Atoms moved out of a supercell during MD calculation by MLFF

[kai_oshiro]

Dear all,

Recently, I trained FF using a slab model of CeO2 based on the following paper.
URL: https://doi.org/10.3389/fchem.2020.601029

The calculation was completed normally, but CeO2 moved out of the supercell during MLFF.
It looks like a strange behavior, as if periodic boundary conditions (PBC) are not applied.

In a supporting information of the paper, atoms that have been displaced from the supercell appear again from the opposite side.
I think it is normal behavior of PBC, but my calculation was not.

Are there any mistakes in my setup of MD or MLFF?

The higher temperature in my INCAR was set according to the melting point of CeO2.

Best regards,
Kai

[martin.schlipf]

The absolute position in a crystal does not have a physical meaning in periodic boundaries. For historic reasons it appears that VASP folds all positions to the primitive cell for MDALGO = 0, whereas the atoms may leave the cell for MDALGO > 0. Both choices are valid and should lead to the same electronic structure. Depending on your goal (visualization, diffusion coefficient, ...) one or the other choice may be more appropriate.

[andreas.singraber]

Dear Kai,

Ferenc and I had a closer look at the machine learning log file and spotted some problems regarding your setup. Overall we have the impression that the system is heated up way too fast and too high. This seems to hinder the ML on-the-fly algorithm to capture the necessary local reference configurations and causes unsatisfactory training errors (quickly rising way above 100 meV/Angstrom for forces). Is it really your intention to reach 3000K, that would be above the melting temperature of CeO2? We would recommend to perform a heating run, starting at a low temperature and ramp up to a temperature still way below 3000K, e.g. setting

Code: Select all

TEBEG = 10
TEEND = 1000

Span this heating run over as many time steps as you can afford and monitor the Bayesian error estimate, the actual error and the error threshold along the trajectory (look for columns 4 in BEEF, ERR, and THRUPD lines in the ML_LOGFILE). Maybe this recently uploaded video can be helpful as well (look at the second half at around minute 46+ where actual simulations with machine learning are discussed). Only if you get a satisfactory ML potential for these lower temperatures you can proceed to higher temperatures by further heating up. The ML force field from the first run can be extended by setting ML_ISTART = 1.

Another issue is the quality of the electronic convergence. In the OSZICAR file one can spot multiple ab initio calculations which did not converge within 200 electronic steps. If converged and unconverged ab initio reference data is combined your training is based on an inconsistent potential energy surface. Please try to make sure that your ab initio calculations are always well converged. For example, switch to ALGO = Fast instead of VeryFast. Slowly heating up may also help in this context.

All the best,

Andreas Singraber

[kai_oshiro]

Dear Andreas Singraber and Martin Schlipf,

Sorry for replying so late.
I am currently confirming the calculation conditions.

Thank you for pointing that out.

Best regards,
Kai