Can't get accurate forces with the trained model.

[Lijieying]

Summary

The properties ones concerned is the accurate forces in research, almost 1E-3 to 1E-2 eV/Å. But the forces in training and validating converged to 1E-3 to 1E-2 eV/Å only, and the rmse of the predicted forces could only reach 1E-2 tp 1E-1 eV/Å. Is that normal? If it is, will it mean different algorithoms of the force calculation between vasp and deepmd-kit? Or not, Is there anything wrong with my model training jobs?

DeePMD-kit Version

2.1.5

TensorFlow Version

2.9.0 (cpu version)

Python Version, CUDA Version, GCC Version, LAMMPS Version, etc

python 3.10.0.

Details

The dataset for my model is generated from an equilibrium state with ab-init MD(npt,press from 0.0 to 1.0, temperature from 300 to 1200). And the accuracy of energy and forces are pretested, 1E-6 eV and 1E-4 eV/Å respectively. Training parameters are shown in attachements. the following lcurve.png is the best outcome of all my trained model.
Looking forward to your response. Thanks~
input.json

[njzjz]

But the forces in training and validating converged to 1E-3 to 1E-2 eV/Å only, and the rmse of the predicted forces could only reach 1E-2 tp 1E-1 eV/Å. Is that normal?

Could you explain this sentence? What is the difference between the training and validating forces and predicted forces?