Phonon Calculation Roadmap Without Phonopy Dependency?

[stefanbringuier]

Curious what the roadmap or intention is for implementing phonon calculations without the Phonopy dependency in the example script? I gave this a shot by implementing the finite-displacement method using ASE as a reference, but performing all calculations in PyTorch. It worked (though it was unexpectedly super sensitive to atom displacements), but the Phonopy-based script was much more performant, so I abandoned the feature.

That said, my implementation doesn't offer much value in the grand scheme of things imo, as it I think it makes more sense to try and leverage autograd to compute more accurate harmonic force constants and higher-order terms. Wondering if this is what the team envisions?

[abhijeetgangan]

• We are planning include a dedicated module for phonons. It will most likely still use the phonopy API. Most expensive calculations is the model call.
• Some model with either single or mixed precision can have varying performance. It can degrade at small displacements.
• Were you performing same number of displacements as Phonopy? If you do all 3N displacements for a crystal with symmetry phonopy would be faster. Either ways we don't have intentions to rewrite phonopy in PyTorch.
• Autograd for force constants is possible if it is implemented in the models. I haven't tested but it should be possible with MACE. We don't plan to modify the models ourselves to implement it.