Thanks Daniel, I recognize that bug hunting for Psi4Corp isn’t always lucrative.
Jet confirmed on github that the problem still occurs even when DK is turned off. He took the input file I posted on Jan 9 and commented out:
#relativistic dkh
#dkh_order 2
and the got the same result
But tonight the plot thickens!
Matthew was originally trying to run the 4Pi state using a basis set that he knew converged to the correct state and then project it to provide a better guess when using a basis set we knew was not converging to the correct state.
The basis set he is using above to start with is the standard non-relativistic cc-pVDZ basis set. When decontracted, the cc-pVDZ and cc-pVDZ-DK sets should give the exact same answer, right?
Could the problem be somehow related to using non-DK optmized basis sets with DKH calculations?
The following input file instead uses cc-pVDZ-DK for the first SCF, then projects into cc-pwCVTZ-DK. It does the projection correctly, but finds the incorrect occupation in both the initial and final ROHF:
memory 8000MB
molecule {
0 4
Fe
H 1 1.566665
}
set {
reference rohf
docc [7, 0, 3, 2]
socc [1, 1, 0, 1]
scf_type pk
relativistic dkh
dkh_order 2
print_mos true
maxiter=500
freeze_core false
}
basis {
assign Fe cc-pvdz-dk
assign H cc-pvdz-dk
}
set scf df_scf_guess false
set scf guess gwh
set ccenergy r_convergence 10
energy(‘scf’)
clean()
basis {
assign Fe cc-pwcvtz-dk
assign H cc-pvtz-dk
}
set scf guess read
optimize(‘cc3’)
Now if I run this input file, which fundamentally is the exact same input file that Justin and I looked at on the github thread, the initial ROHF energy is screwed up (too high by almost 2 Hartrees!) while the projected energy seems reasonable but alas still has the incorrect occupation (off by less than 0.1 Hartrees).
memory 8000MB
molecule {
0 4
Fe
H 1 1.566665
}
set {
reference rohf
docc [7, 0, 3, 2]
socc [1, 1, 0, 1]
scf_type pk
relativistic dkh
dkh_order 2
print_mos true
maxiter=500
freeze_core false
}
basis {
assign Fe cc-pvdz
assign H cc-pvdz
}
set scf df_scf_guess false
set scf guess gwh
set ccenergy r_convergence 10
energy(‘scf’)
clean()
basis {
assign Fe cc-pwcvtz-dk
assign H cc-pvtz-dk
}
set scf guess read
optimize(‘cc3’)
The only difference between the input above and the original is here I am using the PSI4 library version of the cc-pVDZ sets instead of defining it within the input file and then using the “assign” that Lori suggested:
set basis mine2
set basis_relativistic mine2
So, an issue with decontraction maybe?