rreho / yambopy

Python scripts to manage and postprocess quantum espresso and yambo calculation.
Other
0 stars 0 forks source link

Fixing TB dipoles, cpot and development of polarizability. #4

Open rreho opened 4 months ago

rreho commented 4 months ago

This issue is meant to be a thread for discussion about the calculation of the Coulomb potential within the analytical model approach. This problem concerns also the class TB_dipoles. @stevenbos123 recently worked on the issue via a series of commits. The main one is this one 574ca03 . I agree with the changes done in get_eps in wann_H2P.py. Regarding the class TB_dipoles, I find the vectorization of the arrays useful to improve speed-up and reduce the time spent on the double for loop over transition indices.

Lastly. I tested the absorption plot. The shape of real and imaginary part of eps are reasonable, while the magnited (1e-12) is totally wrong. I suspect units. TODO

rreho commented 3 months ago

In devel-tbwannier-modelcpot I re-implemented from scratch the class TB_dipoles. This is the relevant commit cd0b6f6 . I tested it and the code runs with a model Coulomb potentials. Note that I want to keep that k and q grid. It looks a bit useless, but from our point of view this code should compute the excitonic overlap matrix and therefore I need to have these two independent grids.

TODO @stevenbos123

rreho commented 3 months ago

@stevenbos123 spotted a bug where all the F_kcv are the same. Action needed.

stevenbos123 commented 3 months ago

related to the bug where all the 'F_kcv' are the same I pushed this change: #622d8b0 . 'F_kcv' should be oscillator strength of k,c,v for each transition, n. Instead it was oscillator strength of k,c,v for the transition that accompanies this k,c,v. Hence the values where all the same for each element of 'F_kcv'.

rreho commented 3 months ago

The bug spotted above is real. The fix works. The code runs up to get_eps. The values of eps are not "small" anymore and the shape looks reasonable.

TODO

stevenbos123 commented 3 months ago

Benchmark of WS2 6x6x1 kgrid. Comparing our get_eps() implementation to calculate the IPA absorption, with the values acquired from wantibexos. The coulomb potential used is the V2DT2 model potential, with values: v0 = 0.00 lc = 8.0 w = 0.0 r0=1.0 eta=0.08

The plot is normalized to the max of the respective values. From wantibexos the imaginary part of the sp_diel_xx.dat file is plotted which is the single particle dielectric function \epsilon. image

As can be seen in the plot, the second peak position coincides with the result from wantibexos, but the first peak position is not.

To investigate this further I plotted the F_kcv both from our implementation and from wantibexos sp_optics.dat again renormalized: image

stevenbos123 commented 3 months ago

When comparing our results from the dielectric function, with those with wantibexos, I noticed one fundamental difference between the two implementations. I will explain the difference by using the example of setting nc=1 and nv=1 in our yambopy code and doing the same for Wantibexos by specifying NBANDSC= 1 NBANDSV= 1. The way wantibexos uses this input variables is that the first band up or down from the fermi energy is considered to be the valence or conduction band used for the calculation. Now the tricky part is, this is not the same band for each kpoint, so the index of the conduction band might chance for the kpoints. This of course physically makes sense as you are interested in the minimal band gap for each kpoint. Now this is where our implementation differs. In our implementation the band index does not change for the k-points. In other words the bands are not sorted for each k-point on energy. So we do not consider the minimal gap possible for each k-point. This difference of course becomes less of a problem when you consider more bands.

Now a question to ask ourselves, should we keep our implementation the way it is, or copy the wantibexos way? In this case the yambo transition table now does not make sense anymore.

stevenbos123 commented 3 months ago

When comparing our results from the dielectric function, with those with wantibexos, I noticed one fundamental difference between the two implementations. I will explain the difference by using the example of setting nc=1 and nv=1 in our yambopy code and doing the same for Wantibexos by specifying NBANDSC= 1 NBANDSV= 1. The way wantibexos uses this input variables is that the first band up or down from the fermi energy is considered to be the valence or conduction band used for the calculation. Now the tricky part is, this is not the same band for each kpoint, so the index of the conduction band might chance for the kpoints. This of course physically makes sense as you are interested in the minimal band gap for each kpoint. Now this is where our implementation differs. In our implementation the band index does not change for the k-points. In other words the bands are not sorted for each k-point on energy. So we do not consider the minimal gap possible for each k-point. This difference of course becomes less of a problem when you consider more bands.

Now a question to ask ourselves, should we keep our implementation the way it is, or copy the wantibexos way? In this case the yambo transition table now does not make sense anymore.

I looked at the eigenvalues from h2p and model, and they are both sorted per k-point. This implies we compute the same thing as wantibexos does, just our band indexing does not change with it.