Why does i-pi seem to slow down after hundreds of steps

Ziqi Wang

unread,

Sep 4, 2025, 1:42:12 PMSep 4

to ipi-users

I am trying to simulate dp-trpmd with 1024 atoms and using a trained deepmd model as a client to provide the force. And every time I run i-pi, ipi will run a few hundred steps quickly, and then suddenly become very slow. I don't know why. I'm wondering if this is a problem with my setup or is it normal? My input file was attached, and the picture is the GPU occupancy when I perform deepmd as a client.

Any replies would be greatly appreciated.
Thank you for your time and consideration
Best regards.

input.xml

lah3.json

Michele Ceriotti

unread,

Sep 4, 2025, 2:02:40 PMSep 4

to ipi-users

Sorry for the silly question, but have you checked the trajectory? It might be blowing up or collapsing, and that would explain the sudden change in computational effort.

Michele

Ziqi Wang

unread,

Sep 4, 2025, 3:03:21 PMSep 4

to ipi-users

Thank you Professor Michele for your prompt response and your suggestions. I did make a silly mistake. It seems that it was caused by the too fast H motion. Do you think it is because my potential energy model is not well trained or because of other factors? If it is other factors, do you have any good solutions?

Thank you once again for your suggestions.

Sorry to might bother you twice, I chose the option "Reply to the author" at the first time

Ziqi Wang

unread,

Sep 4, 2025, 3:26:21 PMSep 4

to ipi-users

I have found the problem, it is a unit problem. The default unit of i-pi is bohr. The coordinates of my initial structure file are in angstroms. I forgot to specify the unit as angstroms in input.xml. However, I set the cell unit to angstroms, which caused the unit cell to be empty and expanded dozens of times. Sorry for my silly mistake. Thank you Professor Michele for your time and your suggestions

Best Regards

Ziqi Wang

Michele Ceriotti

unread,

Sep 4, 2025, 5:00:23 PMSep 4

to ipi-users

no problema. it's a common mistake and actually we recently allowed to use `ase` as input mode which reads xyz in angstrom (and allows providing the cell in extended xyz mode, that is often more convenient) precisely to reduce the occurrence of these units issues. but do ALWAYS check your output trajectory file, it solves 90% of the problems.