Yambo Community Forum

Dear Haseeb,
yes, if you look at your log files you will see the calculation is allocating a huge amount of memory. Considering you already did the calculation in G space and presumably, you want to extract transitions for the excitation in the low energy part of the spectrum you can largely reduce the bands in the BSE setting a window across the Fermi energy.

Best,
Daniele

Dear Daniele, thank you for helping!

As you suggested, I now lower the band range close to the Fermi level. But even due to these settings, the RAM requirement is beyond my capacity (125 GB)! I even tried running on a single mpi process...

The new input file is attached, how can I tackle this?


Thanks again,

Dear Haseeb,
I see, the problem here is that you have a large number of k points (1944) in the whole BZ. This means that you have an extremely large matrix (dim=Nk*Nc*Nv), which makes the calculation unfeasible.
The only way to perform the calculation is to further reduce the number of bands (let's say 5 occupied and 5 unoccupied). Hopefully they are enough to look at the first excitations.
Best,
Daniele

Dear Daniele, Thanks very much, indeed now it is working for a total of 10 bands only!

But I have a couple of questions:

Q1: If I switch to IP mode in transition space then the calculation completes in a minute or two, even for the full BSEBands range (50-350)! Is there nothing to compute in the IPA approximation if we have a dipole database calculated in G-space?

Q2: What is the meaning of BSENGexx if we do the BSKmod=IP/Hartree absorption calculations, in the Hartree mode, I think it will work like NGsBlkXd but can we include this parameter in IP level also? I don't think it will make any sense in IPA! Am I right here?

Q3: It is very important for me to know, that let's say, I include the same number of bands in the absorption calculation, then should I get the identical spectra from IPA/Hartree in G-Space solving Dyson equation and IPA/Hartree from transition space using BSE?

Q4: What should be the acceptable damping value in absorption calculations? Default is 0.1 eV (uniform)? is it okay or should I decrease it to a very small value?
Thanking you,

Dear Haseeb,
I strongly invite you to have a look at the cheatsheet we prepared for the schools, here you can find the equation and the link to the yambo variables: in this way you can understand the different level of approximations you are adopting.
http://www.yambo-code.org/wiki/index.ph ... heatsheets

A1: In the IP approximation the kernel is zero, the response matrix is diagonal so there is nothing to compute to build the response matrix and diagonalize, just summing up dipoles matrix elements divided by energy differences. (slide 4)

A2: Meaning of BSENGexx, see Eq. 10b. In IP this is no used as the kernel of Eq. 10a is not taken into account

A3: The IP will be identical. In the Hartree approximation, it will be also identical if BSENGexx is set equal to NGsBlkXd. Note that the two calculations are identical if the response function are set both as "resonant". In the case you consider causal Green functions in the transition space you should also include the coupling term, which usually is very small.

A4: A damping of 0.1eV is usually a reasonable choice.

Best,
Daniele

Thanks, dear Daniele.

Dear Daniele, I was doing the RPA optics in the transition space, but my calculations is always stopping at the optics level (BS solvers, I guess) without producing any error! I have attached the log and setup files.



Regards,

Dear Haseeb,
you are trying to diagonlize a very large matrix (dim=58320), are you sure that the run stopped or just what happen is that the diagonalization of such matrix takes very large time?
Usually such large matrix are handled by using iterative methods (Haydock), but in this way you will not have access to the eigenvector.
What I suggest you is on of these options:
1) Reduce the matrix size (suggested!!)
2) Use parallel linear algebra (comple the code using scalapack libraries) and by setting BS_nCPU_LinAlg_DIAGO.
3) Use iterative methods providing the firsts eigenectors (compile the code with slepc library) and setting the proper parameters in input (not well documented).

Option 2 it is not guaranteed to work.

BTW: Are you sure that it is relevant to decompose excitations obtained at RPA level in transistion components? Note that in this approximations you do not get bound excitations. And, in any case in order to analyze the very first excitation very few bands should be enough to have a good description. Note also that the large number in the matrices are driven by the very large k-point sampling you are using, check if it is needed.

Daniele Varsano wrote: ↑Mon Nov 24, 2014 4:13 pm
When you change the states variable form 1-2, 1-3, 1-10, you are calculating the exciton amplitudes for the
excitons wit indexes 1 and 2 (you will have two files)
excitons with indexes 1,2 and 3 (you will have 3 files0
excitons with indexes 1,2, ....up to 10. (you will have 20 files)


Best,
Daniele

Dear Daniele,

I have a couple of questions regarding plotting exciton and understanding ypp output files.

