MLFF descriptor cutoffs

[mike_pols]

Hi all,

I'm working with the MLFFs as implemented in VASP, and currently running hyperparameter optimizations to optimize the performance and accuracy of the models I trained. To do so, I was following the tutorial as found on the VASP website: https://www.vasp.at/tutorials/latest/mlff/part1/

In following this tutorial, I had a question. In the examples, the cutoff values for the descriptors are optimized (i.e. ML_RCUT1 and ML_RCUT2), going from 13 A to 15 A. However, the cells in the training data are relatively small (dimensions of approx. 8 A). As a result of this, the descriptors also include interactions between periodic copies of an atom, or does the MLFF implementation of VASP work around this? Are there any guides to what descriptor cutoffs one should maximally use to train an MLFF with VASP?

Kind regards,
Mike

[ferenc_karsai]

The tutorial example is quite an exaggeration, 13-15 Angstrom cutoff for the two body descriptor barely ever make sense. As you have observed yourself if the cutoff radius is bigger than the cell parameters it's not a good choice. So this is the guideline for the maximum size of the descriptors. Also large radii make the calculations noticably slower. The default of 8 Angstrom for the two-body and 5 Angstrom for the three-body interaction is in most cases a good choice. In some cases it makes sense to optimize these parameters, e.g. if molecules on surfaces or defects when the molecules/defects are further apart from each other than the default parameters. In that case you need to increase the descriptors. Sometimes it makes also sense to decrease the number of the descriptors. For example in liquid water RCUT1=6 and RCUT2=4 Angstrom give the best results.