Inconsistent potential energy between DFT-MD and SCF during NPT_F ensemble simulation

[Denan Li]

Dear CP2K developers and community,

I am running DFT-MD (NPT_F ensemble) simulations in CP2K (version 2024.3) and using a machine learning potential (Deep Potential, DP) to predict energies along the DFT-MD trajectory to validate my DP model. However, I observed inconsistencies between the potential energies calculated by DFT-MD, SCF, and DP, while no such discrepancies were found in NVE or NVT ensembles.

My key question:

In the NPT_F ensemble, some potential energies and forces differ significantly from those calculated by SCF. Is this an expected behavior, or could it be influenced by additional factors (e.g., barostat contributions)?

Any suggestions would be greatly appreciated. Thank you for your time and expertise!

- Denan LI

Below are my test results:

1) I have run a 10-ps long DFT-MD trajectory and compared the energies with those predicted by DP. The overall energy profiles show significant discrepancies (see Fig-1a). DP and DFT-MD energy trends in some regions align well, while they diverge completely in others (e.g., around frame 9600) (Fig-1b).

2) I selected two representative frames and performed OT calculation to obtain the SCF energy:

a) The last MD step (matching trends): DFT-MD and OT  energies agree closely.

b) The 9600th frame (divergent trends): DFT-MD and OT energies show a large dismatch.

c) Note: Both DFT-MD and OT calculations confirmed SCF convergence;

3) I further conducted ML-MD simulations in CP2K using the DeepMD-kit interface and compared the energies/forces between the ML-MD trajectories and DP predictions.

a) In NVE/NVT ensembles, forces/energies from ML-MD trajectory match will with DP re-evaulations:

b) In NPT_F ensemble, ML-MD energies/forces deviate from DP predictions, with errors showing periodic oscillations: