Training Issue: High Force Error

[mcourtiel]

Hello everyone,

I am training a model on a non periodic system (Al)45+ and I am working with deepmd-kit V2.2.10.

The problem I am facing is that even if the error on the energy per atom is reaching 8.5e-3, the error on the forces it is "stuck" at 1.5e-1.

Here you can find an input file :
input.json
And here the lcurve.out :
lcurve.txt

If anyone has encountered this issue or has insights into potential solutions, I would greatly appreciate your advice.

Thanks a lot !

[wanghan-iapcm]

What is the standard deviation of the force labels?

[mcourtiel]

The standard deviation on forces is 0.49924 eV/A.

[njzjz]

Could you plot some figures like #3678?