My Community

Hi,

When performing AIMD runs with on-the-fly MLFF enabled, some SCF loops may not converge for complex systems, such as those with spin or when using meta-GGA functionals such as SCAN (even with very large NELM values).

How are such non-converged SCF loops handled during the on-the-fly MLFF generation?
Obviously, we don't want the training data to contain "bad" labels from non-converged calculations.
I assume that if the SCF loop doesn't converge then the labeled structure isn't added to the training data?
But I couldn't find any information about this when I read the wiki

Dear Dankomaister,

The machine-learning algorithm in vasp can not decide if a supplied
structure is converged in the sense of electronic minimization.
So if the electronic SCF calculation does not converge it still will be taken
for the machine learning data-base.
I understand that you do not want to have those structures in your machine-learning
data-base. The only way to prevent the machine-learning algorithm from picking
up non-converged electronic SCF structures is to adjust your SCF configuration.
If you need help with the settings of your electronic minimization
I would recommend you to send the INCAR, POTCAR, KPOINTS file. And a POSCAR containing an example structure for which the electronic minimization fails to converge.
I hope this is of help.

All the best Jonathan

Hi Jonathan,

Okay that is unfortunate, perhaps this is a feature worth implementing in a later releases of VASP?
Since there is already some logic in there for determining if the SCF loop is converged, it should not be that difficult.

I don't really have a particular calculation to share.
But some of the examples of this that what we have encountered in our group is people using on-the-fly MLFF for magnetic systems.
For example Fe clusters, where a few AIMD timesteps would not converge lets say 1 out of every 100 (the rest of the steps converge fine).
However this "bad" data is still ideal in the dataset which is not ideal.

Dear Dankomaister,

Yes I am aware that this is currently not implemented and would not be so complicated to do. The machine learning part is meant to be usable also without vasp in the future, this is the reason why this is not implemented.
Nevertheless, as a workaround for your problem I would recommend to chop up the training into several small pieces. Let's say, you do 100 MD steps or maybe less. Then you check in the OUTCAR file if there was an electronic SCF which did not converge. If so throw away the ML_ABN file and re-run the MD (you should not use a random seed). Otherwise, if all SCF steps converged you copy the ML_ABN to ML_AB and continue training with the data collected in the ML_AB. Like this you can obtain a data-base containing only converged data. This should be doable in a bash script.
I hope this is of help.


All the best
Jonathan

Hmm feels kinda hacky but I suppose that could work.