Bug: Potential issue when setting sigmas in gap_rss_set_config_sigmas_from_convex_hull.py about workflow

[bernstei]

I guess we can fix that, but it's going to be ugly, since it'll break backward compatibility, even if the new behavior is more correct. Ugh.

[bernstei]

See #37. Is it time to try to formalize something having to do with requiring minimum version, etc, in the config file?

[bernstei]

Is it actually clear how those sigmas should scale, except that they should increase with number of atoms? Linear in N? Linear in sqrt(N)?

Note that this is a non-trivial change to the heuristics - the straightforward solution (i.e. #37 as of now) will increase the energy and virial sigmas by factors of 6-24 (in the gap-rss iteration, more in the defect_MD iterations since cell sizes are bigger), but the force sigmas remaining the same.

[VladCarare]

will increase the energy and virial sigmas by factors of 6-24 (in the gap-rss iteration, more in the defect_MD iterations since cell sizes are bigger), but the force sigmas remaining the same.

But the resulting per atom sigma will be consistent with the force sigma regardless of the number of atoms, no? Sorry I'm a bit confused about what the problem may be with just multiplying the appended energy, virial and hessian sigma by the length of the atoms object.

[bernstei]

Even if you assume that the current heuristic F_sigma = A sqrt(E_sigma) is correct, there's no inherent value for E_sigma or A - it's just a heuristic after all. When we make E_sigma larger than it used to be by multiplying by sqrt(N), we have to decide if we want to keep the typical E_sigma the same (so reduce the value that's multiplied by sqrt(N)), or allow it to be larger, and then independently decide if we also want the F_sigma to be larger, or if we want to reduce A so that F_sigma is similar to what it was before.

[bernstei]

closed by #37

[bernstei]

@VladCarare Any problem if I close this issue? I claim that #59 changed the heuristic to one that is more consistent with what we want.

[VladCarare]

No, it can be closed. I stopped testing with having a sigma-setting function some time ago, so I don't have anything to add.