VASP 6.3 Machine learning

Question on input files/tags, interpreting output, etc.

Please check whether the answer to your question is given in the VASP online manual or has been discussed in this forum previously!

Moderators: Global Moderator, Moderator

Post Reply
Message
Author
suhong_zhong
Newbie
Newbie
Posts: 1
Joined: Tue Aug 03, 2021 3:58 am

VASP 6.3 Machine learning

#1 Post by suhong_zhong » Mon Feb 07, 2022 5:48 am

Hi,
Parameters in Machine learning, How do I set the "RANDOM_SEED" ?

ferenc_karsai
Global Moderator
Global Moderator
Posts: 422
Joined: Mon Nov 04, 2019 12:44 pm

Re: VASP 6.3 Machine learning

#2 Post by ferenc_karsai » Mon Feb 07, 2022 7:31 am

It's not clear for what you need the random seed for in the machine learning calculations, but I try to answer the possible reasons.
The machine learning module doesn't use any random number except for a small unimportant feature that is hidden from the users and therefore should be avoided.
If you need a random seed for the MD used together for on the fly learning then please use RANDOM_SEED:
wiki/index.php/RANDOM_SEED

A possible random seed looks like:
RANDOM_SEED = 248489752 0 0

It's a good habit to specify the random seed for your MD calculations.

bhubnesh
Newbie
Newbie
Posts: 5
Joined: Tue Jun 21, 2022 8:34 pm

on-the-fly learning

#3 Post by bhubnesh » Wed Aug 03, 2022 3:38 pm

When we train the machine on-the-fly-learning with molecular dynamics to generate the Force field, can we train them in any ensemble i.e. NVT, NVE or NPT, or do we have to follow particularly one of them? And also after doping our system with a new element, do we have to generate the new force field or we can use the force field already generated for the pure system?

ferenc_karsai
Global Moderator
Global Moderator
Posts: 422
Joined: Mon Nov 04, 2019 12:44 pm

Re: VASP 6.3 Machine learning

#4 Post by ferenc_karsai » Thu Aug 04, 2022 1:34 pm

In principle one can train in any ensemble, but the question is, does it make sense.

The NVE ensemble visits too little phase space so I would never train there.

The NVT is good for training, but if possible I would always train in the NpT ensemble, since training on the volume fluctuations add additional stability to the force field.
Important: if you use an NpT ensemble always use an ENCUT that is at least 30% higher than for NVT calculations. This is needed since the volume can increase and then one needs an increased basis.

Concerning the elements:
The force field is not transferable from one element to the other and you will need training on the new element too.

xiaoming_wang
Jr. Member
Jr. Member
Posts: 52
Joined: Tue Nov 12, 2019 4:34 am

Re: VASP 6.3 Machine learning

#5 Post by xiaoming_wang » Thu Aug 04, 2022 4:16 pm

Concerning the elements:
The force field is not transferable from one element to the other and you will need training on the new element too.
If for intrinsic defects (no new elements), do we need further training?

ferenc_karsai
Global Moderator
Global Moderator
Posts: 422
Joined: Mon Nov 04, 2019 12:44 pm

Re: VASP 6.3 Machine learning

#6 Post by ferenc_karsai » Mon Aug 08, 2022 8:20 am

I would say always yes. The formation energies of defects need as good accuracy as possible. The error in energy per atom comes from your training and will be the same in the production runs in larger structures. Since the total error in energy for a structure is the error in energy per atom times the number of atoms. So with increasing cell size the magnitude of error of the total energy becomes the same as the defect formation energy of a single vacancy and hence that size cannot give reliable results. So you need to get as accurate as possible in your training. That requires learning on defect structures too.

Post Reply