 94, 035118 (2016).", and the crucial stuff seems to be (beside the number of empty bands) also the usage of local orbitals for unoccupied states and the maximum angular momentum for such orbitals, how can I select this in the calculations?

]]>thanks for the advice, I'm trying it now, just few more questions.

1) The nempty parameter (in both groundstate and gw) seems to be truncated somehow, to the value of "Minimal number of LAPW states"

how can I get more LAPW states, do I need to increase the basis (rgkmax) or something else?

2) Is is possible to do xc calculations on top of existing GW without redoing it?

e.g. something equivalent to do="skip" parameter for the groundstate?

3) I did some reading of the

I also had a look at the article "D. Nabok, A. Gulans, and C. Draxl, "Accurate all-electron G0W0 quasiparticle energies employing the full-potential augmented plane-wave method", Phys. Rev. B 94, 035118 (2016)." which also article, and the crucial stuff seems to be (beside the number of occupied bands) also the usage of local orbitals for unoccupied states and the maximum angular momentum for such orbitals, how can I select this in the calculations?

I think your problem is related to the small number of empty states that you are including in the *GW* calculation. 100 is really not enough for comparing your results with published data. The exact number of empty states needs to be converged carefully and depends on the specific system, but typically several hundreds of states are needed to achieve reliable results. I suggest you to focus on the *GW* step since indeed the parameters for the solution of the BSE seem reasonable to me.

I hope this helps!

Best regards,

Caterina Cocchi

(` exciting` team)

while learning exciting, I've tried to do simple c-Si calculations with G0W0 and BSE and I have increased the numerical parameters hoping to get a good convergence and agreement with experiment. However despite all my efforts I still get the QP energies ~0.5eV underestimated (the shape of the dielectric function seems almost correct now so the BSE parameters should be fine). This is surprising as some reports show that it should be possible to get much better agreement. For example Rohlfing and Louie 1 reported much nicer agreement (almost perfect at least with respect to the QP shift of G0W0) that mine with the same method (G0W0 on top of LDA and BSE). The only difference I can see is that they used the plasmon pole model while I use the default RPA for W (which should be in theory even better?).

My input.xml bellow, do I still need to increase some parameters to get better QP energies or where can the problem be? Any ideas?

` ``<input> <title>cSi-BSE</title> <structure speciespath="/home/ondracka/exciting/species/"> <crystal scale="10.262"> <basevect>0.5 0.5 0.0</basevect> <basevect>0.5 0.0 0.5</basevect> <basevect>0.0 0.5 0.5</basevect> </crystal> <species speciesfile="Si.xml"> <atom coord="0.0000 0.0000 0.0000" /> </species> <species speciesfile="Si.xml"> <atom coord="0.25000 0.25000 0.25000" /> </species> </structure> <groundstate ngridk="20 20 20" gmaxvr="14.0" nempty="100" rgkmax="8.0" do="fromscratch"/> <gw taskname="g0w0" ngridq="10 10 10" nempty="100" ibgw="1" nbgw="22"> </gw> <xs xstype="BSE" ngridk="10 10 10" vkloff="0.097 0.273 0.493" ngridq="10 10 10" nempty="30" gqmax="3.0" broad="0.002" scissor="0.0" tevout="true"> <energywindow intv="0.0 1.0" points="2400"/> <screening do="fromscratch" screentype="full" nempty="100"/> <BSE bsetype="singlet" nstlbse="1 4 1 7"/> <qpointset> <qpoint>0.0 0.0 0.0</qpoint> </qpointset> </xs> </input>`

1) 1. Rohlfing, M. & Louie, S. G. Electron-hole excitations and optical spectra from first principles. Phys. Rev. B 62, 4927–4944 (2000).

]]>