[BUG] Unable to calculate relative force deviation.

[xiaohe-njust]

Bug summary

When using dpgen for structural exploration, if an external LAMMPS template is used, the relative force model deviation cannot be calculated, and the output model_devi.out file information is the absolute force deviation. When no external LAMMPS template is used, everything is normal.

DP-GEN Version

0.12.1

Platform, Python Version, Remote Platform, etc

No response

Input Files, Running Commands, Error Log, etc.

param.json

"use_relative": true, "epsilon": 1.0,
If an external LAMMPS template is used, regardless of whether the code for calculating relative force deviation is used, the output model_devi.out file information remains the same. Trying to change the value of the keyword epsilon still has no effect. When no external LAMMPS template is used, everything is normal.

Steps to Reproduce

input.zip

Further Information, Files, and Links

No response

[njzjz]

This is actually the same problem as #1608

[cheng2t]

Hello, I'm also experiencing this problem, have you solved it yet?

[xiaohe-njust]

Hello, I'm also experiencing this problem, have you solved it yet?

No, it's not settled.

[njzjz]

Hi @cheng2t, we don't have enough people to maintain this project. If you are willing to contribute to the project, don't hesitate to do it!

[cheng2t]

Thanks for your reply，Do you have any problems with the phenomenon that the accuracy rate has not been improved during the iteration process?

[xiaohe-njust]

@cheng2t Yes, I encountered this issue and discovered this bug. Incorrect use of absolute force deviation can lead to the newly added structures in each iteration being unreasonable. If your configuration involves large force values, you can try increasing the value of model_devi_f_trust_lo. The numerical setting for relative force deviation is not suitable for setting absolute force deviation.

[cheng2t]

Is it convenient for you to tell me how much you set the value of model_devi_f_trust_lo and model_devi_f_trust_hi?