Training strategies in MLFF

[jun_yin2]

Dear all,

I have questions about training in Machine learning force field. If I want to explore more local reference configurations during temperature ramp from 300 K to 500 K, in two strategies, which training strategy is a better choice?
1. First train a force field from 300K to 400K by 10 ps with 0.5 fs per step, then continue training from 400K to 500K by 10 ps with 0.5 fs step.
2. First train a force field from 300K to 500K by 10 ps with 0.5 fs per step, then continue training from 300 K to 500K again with the same initial structure by 10 ps with 0.5 fs per step.

I hope you could help me. For now, I think the first may be a better choice. But I could not confirm.

[marie-therese.huebsch]

Hi,

Dividing long trajectories into smaller parts is always a good idea, i.e., use option 1. One reason is in an NpT ensemble, the volume would likely change during the simulation, and it is good to reinitialize the PAW basis for the electronic minimization.

But this won't significantly affect exploring more local reference configurations. Instead, you should look into parameters such as ML_SCLC_CTIFOR, ML_CX, ML_EPS_REG, etc.

Marie-Therese

[jun_yin2]

Very thanks to your reply!

[jun_yin2]

Hi,

I also want to ask about that whether ML_SCLC_CTIFOR can only be used in ML_MODE=select?

[ferenc_karsai]

ML_SCLC_CTIFOR is always applied when local reference configurations are selected.
That means for ML_MODE=TRAIN and ML_MODE=SELECT.