1: I have done BSE with the energies 0-4.5 eV only, but in the o.exc_qpt1_E_sorted and o.exc_qpt1_I_sorted files, energies more than 4.5 eV is also there? How it is possible and what is it's meaning?


2: In the same output file, at the top, we have something like this: So, what these k-points correspond to? In a single file, I have more than 10 these k-points? So, should we extract the k-points of the max weight from this list or it will be extracted from the info given after this list of points, like Kv-q ibz, etc... (my next question)? And what is this iku unit in which k-points are reported?

3: I want to know the structure/meaning of the following:
Band_V Band_C Kv-q ibz Symm_kv Kc ibz Symm_kc Weight Energy

columns 1,2 and last are understandable, while the third column is k point or q point of valance band of the corresponding transition entry? And why not it is Kc-q ibz in the fifth column Similarly where will Symm_kv and Symm_kc can be used? And does the Weight max value is 1/normalized?

4: As a lot of users already asked about the meaning of states! And above you have also answered this question,

excitons with indexes 1,2, ....up to 10. (you will have 20 files)

I didn't get it, If I put the 1-10 states, should not I get the 10 files instead of 10?, I file for each excitonic state? I know the index will come from sorting the energy/intensity? So, does it index corresponds to # first 10 states or something else? As, I was doing 1-3 states, and I was obtaining 4 files? I was confused then!

5: I am attaching all the relevant files, in fact, the bandgap (electronic) of the material is 2.32 eV, so I wanted to collect all the major excitons before this energy? But can we also calculate the transitions after the bandgap, I mean the main peak comes at 4.2 eV so, by using ypp can I guess the states due to which this transition will be happening, although this is beyond the bandgap?

6: In my absorption curve, before the fundamental bandgap, the abs is not zero but there is not any peak either! Instead, absorption is continuously increasing from 0 eV. So, should I conclude I have dark excitons in the system? And I found that by looking at the RPA and converged BSE result, that BSE spectra peak is almost 1 eV before than that of RPA, which shows that binding energy of 1 eV? How can I guess the same from ypp results?

7: And lastly, I had planned to apply the scissor operator manually at the end by hand, I have applied the scissor corrections neither in the dielectric screening nor in the BSE! So, is this approach good or should I always apply the corrections in the input file?

8: (I have added this later), In the ypp -e s 1, the energy was sorted and I can see couple of excitations before the bandgap as follows: But if I do the states 1-1 in ypp -e a

I got the following result:
Here the Energy in the last column is clearly above the bandgap (2.3 eV) while in the energy sorted file was showing the transition index 1 corresponds to 2.002 eV!
I apologize for a lot of questions!!

Thanking you,

Dear Haseeb,

1. You have specified the interval of energy where the spectrum is plotted, in the o.exc* are reported the eigenvalues and intensities for the complete spectrum of your BSE independently to the energy interval the spectrum is plotted.

2. K points participating in the excitation, where the v->c transition at that k point are summed up. Details on single v->c transition are given below. iku is an internal units, you can discard it.

3. In the BSE you have the contributions for the entire BZ. The entire BZ is decomposed in k point in the IBZ (3rd and 5th column) and a symmetry operation applied on it (4th and 6th column). The symmetry is indicated with an index, the crresponding operation is found in the report where the symmetry operations of the system are listed. The weights are normalized to one, if you have degenerate excitons (up to a threshold you can control in input), they are summed up and then the normalization is the number of the degeneracy. If degenerate exciton are found it is reported in the report. If you set a degeneracy threshold equal to zero you consider just one exciton.

4. One file for each state, in the top of each file you have the legend: What do you have in the 4th file?

5. Yes, you can look at the contribution of any of the excitations, just insert the corresponding index.

6. Not really clear what you are asking here. The spectrum is built by considering Lorentzians around each excitation according to its intensity, if you have a dark exciton the intensity is almost zero and you will not see it in the spectrum. Binding energy can be calculated looking at the excitation energy (value reported by ypp) and the quasiparticle gap. You can calculate binding energy also for dark excitons that are not visible in the spectrum.

7. To be consistent scissor operator should be applied in the BSE only unless you want to apply a sort of eigenvalue self-consistency.

8. The energy in the last column is the KS energy difference (unless you added a scissor in ypp, I do not know what happen in that case) of each transition participating to the exciton The exciton instead is a result of a linear combination of all the transitions participating. You need to look at your QP gap and not that energies.

Best,
Daniele