How can we determine the reliability of a machine learning force field (ML_FF) for use in production?

[dianwei_hou]

After obtaining a machine learning force field (ML_FF) using VASP, and deciding to use it for production, what factors should be considered during the production run? Additionally, what is an acceptable range of BEEF value fluctuation?

[mike_foster]

Read the documentation.
https://www.vasp.at/wiki/index.php/Best ... rce_fields

[jonathan_lahnsteiner2]

Dear dianwei_hou,

This is a very general question and therefore not so easy to answer.
In principle there are lot's of techniques to estimate errors of machine
learning approaches. A very common one is to use a test set error estimation.
To do this, you can run a molecular dynamics run under the conditions
(pressure, volume, temperature) you want to study your system. Then you can
extract some of the structures produced during the MD run and compute their
energy, forces and stress tensor with the DFT approach you trained your force field on.
Then you can estimate a test set error between force field and DFT values.

The size of error that you can tolerate during your production molecular
dynamics runs also strongly depends on the quantity you are trying to
predict.

In general it is always good to have the error of your force field as low as possible.
To minimize your errors, it is advisable to retrain your force field with the ML_ISTART=1 and optimize
your hyper-parameters after picking up reference structures.
This is described here https://www.vasp.at/wiki/index.php/ML_LFAST

Maybe also take a look at our best practices page to get more information:
https://www.vasp.at/wiki/index.php/ML_LFAST

There are also lot's of papers discussing the errors in machine learning
force fields. Maybe take a look at Ceriotti
https://arxiv.org/pdf/2011.08828.pdf
or Carla Verdi
https://www.nature.com/articles/s41524-021-00630-5.pdf
But you should also be able to find other references.

If you need further information, please contact us again and maybe try to
be more specific with what you exactly want to with your force-field.

I hope this helps.


All the best Jonathan