Energy Deviation with DeepMD-Kit

[Maia Courtiel]

Dear i-PI community,

 

I am currently running PIMD simulations with i-PI (v3.2.0) and a deep neural network potential trained with DeepMD-kit (v3.1.3).

I have observed a discrepancy: the potential energy of a specific bead printed in i-PI's output (using potential(bead=0)) differs from the potential energy I get when performing a single-point (SP) calculation directly on the exact same frame extracted from the .xyz trajectory (using the same potential outside of i-PI).

Attached are input files, as well as an "energy.csv" file comparing the values obtained both ways (which highlights the difference).

I have the exact same problem with a totally different system (and so a different potential).

Could this be related to any kind of truncation (precision loss when i-PI writes the trajectory), an issue passed through the socket, or something else I might have missed?

 

Thank you in advance for your insights.

 

Maïa