Some questions about using DeePTB-E3 in liquid system

[hwf2002]

Details

When I use DeePTB-E3 to train the model of liquid water box with 64 water molecular. It seems like the band predict is not fix with the band of dft. Thought the train_loss has been to 0.0005. So I want to know how to fix this problem.

[hwf2002]

https://zenodo.org/records/19042378
I upload the file on the zenodo, and the train dataset and the model(tast and nnsk) need to be unzip.