LargeSigmaHandler set ISMEAR=0 if sigma reaches 0.02 minimum

[janosh]

Half the jobs in our atomate2 r2SCAN trial runs fizzle due to LargeSigmaHandler. The failed runs are mostly metals.

After starting with sigma=0.2 and decreasing 3 times, we reach the minimum of 0.02 allowed by Custodian at which point the LargeSigmaHandler aborts the job.

VASP wiki says Gaussian smearing leads to reasonable results in most cases. The final summary reiterates if you have no prior knowledge of the system, use ISMEAR = 0.

Alternative suggestion from @esoteric-ephemera is to increase the max allowed entropy from 1 meV/atom even though VASP docs recommend that value. They probably didn't benchmark that value for larger structures. Or better yet, make it a user-overridable entropy_tol kwarg.

custodian/custodian/vasp/handlers.py

Line 1264 in 202b941

if entropy_per_atom > 0.001:

[Andrew-S-Rosen]

I agree that ISMEAR = 0 is a good fallback option. It is a safe bet in virtually all cases and is likely to be more robust a solution. However, there is one important caveat. If the user has specified ISMEAR = -5, they almost certainly want to only do ISMEAR = -5. Changing ISMEAR = -5 to ISMEAR = 0 would likely cause hidden issues for the user that they likely would not know about. So, my personal $0.02 is to implement a change like this but only if vi["INCAR"].get("ISMEAR", 0) > 0.

Edit: Actually, SIGMA is just irrelevant with ISMEAR = -5 so none of this should matter then.

[janosh]

@arosen93 Looks like the handler only applies if ISMEAR is > 0 to begin with.

custodian/custodian/vasp/handlers.py

Line 1254 in 202b941

if incar.get("ISMEAR", 0) > 0:

[Andrew-S-Rosen]

@janosh: Beautiful. Then I think this makes sense.

I would also recommend explicitly setting a value for SIGMA if you set ISMEAR = 0. Perhaps 0.05, which is consistent with other parts of Custodian?

custodian/custodian/vasp/handlers.py

Line 1057 in bfc7049

actions.append({"dict": "INCAR", "action": {"_set": {"ISMEAR": 0, "SIGMA": 0.05}}})

Per the VASP manual:

If you have no a priori knowledge of your system, for instance, if you do not know whether your system is an insulator, semiconductor or metal then always use Gaussian smearing ISMEAR=0 in combination with a small SIGMA=0.03-0.05.

[esoteric-ephemera]

I second SIGMA = 0.05, that's pretty safe