Fluctuating energy levels in band structure plots - DeepTB-SK

[SoumyajeetMazumdar]

I have noticed that the band structures produced—both for the initial SKTB baseline and the final trained models—exhibit large, inconsistent vertical shifts relative to the Fermi energy specified in the band.json file.
This behaviour is also raising concerns about the reliability of using band windows for TB parameterisation when the absolute energy levels appear to fluctuate so much during the training process.

For a Fermi energy of ~3 eV in the actual Band structure (blue):

While the initial SKTB baseline model would understandably be shifted, the original plot also shows the DFT reference shifted by ~23eV, while the E_f = 3eV.

These shifts are also present during the training process, as shown for one of the intermediate checkpoint files:

I would appreciate any input on this or similar issues that others may face.