behavior of the force error on different systems

[martino95]

Hi,
I'm using deepmd to learn an all atom potential for organic molecules in water. I have this issue: in the long run the force error curve splits in a number of curves equal to the number of systems I am using: the model perform systematically different on different systems (systems contains the same molecules in different configurations). Energy seems to behave well. I can't get an explanation for this phenomenon: why is it limited to forces? Is this a known issue?
Greetings
Riccardo