Replies: 2 comments 12 replies
|
Hi Gwyn, Good question — this comes up a lot. A few thoughts: All-electron vs. pseudopotential basis. In general, WIEN2k as an all-electron (L)APW code will give you a more accurate starting point: a more complete basis, proper treatment of core states, no pseudization. So at the DFT level WIEN2k+TRIQS is the "gold standard" reference, and for strongly hybridized systems that baseline accuracy does matter. That said — a carefully set up VASP calculation should match WIEN2k closely. The key is to use POTCARs with as much valence as possible: the Beyond that, the velocity (momentum) matrix elements that enter the transport kernel are always evaluated on a finite k-mesh, regardless of the code, and the results have to be converged with respect to that mesh. This convergence is tedious and is often the dominant source of discrepancy — more so than the basis itself. So before comparing codes, make sure each calculation is individually k-converged. For the VASP+W90 path specifically, everything rests on the Wannier Hamiltonian reproducing the KS eigenvalues inside your energy window. It's worth verifying this explicitly — the band interpolation should lie essentially on top of the original KS bands in the window of interest. If the disentanglement/projection is poor, everything downstream (including velocities) inherits that error. One word of caution: the Wannier route only captures transitions between the bands you wannierized. Transport is dominated by the low-energy states around the Fermi level, so this is usually not a severe restriction — but if relevant transitions reach to states outside your window, those are simply missing. So just make sure your window is generous enough to cover the energy range that actually contributes to the transport quantity you care about, rather than only the narrow correlated subspace. The flip side is that, once you have a faithful H(R), you can interpolate to H(k) on an arbitrarily dense grid very cheaply and precisely. This makes the k-convergence above much more tractable than re-running expensive all-electron optics on denser meshes — and it's really the superpower of the Wannier interface: the interpolation is flexible and trivially convergeable, so it's a feature rather than an accuracy liability. Recommendations:
Bottom line: VASP+W90+TRIQS can be perfectly adequate for transport provided you (a) use hard enough potentials, (b) confirm the Wannierization is faithful, and (c) include sufficient bands. The interpolation is a feature, not a liability. Best, |
5 replies
|
Hello Alex, Thank you for your thorough response. I have one question about enabling site symmetry. When I analyze the I have considered post-processing the symmetrization, but that is just a one time fix and would not translate well to a CSC calculation. Eventually I will enforce symmetry for self energies in dmft, but I wonder if this H(r) is not too clean to begin with, adding some room for error. Please let me know your thoughts, Gwyn |
|
I think the question really is how bad the symmetry breaking is. As far as my testing goes, important is the 2nd to last column, which is the residual of the eigenvalues between iterations. This will tell you how well the wave functions are converged. This should be at least 1e-6 before you run the wannierization or locproj. Is this the case (I typically set Then also make sure that the disentanglement achieved very high precision. If you can share some input and output files I can have a look. Best, |
|
Hello Alex, I apologize for my delayed response. I have been trying different energy windows and orbitals for projection to get the cleanest overlap between wannier bands and DFT bands. I wanted to ideally choose the s,p and d as projection orbitals for hcp iron (following your suggestion to keep the energy window large enough so that the relevant wavefunctions for transport calculations will be included), but that gave me a lot of oscillations when plotting the band overlap. I believe it is due to higher energy from the conduction band being mixed in. I get a better overlap when I only project to the s and d orbitals. Before running wannier90, I ran a VASP SCF step that shows the output: Then for wannier90 this is what my output shows: I used these specifications for my .win file: When I analyze my hr.dat file I see the break in symmetry.. for example, from the D3h symmetry for hcp I expect my energy levels to have the degeneracy of 1,2,2, but instead I see: when I try to convert wannier to dmft, it also gives an error of
I am also attaching an image of my band overlap for the s and d orbital and the DFT orbitals. I want to add that the reason why I kept num_iter as 0 is so that I could just project to the orbitals, and specify the d orbital as correlated in the dmft step after wannierization. I could just be completely missing something leading to this asymmetry. Please let me know if you need any other details. Thank you. Gwyn |
|
Hi @gwynchoi, the band overlay plot is very telling. You can see the oszillationsat K and H for example? This is a strong sign of disentanglement not working properly. Also I think you highlighted the frozen window but even in the frozen window the band energies do not match everywhere which is not good. You did set thanks! |
|
Hi Alex, I am attaching the wannier90.wout file. The disentanglement criteria was not satisfied, even at higher Thank you! |
|
Yeah one can see that the Wannier functions 1 and 7 are not good at all. Their spread is enormous. The convergence of DFT eigenvalues is okay now but something is not right. The fact that the disentanglement does not converge is not good. Let's try to reduce the outer window from 40.22 to 30.22 . Also maybe for the Wannier step you can do Best, |
7 replies
|
Would it be possible for me to email the files to you? |
|
yes certainly my github profile reveals my mail address. Use that one. |
|
Hi Gwyn, I figured it out. The problem is that the then I get (32 bands in VASP are sufficient): that should fix your issues. I somehow did not think about this before that the s-orbitals hybridize here strongly and to get the character correctly we need the antiboding character like bands. Let me know if that gives for you similar results. Here is a plot of the band structure now without the strong oszillations: I think you cannot make this work without using something like SCDM / SVD for the initial projection. I would not use the wannier hamiltonian without localization. The s-orbitals need the MLWF algo. If you only need the d orbitals just drop the s projections. That will work now. Best regards, |
|
Hello Alex, With the settings you specified, I was able to reproduce your results. I really appreciate your help. Sincerely, Gwyn |
|
I am not sure I fully understand your question here. In principle if run through solid_dmft in a CSC calculation in each VASP (or whatever DFT code you are using) the projection <Xi_m|Psi_n> are calculated again in every step. That is necessary to convergence also the DFT charge density rho together with the local Green function. This is all done automatically for you. No matter if you use LOCPROJ (PLOs) or wannier functions via the w90 interface. If you do want to use w90 for solid_dmft please read: https://triqs.github.io/solid_dmft/latest/md_notes/vasp_csc.html#enabling-csc-calculations-with-wannier90-projectors this section. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Hello everyone,
I am calculating CSC DFT+DMFT and postprocessing to get transport properties using VASP6.5.x+Wannier90+triqs. I was comparing this methodology to WIEN2k + triqs, which is the more standard way of calculating transport properties.
I was wondering if, even with careful selection of energy windows to capture the correlated subspace, the WIEN2k +triqs methodology would still yield much more accurate transport properties for highly correlated and hybridized systems due to some approximations of VASP6.5.x+Wannier90+triqs, such as how the velocity matrices are computed.
I would like to stick to VASP as I am used to it for DFT, but of course, accuracy is paramount. What are your thoughts on this, and are there any ways you would recommend for me in this case for more accurate results?
Thank you.
Gwyn
All reactions