Does delta machine learning have been implemented?

[min_gao]

Dear Developer and users

Does delta machine learning have been implemented in the latest version?
If yes, is it possible to share an example for that?

Best wishes,
Min

[andreas.singraber]

Dear Min!

Good point, thanks for bringing that up! I have not performed delta-ML calculations myself but as far as I know all computational parts necessary are available in the latest version of VASP. However, it may be necessary to write some scripts to extract or combine data. The steps for delta machine learning are outlined in detail in this paper (see also the supplementary information):

https://journals.aps.org/prb/abstract/1 ... 05.L060102 (or, on arXiv: https://arxiv.org/abs/2109.06282)

I will suggest to my colleagues to upload a description of a workflow for delta machine learning to our Wiki and report back...

Best,
Andreas Singraber

[andreas.singraber]

Dear Min!

I discussed this issue with my colleagues. Unfortunately, we do not have a workflow ready which we can immediately post on our Wiki. While delta learning did work very well in the paper I mentioned it is still an experimental technique and has not been applied directly by members of the VASP team. However, we agreed that it should be documented in the future and we would like to generate a tutorial which may then presented in our next workshop (presumably in September). I am aware that this does not help you now, sorry that we cannot give you more at this point. Please try to follow the individual steps described in the publication and do not hesitate to ask details about how to set up these calculations here in this forum!

Best,
Andreas Singraber

[prasad_mvd]

Dear MLFF Developers,

Please suggest how to select the subset of structures from training data set (T(2) as referred in DOI: 10.1103/PhysRevB.105.L060102) using CUR algorithm. I need this carefully chosen set of structures to generate Delta-MLFF using HSE forces and energies.

Regards,
Prasad Matukumilli

[andreas.singraber]

Dear Prasad Matukumilli,

I did not participate in the work on this paper, so please interpret my answer as suggestion only. If you need a more in-depth answer please contact the authors of the article.

I would set up the T(2) data set as input ML_AB in a separate run of VASP with

Code: Select all

ML_MODE = select

and tune the ML_EPS_LOW parameter in such way that the desired reduction is achieved. In the paper, the original T(2) set was compressed from 1275 to 168 structures. I assume that this was done by applying the "select" mode and then picking only the structures from the resulting ML_ABN file for which local reference configurations were selected. This requires some scripting work, because these structures are not directly extracted from the data set by VASP. However, you can find the structure indices in the ML_ABN file in the first column of the "Basis set for ..." section. For example, let's assume your T(2) set looks like this

Code: Select all

...
**************************************************
     The number of configurations
--------------------------------------------------
          1275
...

**************************************************
     The numbers of basis sets per atom type
--------------------------------------------------
        3212
**************************************************
     Basis set for H
--------------------------------------------------
          1      1
          1      2
          1      6
          2      1
          2     17
         29     13
...

Here you can see that structures 1, 2, 29,... contain local reference structures, i.e., atoms selected as reference environments. In this examples these atoms are:

Code: Select all

structure 1: atom 1
structure 1: atom 2
structure 1: atom 6
structure 2: atom 1
structure 2: atom 17
structure 29: atom 13
...

After running with ML_MODE = select you should find a reduced set of local reference configurations in the ML_ABN file, e.g.:

Code: Select all

...
**************************************************
     The number of configurations
--------------------------------------------------
          1275
...

**************************************************
     The numbers of basis sets per atom type
--------------------------------------------------
        523
**************************************************
     Basis set for H
--------------------------------------------------
          1      6
         64     23
...

Now, only structures 1, 64,... contain reference atoms whereas structures 2 and 29 were removed by CUR selection. So the task at this point is to extract the configurations with the given indices and convert them into the POSCAR format with a script. Then, run your low- and high-level ab initio calculations and construct the T(3) data set.

Hope this helps, if there are further questions please open a new topic!

Best,
Andreas Singraber

[prasad_mvd]

Thank you very much, Andreas Singraber for the detailed work-flow!

Regards,
Prasad