Energy shift of all the datasets that iteratively added from the first-principle calculation

[DXiming]

Summary

Energy shift of all the datasets that iteratively added from the first-principle calculation, while for the initial data, the prediction is well.

DeePMD-kit Version

v2.2.0b1

TensorFlow Version

2.11.0

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

No response

Details

Hi, I have some issues with the training procedure using multiple datasets generated from dp-gen. The results show good prediction with the initial dataset, while for the iteratively added datasets, like "iter1-0", "iter0-1", and "iter0-0", there is a shift for the predicted energy. The training curve shows that the training procedure seems separated for each dataset with abnormal fluctuation of some training steps. I have tried to merge all the datasets into one, but the problem still exists. Can you help me solve this problem? I really appreciate any help you can provide.

input.json

[njzjz]

You might check if you use the same first-principle level for initial data and added data or any different things. From Figure 1, it's obvious that the two data sets are different.

[DXiming]

Thanks for your suggestion. This discrepancy is due to the omission of the VDW correction of my first-principle calculation.