Force RMSE Threshold required for using u-MLIP/fine-tuned for geometry relaxation Vs MD simulation.

[amannsonii3]

What is the threshold force RMSE(kcal/mol•Å or in eV/Å) I should consider to use my MLIP for geometry relaxing my systems to the ground state structures? I know DFT has an inherent error of ~1 kcal/mol for energies, but I don't know about forces. Similarly how does it go for when I want to run a MD simulation?

Do I have to fine-tune my model to the point it matches my ASE optimiser's force threshold (0.05 eV/Å or 0.02 eV/Å)?
What should this range be in case of high-throughput workflows and high-quality workflows?

(It would be really amazing if someone could cite any paper for this, rather than relying on just intuition)

[ilyes319]

@amannsonii3 You mean the force threshold for the optimiser? Usually, we would use 0.001 eV/A, but for high-throughput workflows one can use 0.02 eV/A to speed up things (see for example what people use in the matbench discovery workflow).

[amannsonii3]

No,

I'm asking about when fine-tuning a foundational model, till what error range do I fine-tune it on the reference DFT? So that I can use that fine-tuned potential as the DFT for doing geomtery relaxations for set of chemical systems.
And to decide if we even need fine-tuning to begin with, and just use the foundational model instead.

If the MAE, RMSE isn't the correct metric to look at in this case (ground state structure relaxation) then what should it be?
(relative energies or something...)