My Community

Hi,

I am running an VASP MLFF to characterize naphthalene molecular crystal. My training dataset has only samples from 295K while the testset has structures from 80K, 120K, 220K and 295K. By looking into the DFT energies and volume of my test set, I realized there is a Pulay stress (see the attached figure). On the other hand, MLFF also recover the same Pulay stress trend (see the attached figure).
My question is that as MLFF did only saw 295K structure, how it can demonstrate Pulay stress effect? I expect it to be a linear model. Is there any internal thing that is done to create such an effect? Can this be explained by the fact that VASP also learns the stress?

I would appreciate any help.

Regards,
Burak

Dear Burak,

The high temperature structural phase space is a superposition of distorted low temperature structures. So the machine learning model will recognise the low temperature structures, since it will sample some of them. However, there is no guarantee that this will work, since it is strongly depend on the material.

Regarding the model: a polynomial kernel is employed, which is definitely non-linear.

Finally, in prediction mode VASP will compute the stress tensor using the Virial theorem. Stress is computed from pair forces and you should be able to predict the right stress even for low-temperature structures if these were also explored together with high-temperature structures.

Kind regards,
Pedro

Dear Pedro,

thanks for your response. That is what I thought, but the training only sees one unit-cell which is 295K. It is true that the training set has low temperature structures, but the volume is the one of 295K as I set ISIF =2.
How can then MLFF predict the same Pulay stress effect as DFT by not seeing that volume, i.e. planewave set? Is there any chance there is an internal correction from VASP?

On the other hand, although VASP learn the stress, I can not be sure it will be enough. I now refit my ML_AB by setting stress weight to zero and can test this.

Regards,
Burak