search

materialsproject / atomate2

atomate2 is a library of computational materials science workflows
https://materialsproject.github.io/atomate2/
Other
136 stars 76 forks source link

Extremely tight phonon k-point mesh #271

Closed utf closed 10 months ago

utf commented 1 year ago

Hi @JaGeo.

I just tried running the phonon workflow and noticed that the k-point density used is extremely high. It's set as grid_density=7000 which equals a reciprocal density of around 600 for silicon. Just for reference, the default reciprocal density in Atomate1 is 64 for insulators and 200 for metals. So this works out almost 10x the default k-point density.

For silicon, this means a supercell with lengths 22 Å is being run with a 2x2x2 k-point mesh!

I'd be very surprised if we need to go this high to get converged results. My feeling is that we should could use reciprocal_density=100 and still get converged results. From my previous experience, I usually run displaced supercells with the same k-point density as the relaxation and this normally works well. My experience is that it is usually the force and energy convergence that are essential to get good phonons.

What do you think about:

  1. Switching from grid_density to reciprocal_density, this means the number of k-points is independent of the number of atoms and only depends on the cell size.
  2. Dropping the density to reciprocal_density=100

Obviously, we should do some tests to make sure this gives reasonable results still.

JaGeo commented 1 year ago

Thanks for the suggestions!

Switching to reciprocal_density sounds very resonable and I actually follow a similar strategy when optimizing the phonons by hand. However, PbTe would potentially be a compound, for example, where one might need such a high number of kpoints, if I remember correctly.

Maybe, we should again do around 10 test computatations for smaller and larger compounds? I am not sure I will have time for it immediately. (I am currently again working on the Lobster schema. I realized several points I would like to improve with regard to usability of the outputs while thinking about a potential tutorial ...).

JaGeo commented 1 year ago

Overall, I agree that this setting is on the very tight side. I will talk with the people using the code in my group (currently with less tight k-point settings 😅) to make more tests. Potentially also more improvements on the schema are needed when would like go towards the quasi-harmonic approximation.