fingerprint values

[Alejandro Rivero]

Dear Dr. Artrith,

I am currently using aenet to describe the O2-graphite system, i obtained very good results using Chebyshev basis set, but testing with different Behler basis set the results are not so good. I tracked the values of the FP using:

trnset2ASCII.x --raw file.train OUT.txt

The OUT.txt file contain all the information needed, i suppose that the SF values are order in the same position that its appear in the atom.fingerprint.stp, so, i don't understand why for a given structure with these features:

----------------------------------------------------------------------

                Adding structures to the training set

----------------------------------------------------------------------

#     N nAtoms  nTypes      E/atom structure file (xsf)

     1 56       2 -9.16756866E+00  ../structures/cfg-1043.xsf

     1 O     37

     2 O     28

     3 C     76

     4 C     76

     5 C     76

     6 C     77

where the O atom have several C neighboring atoms, the value of some of the radial SF for the O atoms is zero, and the values of non-zero radial SF match with the value of the SF if only another O atom be present in the rcut sphere, given my FP setup:

20

 8.76868890933318545E-01  8.42086699797561788E-01 7.74093779581538355E-01  6.49737499461713996E-01 0.00000000000000000E+00 0.00000000000000000E+00  0.00000000000000000E+00 0.00000000000000000E+00 6.89485081570356262E-01 1.01209674481052295E-01  4.85904544242246994E-01 1.15060723057684635E-03 3.34070711968992562E+00 2.12569996035534370E+00 8.73187364613008388E-01  1.60027506000764985E-01 9.31238545877092982E-01 8.41920073923692591E+00 6.90856628519202157E-03 6.37733517930427762E+00

I used (rc=6.5 A, rs=0) and a Behler basis set with 8 radial functions with etas (0.024, 0.049, 0.102, 0.213) and 12 angular functions with eta (0.005), zeta(1,4) and lambda (-1,1).

I obtained similar kind of results for all the configurations that i used (always some of SFs are zero) and i have a warning like that in the training process:

Warning: Invalid scaling encountered in 'stp_normalize()'.

         This means at least one fingerprint function for O

 is always equal to zero!

Maybe i am misinterpreted the results. I will appreciate so much your  advice about this issue.

Best regards,

Alejandro Rivero Santamaría

Postdoctoral Researcher

Gas/Solid Interfaces group

Centro de Física de los Materiales

Spain

[Nongnuch Artrith]

Dear Alejandro,

Thank you for your email.

To clarify, the order of the fingerprint values in the training set is not the same as in the input file (atom.fingerprint.stp).  First all radial function values are listed, and then the angular functions follow.

Could it be that you defined one or more angular symmetry function(s) with O as central atom and O as coordinating atom? 

If you have only a single O2 molecule in your system, these angular symmetry functions would always be zero since there are only two Oxygen atoms (i.e., there is no O-O-O angle).

Note that the complexity of the symmetry function setups is precisely the reason why I have completely migrated to the Chebyshev descriptor by now.  The Chebyshev descriptor is also computationally more efficient.

Please let me know if you have any further questions.

Best,

Nong

[Alejandro Rivero]

Dear Dr. Artrith,

Thank you very much for your answer. After understanding how the fingerprint values are printed I confirmed that, as you said the zero symmetry functions for O atoms correspond to the angular functions with O as central atom and O as coordinating atom. In my case i can not avoid this issue because i have only two oxygen atoms in the system. I will follow your advice to use the Chebyshev descriptor.

Best regards,

Alejandro