AIMD parameter selection with MLFF

[dominic_varghese]

Hi everyone,

I am generating training data for an on-the-fly MLFF (ML force field) using ab initio MD for a KTaO₃ system. The goal is to sample a sufficiently broad configuration space while maintaining stable and physically meaningful dynamics.

In this context, I would appreciate guidance on choosing appropriate MD parameters. For example, I am currently using values such as:

LANGEVIN_GAMMA = 10.0 10.0 10.0
LANGEVIN_GAMMA_L = 10.0
PMASS = 100
POTIM = 2.0

My questions are:

Is there a recommended strategy or physical criterion for selecting these parameters when the primary aim is to efficiently span configuration space for MLFF training?

How sensitive is the quality/diversity of the training data to the choice of friction coefficients (LANGEVIN_GAMMA) and time step (POTIM)?

Are there best practices (or typical parameter ranges) that people follow for on-the-fly MLFF generation in oxide perovskites or similar systems?

Any advice or references would be greatly appreciated.

Thanks
Dominic

[dominic_varghese]

Hi,

Also a general doubt. Is it okay to use Gaussian smearing instead of tetrahedron smearing for insulators in NPT/NVT runs? I was getting some errors when the tetrahedron method was used.

Thanks
Dominic

[ferenc_karsai]

Results should be independent of the friction coefficients (LANGEVIN_GAMMA, LANGEVIN_GAMMA_L) as long as you don't choose them too "extreme". To small coupling coefficients can lead to very long relaxation times and too strong coupling can dampen your system too much, resulting in strongly distorted trajectories.

Empirically we have found that 5-10 is a good value for LANGEVIN_GAMMA and LANGEVIN_GAMMA_L. PMASS acts similarly as LANGEVIN_GAMMA_L and values between 100 and 1000 are in most cases fine combined with the mentioned values for LANGEVIN_GAMMA_L.

POTIM depends on the weight of your elements. Typically for heavier elements it can be increased. Since Oxygen is the lightest element you have a time step of 2 ps is most likely safe. Maybe you could increase it even more, but that needs to be tested.

If you are unsure test all these parameters (POTIM, LANGEVIN_GAMMA etc.) once you have a force field that is not completely bad. For that run few tens of picoseconds with varyiing parameters and observe how important quantities like e.g. the pair-correlation function changes.