Fix DriftErrorHandler or remove it from default handlers

[mkhorton]

So drift in forces is obviously important to catch, since it means that you might be misled into thinking you are more force converged than you actually are. However, it is not a "failure" in the traditional sense: the calculation can still finish and be otherwise accurate. One could imagine a warning in, say, the atomate drone if ingesting a calculation where drift > desired force convergence, and warn there, instead of causing outright failures.

Currently, the DriftErrorHandler fails to actually fix drift issues, and instead causes a large quantity of jobs to fail since it's in the default handler group... I've seen 1.4k issues in the last month.

Tagging @shyamd since I know it's your creation -- do you have any suggestions to fix drift issues, beyond the strategies in the handler?

[shyamd]

The handler group is part of atomate, so I would say the change has to happen there to not include the error handler if you don't want it. There are really no other good ways to fix "drift" issues. Fundamentally it's a lack of accuracy in the calculation of forces and the only way to up that is by increasing the density of the augmentation grid.

[mkhorton]

Sure, I appreciate the handler group is in atomate -- but ideally I would like a DriftErrorHandler that fixes drift, which would be a change here.

Surely there are other ways to fix drift ... increasing ENCUT is probably a good way to start. It's not just the augmentation grid.

[mkhorton]

EDIFF also

[mkhorton]

@utf any thoughts on if this all these drift issues could be related to the ediff/atom scheme we're using? I haven't investigated myself yet but it seems like it could be another reason to move away from that

[shyamd]

Updating ENAUG is a way to get the effect of increasing ENCUT without adding extra bands. Rather than increasing EDIFF, I would say the drift tolerance should be limited to something like 2*EDIFFG and at least 10X EDIFF or something like that.

[mkhorton]

See slide 28 of https://www.vasp.at/vasp-workshop/slides/accuracy.pdf also -- handler currently doesn't touch LREAL, ROPT