How to train if I have only force data and no energy data?

[utkarshp1161]

How to train if I have only force data and no energy data?

[Linux-cpp-lisp]

Hi @utkarshp1161 ,

Just omit total_energy from loss_coeffs. Everything else is the same.

[Linux-cpp-lisp]

No, you need to remote total_energy from loss_coeffs entirely, while leaving forces (and whatever other properties you do want to train on)

[Linux-cpp-lisp]

Setting the coefficient to zero has the same effect, but will still cause an error if there are no total_energy labels in your data at all

[Andy6M]

Hello @Linux-cpp-lisp,

I came across the previous discussion where @utkarshp1161 inquired about training with only force data and no energy data. You mentioned that they would need to remove total_energy from loss_coeffs entirely while leaving forces.

In my situation, I have the opposite problem: my dataset contains energy data but no force data. Could you please guide me on how to train with only energy data and no force data? Any advice or recommendations would be greatly appreciated.

Thank you for your time and assistance!

[Linux-cpp-lisp]

Hi @Andy6M,

https://github.com/mir-group/nequip/blob/main/configs/minimal_eng.yaml should help get you started on that direction.

[Andy6M]

@Linux-cpp-lisp Thanks.