The trained forces are much better than energy. How could I improve the results of energy?

[I10140317]

I have done some trainings on a small molecule and the obtained results is shown in the attached figure.
The input.json is also attatched.
As can be seen, the energy is much worse than force. I have tried to increase the start_pref_e and limit_pref_e as suggested in some discussions. However, both energy and force become worse than using the parameters given in the examples.
I wonder whether there is any other suggestions to improve the results.

Thanks!
input.json

[njzjz]

You may consider increase the prefactor of the energy loss, i.e. loss/start_pref_e and loss/limit_pref_e. The current value is quite small.

[njzjz]

By the way, you need to ensure the DFT accuracy is enough before doing everything else.

…

On Fri, Feb 9, 2024 at 2:41 PM Jinzhe Zeng ***@***.***> wrote: You may consider increase the prefactor of the energy loss, i.e. loss/start_pref_e and loss/limit_pref_e. The current value is quite small. — Reply to this email directly, view it on GitHub <#3251 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACIOQ7WUH6M47D366JNYQ33YSZ3VLAVCNFSM6AAAAABDA3UHROVHI2DSMVQWIX3LMV43SRDJONRXK43TNFXW4Q3PNVWWK3TUHM4DIMRSGY3TA> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.*** .com>