Is it possible to run GGA DFT with fractional occupations (Fermi smearing)?
“Fractional occupations” and “Fermi smearing” are not synonymous, and the answer to your question depends on what you mean.
Psi supports user-specified fractional orbital occupation. It does not support introducing a temperature to automatically compute fractional occupations of all orbitals from a Fermi distribution applied to single-particle states, i.e., Fermi smearing.
For an example of fractionally occupied code, please see here and here.
Thank you. I was thinking of an automatic solution, i.e. Fermi smearing. I am interested in DFT calculations on small metal clusters and the default settings lead to lack of convergence (oscillatory behaviour in the SCF loop). In other codes, high damping + level shifting (and, as a last resort, Fermi smearing) usually did the job. What else would you suggest, apart from damping?
I don’t understand why you request solutions “apart from damping” if you were willing to accept high damping + level shifting in other codes. Psi has this combination as well in 1.4. Is it not converging?
For a list of available tricks in Psi, see this section of the manual. If I need a low-fuss solution and can afford to wait a bit, I’ll just use
soscf. MOM may be useful in your case - I don’t have enough information to know. If those two fail, manually playing with DIIS, level shifting, and damping is the best I can offer.
My own research group is having problems with SCF convergence involving transition metal systems, so I can be very easily convinced to add new SCF tricks into Psi. I’m currently committed to adding EDIIS and ADIIS, but it’s not the only project I’m responsible for, so it’s not going as fast as I would like.
Thank you very much. I tried several things (damping 10-80%, SOSCF true/false, scf_type pk or df, lanl2dz or def2-svp or def2-tzvp ecp&basis, mom on/off). I also increased the grid to dft_spherical_points 590 and dft_radial_points 99
The oscillations I mentioned before are gone but still the convergence appears to stall, e.g.:
@RKS iter 370: -595.69533209474992 1.47793e-12 3.08330e-03 DIIS/MOM/DAMP=40%
@RKS iter 371: -595.69533209475264 -2.72848e-12 3.08330e-03 DIIS/MOM/DAMP=40%
@RKS iter 372: -595.69533209475071 1.93268e-12 3.08330e-03 DIIS/MOM/DAMP=40%
@RKS iter 373: -595.69533209475048 2.27374e-13 3.08330e-03 DIIS/MOM/DAMP=40%
@RKS iter 374: -595.69533209475162 -1.13687e-12 3.08330e-03 DIIS/MOM/DAMP=40%
@RKS iter 375: -595.69533209474935 2.27374e-12 3.08330e-03 DIIS/MOM/DAMP=40%
What else can I try? The geometry is, I believe, close to equilibrium and the spin state should also be OK (I took the structure from a paper, it is just a cluster of a handful of metal atoms)… Any suggestions will be welcome.
What Psi has is what’s in the manual. The only options are that you haven’t found the magic combination, there’s a bug somewhere in the SCF code, or the features we have just aren’t enough.
We’re collecting difficult-to-converge SCF cases so we can assess the weaknesses of the current SCF code. That’s as much as I can do.
Which version of Psi4? IIRC there have been some problems with the ECP code…
Version 1.3.2, I’m aware that there is some work underway on the ECP code. Anyway, we managed to converge this system in PySCF. I am more than happy to share the details of this case as soon as we publish.