Application of potential to structures of n+1 atoms

[nikita...@aggiemail.usu.edu]

Dear all, 

Is it possible to use aenet network to predict energies for structures having n+1 atoms? I mean, for example, I have pretty big dataset (energies and geometries) of clusters consisting of 20 equivalent atoms. Can I apply extracted potential for systems build of 21 atoms? Can also train network on clusters consisting of different number of atoms? 

If I am only interested in energies, can I just omit forces input or set it, for instance, to 0? 

I am sorry for probably very naive questions. I am a complete novice in NN. I would appreciate any help. 

Thank you, 

Nikita. 

Nikita Fedik
PhD student
Department of Chemistry and Biochemistry,
Utah State University

Maeser Lab 287

+1-435-5126637

[Nongnuch Artrith]

Dear Nikita,

Thanks for your email.  Yes, Aenet uses high-dimensional ANN potentials that are not atom-number specific.  Essentially, an ANN potential predicts the energy of a single atom, and the total energy is the sum of all atomic energies.

Note: If your training set only contains clusters with N atoms, you still have to make sure that the prediction for N+1 atoms (or any other number of atoms) give reasonable results. It could be required to generate additional reference structures.

Best,

Nong

[nikita...@aggiemail.usu.edu]

Dear  Professor Artrith, 

Thank you for your time and reply. I also have sets from 1 to n atoms. Could you advice me if should I feed them to the network all together specifying path to all structures in one input file? 

And another one question: are components of force vectors necessary to build the potential?  We calculated only energies. Can we use only energies if are interested only in them? If yes, should I just omit columns 4-6 in input files or write zeroes, for example? 

Thank you again for all your help, 

Nikita. 

 

[Nongnuch Artrith]

Dear Nikita,

Yes, I would suggest to use all of your structures with 1 to n atoms together to train one ANN potential.

If you do not have force information, you have to fill the columns 4-6 in the XSF files with 0.00  0.00  0.00.  The forces are not used for the potential fit (but this will be an option in the next aenet release).

Example:

# total energy = -34.57516771 eV

CRYSTAL

PRIMVEC

        4.09780934    -0.00000099    -0.09460307

       -0.00000045     4.03823477    -0.00000246

        0.11043487    -0.00000244     5.17229730

PRIMCOORD

8 1

Cu      1.07966999     1.00955702     2.56249736    0.00   0.00   0.00  

Cu      1.02445171     3.02867571    -0.02365250    0.00   0.00   0.00     

Cu      3.07335725     1.00955839    -0.07095344    0.00   0.00   0.00    

Cu      3.12857369     3.02867420     2.51519451    0.00   0.00   0.00    

O       4.18063370     2.01683724     3.78462001    0.00   0.00   0.00     

O       2.07651506     0.00227610     1.24577218    0.00   0.00   0.00   

O       2.13172875     4.03595528     3.83191980    0.00   0.00   0.00    

O       0.02761002     2.02139390     1.29307187    0.00   0.00   0.00   

Please let us know if you have any further questions.

Best,

Nong

[nikita...@aggiemail.usu.edu]

Professor Artrith, 

Thank you very much for your detailed and very helpful answers. 

How should input .xsf files look like for module predict.x? I suppose the structure of file is completely the same, but what should I do with the line with energy value?

Thank you, 

Nikita. 

[Nongnuch Artrith]

Nikita,

‘predict.x’ does not expect energy or forces in the xsf files.  So you can just omit the line with the energy and the columns with the forces.  Energy and forces are output of predict.x. 

However, 'predict.x' can also read XSF files that already contain the energy and the atomic forces, for example if you want to benchmark a potential.

All the best,
Nong

P.S.  Just to let you know, I do not have professor title :-)