6
$\begingroup$

I'm trying to estimate some crude gamma-point phonon frequencies for a material with a large unit cell (>150 atoms). My goal is to verify some general trends (e.g., redshift) between 3-4 material variants, rather than achieving highly accurate values. Very crude and approximate results would be fine.

Realizing that the large size makes it unfeasible to do a proper phonon calculation (using supercells and high k-point resolution), I've instead attempted a very crap quality one with Quantum Espresso's ph.x, with minimal k-point sampling, a 1x1x1 "supercell" (our unit cell is already kind-of a 2x2x1 supercell), and loose convergence criteria. However, even this configuration proved too computationally demanding. There are almost a thousand modes to be considered, and we estimated that it would take nearly six months on our local cluster - for a calculation that has a good chance being meaninglessly poor.

Given these limitations, I’m wondering if there are more computationally efficient methods or alternative approaches that could give rough phonon trend estimates. Do you have any suggestions for software or methods that could estimate, even if roughly, gamma-point phonon frequencies of a bulk material with a large unit cell? Alternatively, a method estimating phonon density (over any region of the brilluin cell) is equally good.

Improve this question
$\endgroup$

1 Answer 1

Reset to default
3
$\begingroup$

This may be a good use for machine learning force fields. You can fine tune the MACE-MP foundational model on a bit of your data (geometry optimizations, smaller unit cells that are similar etc) then use it with ASE and phonopy to compute phonons. If you are using PBE or PBE-D3 you may not even need to fine tune (but may need to perform geometry optimization).

Improve this answer
$\endgroup$
2
  • 1
    $\begingroup$ This is actually one of the directions we are exploring currently! However, MACE itself is quite new, and the only citation using it for phonon I’m aware of (Lee et. al., in Arxiv) required 15k DFT calculations on structures with 100-200 atoms per unit cell each. Granted, their aim was an universal force field for phonons, but I wonder if we’d need ~thousand such calculations even if we restrict ourselves to a specific class of materials to be sufficiently accurate (i.e. avoid nonsense). Do you have experience with the minimal required number of structures for finetuning in this case? $\endgroup$
    – Neinstein
    Commented Nov 14 at 19:06
  • 1
    $\begingroup$ I am not sure, but analytical hessians are available and you could just use MACE-MP's geometry to avoid fine tuning at all. Please reply here if you ever publish something in this direction $\endgroup$
    – Tristan Maxson
    Commented Nov 15 at 18:57

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .