Slow performance of NequIP potential in LAMMPS

[wangxuan1586]

Dear NequIP developers,

Hello. I have trained a NequIP potential that performs quite well. The training process took about four days. However, when I use this potential to simulate the growth of large-size materials on a substrate, with around 400 atoms or more, the simulation speed is very slow. It takes several hours to complete the deposition process for a single atom.

Are there any ways to improve the simulation speed?

Also, what is a reasonable size for the training set when training a NequIP potential? At the moment, I am using a training set of about 180,000 frames, some of which come from structural optimization data. Is the training set too large, and could this be the main reason why the trained potential runs slowly in LAMMPS?

Thank you very much for your help.

[cw-tan]

Are you using GPUs? See https://nequip.readthedocs.io/en/latest/guide/accelerations/openequivariance.html , to be used with https://nequip.readthedocs.io/en/latest/integrations/lammps/mliap.html

Inference speed is only depends on the model details, it's independent of training set size (unless of course you need a bigger model to do better for a bigger, diverse dataset).

[wangxuan1586]

Thank you for your reply!

I am currently not using GPUs and have been running the NequIP potential in LAMMPS on CPUs. I will try the methods you mentioned, including OpenEquivariance and the ML-IAP LAMMPS integration.

Thanks also for clarifying that the inference speed depends on the model details rather than the training set size.