My Community

I am in the midst of generating several MLFF for different materials and I am curious as to the best way to utilize these FF for simulating the structure of a solid upon cooling from a melt. In particular, I wonder about the wisdom of using only the force field without including the possibility of any ab-initio calculations. Is it possible to use Bayesian error to induce a ab-initio calculation when using the FF in a non-training mode. In particular, I mean the "threshold" status as indicated in the ML_LOGFILE (e.g. not learning, but occasionally doing a ab-initio calculation to improve the accuracy of the result. What is suggested?

Dear paulfons,

In the current version of vasp it is not possible to utilize ab-initio calculations during the inference mode (MO_MODE=run, non-training mode) of MLFF. You first have to finish on the fly training to be able to switch to inference mode.
So I would recommend to first train the MLFF in all phases which are important for you. And to then run the MLFF for doing your production runs. Maybe the best practice page is interesting for you.

All the best Jonathan