Option to use a stress tolerance instead of virial?

[Technici4n]

The stress is (minus) the virial divided by the volume. Since the volume can change during optimization, and stress is the variable that is usually used for DFT geometry optimization, it would be nice to be able to directly give a stress tolerance to the geometry optimization.

Additionally, I find it a bit weird that the pressure is computed from the trace of the virial without any division by the volume:

GeometryOptimization.jl/src/callbacks.jl

Line 54 in 8701710

pressure = -austrip(tr(geoopt_state.virial)) / 3

Could that be a bug?

[mfherbst]

I used the virial, because we used it in the interface, but have not given it much thought beyond that. I agree under volume changes stress may be better.

Yes, I agree, the pressure looks wrong.

[cortner]

under volume changes stress may be better

maybe but what's the basis for this? I can use an arbitrary scaling and get a descent direction? Is there experience from established codes like ASE etc?

(or are you thinking of this only as a termination criterion rather than a search direction? of only termination, then I don't have a strong view)

[Technici4n]

only as a termination criterion rather than a search direction?

That was at least indeed my thinking.

[jameskermode]

ASE currently uses virial/cell_factor as the search variable, with cell_factor scaling as the mechanism that makes a single force tolerance double as an effective stress tolerance. We are planning to change this though, and add seperate force and stress tolerances.

[jameskermode]

stress is the instrinic quantity, so a single tolerance would be transterable across system sizes, unlike a virial-based tolerance which would be extrinsic.