Contradiction between Energy and Force Accuracy in DPGEN Training

[zhangyu0104]

Dear all,

I am encountering an issue with the model accuracy while training iteratively using DPGEN for a Fe-H binary system, where the magnetism need to be considered. After several iterations, the accuracy ratio has improved from 70% to over 95%. However, when testing the model with dp_test, I noticed that while the root mean square error (RMSE) for force is relatively small, the RMSE for energy is significantly large.

I am not sure whether this issue is related to the incorporation of magnetic effects. However, regardless of whether I fix the total magnetic moment or only set the initial magnetic moment, the energy RMSE continues to increase gradually with each iteration, while the force RMSE decreases. I also attempted to adjust the prefactors of energy and force during training, but this has not led to any improvements.

My input parameter file is attached below. I would greatly appreciate any insights or suggestions on how to address this issue.

param.json

Best regards,
Yu Zhang

[wanghan-iapcm]

Dear Yu,

The model trained during dpgen iterations pay little attention on the accuracy of energy. Note that we select configurations for labeling based on force criteria.

For a production model, one needs a long-train and perhaps a few rounds of training (with --init-model option) with larger energy loss prefactors.

Best,
Han