Energy drift in MLFF MD simulations

[MCV_IFPEN]

Hello,

I have read reports in this forum about energy drifts using the Nose-Hoover (NH) thermostat in AIMD simulations (see for instance this link). What about simulations with MLFF?

I have this problem with a MLFF potential I've made following the best-practice steps in the wiki . Like other users running AIMD simulations with the NH thermostat: E0, T, EK and SK fluctuates around an average value, but there is a constant increasing drift in both E and SP.

People running AIMD simulations seem to solve this issue by increasing the precission of the wavefunction calculations (which is a bit surprising to me, as the issue seems to concern the thermostat rather than the electronic energy). What should I do in a MLFF simulation?

Thanks.
Best regards,
MCV

[jonathan_lahnsteiner2]

Dear MCV IFPEN,

Energy drifts during molecular dynamics simulations can have a lot of reasons. For example

• starting structure is not equilibrated.
• in ab-initio calculations forces might not be computed with sufficient accuracy
• too large time step
• bad machine learning force field
• thermostat settings
  and so on

The issue might not be directly related to the Nose-Hover thermostat but can occur with any thermostat. It is not possible to tell what issue causes the energy drift in your situation without supplying the used input files and the generated OSCIZAR file and ML_LOGFILE.

All the best Jonathan