Density of states (DOS) calculation with HSE06 not coming to an end

[Nishat Tasnim Hiramony]

Dear all,

I am calculating the DOS of a system using HSE06 functional. The calculation converges but it does not end. It gets stuck at "final diagonalization" and does not stop thereafter. The input, minimal output and submitted job files are attached here.

dos_io.zip

I did HSE06 DOS calculations for slightly smaller systems earlier and had no such issues. Is it possible that the code takes too long to write the wavefunction since the system is large? In the INCAR file that I have attached here, you will find ALGO = D ; LDIAG = .TRUE., LWAVE = F. However, I have also tried the following combinations:

1. ALGO = D ; LDIAG = .TRUE., LWAVE = T
2. ALGO = D ; LDIAG = .FALSE., LWAVE = F
3. ALGO = Normal ; LDIAG = .TRUE., LWAVE = F
4. ALGO = All ; LDIAG = .TRUE., LWAVE = F

For 3. ALGO = Normal, the convergence is faster but again, the calculation does not end and continued to run when I left it that way for 1.5 more days after reaching the convergence criterion. Could someone kindly suggest what I can do about this? And I know that the structure looks odd, I'll look into that, but apart from that, is there anything else I can do to solve this case?

By the way, I left out the POTCAR, should I add it here as well?

[martin.schlipf]

Did you run the same calculation with PBE? If writing the results is the issue, we would expect to see the same problem with PBE. If you do not see the issue for smaller systems you could see if there is a critical size of the system where it does not work anymore and check if at this stage the size of an array is comparable to the memory of your system. You also used a very dense grid to sample the DOS, perhaps a smaller NEDOS could reduce the load. Similarly if you do not need m-resolved DOS, you could try to set LORBIT = 10 instead of 11 to reduce the output further.

I also noticed that you used the old GPU port. Please be aware that this is deprecated and the OpenACC one should be used instead. I would also advise to look into KPOINTS_OPT, which is a new way to produce results in a non-scf calculation. You may need to set HFRCUT for this to work properly, though.

[Nishat Tasnim Hiramony]

Thank you for those suggestions. I will first retry with smaller NEDOS and LORBIT = 10 and get back here. I did scf earlier with PBE which had no problem converging but I have not tried DOS with PBE yet. If the first thing does not work, I will look into this.

I have also done the calculations with a cpubase instead of gpubase, but the same thing happens there as well. But I will try your recommendations later on since I will have to ask someone else to rebuild vasp on my system since I don't have the permission to do so myself.

[martin.schlipf]

You may want to try PBE first, just to save computational time. It is often a good idea to start a hybrid functional calculation from a PBE WAVECAR to reduce the steps to solution, so it would not be wasted time.

[Nishat Tasnim Hiramony]

I did start from a PBE (scf) WAVECAR. Do you recommend starting from a PBE dos WAVECAR instead?

[martin.schlipf]

No that is not necessary. Just keep in mind that HSE is an orbital dependent functional. When you restart it is not sufficient to have just the charge-density, your k-point mesh also needs to be a regular mesh in the Brillouin zone. This is particularly important for band structure calculations.