Invalid scaling encountered in 'stp_normalize()'

[Williams1996]

Dear AENET group,
I am training a potential for a perovskite structure with the xsf file attached below. But I always meet the problem
"Warning: Invalid scaling encountered in 'stp_normalize()'.
This means at least one fingerprint function for Mn is always equal to zero!
Maybe an atomic species is not present in the reference set?
type = Chebyshev 0
covariance = 1698.3182700627203
average = 41.210657238907515
min, max = 41.210657238560906 41.210657239138584
"
The training result is NaN then.

I tried to adjust the cutoff radius and radial_N, but I still cannot solve this.
Can you help me find out the problems?
The structure and *.stp file I prepared are as follows.
Thank you so much!

One of the XSF files:

total energy = -132.36555 eV

CRYSTAL

Primitive lattice vectors in Angstrom

PRIMVEC
5.32280874250000 0.00000000000000 0.00000000000000
0.00000000000000 7.30292844770000 0.00000000000000
0.00000000000000 0.00000000000000 5.12996721270000

Cartesian coordinates in Angstrom.

PRIMCOORD
20 1
Mn 0.00000000000000 0.00000000000000 2.56498360635000 0.0 0.0 0.0
Mn 2.66140437125000 3.65146422385000 0.00000000000000 0.0 0.0 0.0
Ni 0.00000000000000 3.65146422385000 2.56498360635000 0.0 0.0 0.0
Ni 2.66140437125000 0.00000000000000 0.00000000000000 0.0 0.0 0.0
O 2.56119739209885 5.47719633577500 0.42036145084234 0.0 0.0 0.0
O 1.58628466127937 3.95780589624450 3.60607942099579 0.0 0.0 0.0
O 1.58628466127937 6.99658677530550 3.60607942099579 0.0 0.0 0.0
O 0.10020697915115 1.82573211192500 2.98534505719234 0.0 0.0 0.0
O 1.07511970464782 0.30634167239450 1.04109580951583 0.0 0.0 0.0
O 1.07511970464782 3.34512255145550 1.04109580951583 0.0 0.0 0.0
O 3.73652408122063 0.30634167239450 1.52388779170420 0.0 0.0 0.0
O 2.76161135040115 1.82573211192500 4.70960576185766 0.0 0.0 0.0
O 3.73652408122063 3.34512255145550 1.52388779170420 0.0 0.0 0.0
O 5.22260176334885 5.47719633577500 2.14462215550766 0.0 0.0 0.0
O 4.24768903785218 3.95780589624450 4.08887140318417 0.0 0.0 0.0
O 4.24768903785218 6.99658677530550 4.08887140318417 0.0 0.0 0.0
Sr 2.35149648985713 1.82573211192500 2.48822500988932 0.0 0.0 0.0
Sr 5.01290085578432 1.82573211192500 0.07675859133071 0.0 0.0 0.0
Sr 0.30990788139287 5.47719633577500 5.05320862136929 0.0 0.0 0.0
Sr 2.97131225264287 5.47719633577500 2.64174220281068 0.0 0.0 0.0

One of the *.stp file in generate.in:
DESCR
N. Artrith and A. Urban, Comput. Mater. Sci. 114 (2016) 135-150.
N. Artrith, A. Urban, and G. Ceder, Phys. Rev. B 96 (2017) 014112.
END DESCR

ATOM Mn

ENV 4
Mn
Ni
O
Sr

RMIN 0.55d0

BASIS type=Chebyshev
radial_Rc = 8.0 radial_N = 20 angular_Rc = 8.5 angular_N = 16

generate.in:
OUTPUT aenet.train

TYPES
4
Mn -0.0 ! eV
Ni -0.0 ! eV
O -0.0 ! eV
Sr -0.0 ! eV

SETUPS
Mn Mn.fingerprint.stp
Ni Ni.fingerprint.stp
O O.fingerprint.stp
Sr Sr.fingerprint.stp

[alexurba]

Dear @Williams1996,

You would see this warning if your entire data set only contains a single structure, the one that you attached to the issue above. Could this be the case?

For training machine-learning interatomic potentials, many reference structures are required. A single example is not sufficient. Could you please elaborate on your objective? If we understand it better, we might be able to help.

[Williams1996]

Thank you so much for the timely reply. Currently, the training data set contains 5 structures.
Is it because the total number of structures is too small?
Currently, I start with a small number to have a try, and will increase the number of structures with a larger supercell size.
I want to run the code abICS (https://github.com/issp-center-dev/abICS) to get a crystal structure with lowest energy for a specific ratio of compositions, like SrMn0.5Ni0.5O3. The abICS code use the aenet potential to predict the energy.

Best regards,

[Williams1996]

I find the problem is in the scaling process,
sfsetup.f90: subroutine stp_normalize, (line 1009)
s = sqrt(stp%sfval_cov(isf) - shift * shift)
The sqrt() will give a NaN value, if the input is negative. I am not so clear why the inside will be negative.
But converting the inside to the absolute value will solve the problem in my case.
s = sqrt(abs(stp%sfval_cov(isf) - shift * shift))

I think this will not affect the ML training and prediction too much.
Finally, the code works.
Thanks for your code and best regards.