Improvement of preconvergence step lobster workflow

[materialsproject]

Improvement of preconvergence step lobster workflow

[utf]

I have a couple of thoughts having looked more closely at the lobster workflow:

1. I wonder if the need for many more steps is because NBANDS is only set after the preconvergence job? If you could do the pre-convergence with the same higher NBANDS that might make things faster since it would have converged all necessary bands?
2. Can you use a non-scf uniform input set for the LobsterStaticMaker? Obviously this won't work for hybrid calculations but that should allow you to go to very dense k-point meshes without much slowdown. If we did this, we should make the preconvergence step compulsory. In fact, an even better approach would be to replace both the preconverge_static_maker and lobster_static_maker with a single maker lobster_uniform_maker that is actually a flow maker. E.g., some variation of UniformBandStructureMaker but with symmetry disabled for the second job.
3. Can we switch to using reciprocal_density instead of grid_density for the LobsterStaticMaker?

[JaGeo]

Strategy 2 seems to work now:

51 s with smaller number of reciprocal points (reciprocal_density: 100) and without symmetry (50s), uniform band structure computation with wavecar writing (reciprocal_density: 400, 150 s)

400 s with only one static computation (400 s, reciproal_density: 400)

I will now try to confirm that the Lobster outputs also still look fine based on this.

[JaGeo]

1. Yes, this is also my current assumption. I have to test it in detail for some cases once our cluster is back up. I hope it will lead to additional speed-up.
2. I haven't performed Lobster computations in this way. This would also be something we need to check.
3. Sure. I think it is much better

[JaGeo]

I now changed NBANDS in the preconverge step:
For my test case Silicon on 144 cores:

• Without preconverge step, the static run takes 112.338 s
• Preconverge step: 9.674 s, 94.304 s. Copying the WAVECAR also takes a bit *of time so that the speed-up is still not large but exists at least. (btw, maybe linking could be an option as well)
• Preconvergence step (same number of kpoints as in static run): 17.306, static run: 85 s. This is again slightly faster.

I will now make a test for a larger but very symmetric structure.

SrTiO3:
preconv (same number of kpoints): 81.158 s, static: 220 s
static alone: 274 s

This is now slower. I would probably need to test more systematically...

Maybe, we should really go for the lobster_uniform_maker but this would take a longer time to implement. I naively thought that the number of iterations should be much lower when starting from a pre-converged wavefunction. I still get more than 10 in each of the cases.

[JaGeo]

It looks very similar. Just VASP sets the band gap differently in VASP 6 and in this new maker ISPIN is set to 1 (i.e., plot 1 shows both spin channels but they are on top of each other, plot 2 shows the non-spinpolarized computation).

I also compared ICOHPLIST.lobster and again the agreement is very good.

[utf]

Nice! That looks promising.

By the way, I think if you upgrade to the latest VASP, they fixed the issue with setting the Fermi level so it goes back to how it used to be.

[JaGeo]

That's great. I guess we will then compile a new version soon!

[JaGeo]

I will now create new test data and fix the tests and also check if all folders in the output schema have been included.

[JaGeo]

I will close it as the mentioned pull request has been merged.