PES Scan converge issue in HF molecule

Hi, all friends,

I have got PES Scan converge issue in HF molecule, the input is like below:

memory 8 GB

molecule HF_mol{
0 1
F 1 R
#Symmetry c1



set {
basis aug-pcseg-3
guess SADNO
soscf false
reference rhf
scf_type PK 
e_convergence 10
dft_spherical_points 590       # Often needed
dft_radial_points 135          # Often needed
dft_radial_scheme treutler     # Rarely needed
dft_pruning_scheme robust      # Generally safe and will speed things up

E_Dipole = {}
E_energy = {}

for R in Rvals:
   HF_mol.R = R
   E_energy[R] = energy('PBE')
   E_Dipole[R] = properties('PBE', properties=['dipole'])

for R in Rvals:
   e = E_energy[R]
   d = E_Dipole[R]
   psi4.print_out("%3.1f  %10.6f  %10.6f\n" % (R, e, d))

previous points (when R < 2.2), the SCF are converged very well, after that, converges are very difficult. the sysytem Occupation by irrep changed frequently,PSI4 can not get a converge results.

I have tried many many combinations: change the initial guess, Damp,levelshift,MOM, trunoff Symmetry …

SOSCF could help in some case, but also failed at R=2.5, and there is drawback that SOSCF can not use in meta-gga?

By the way, use the very same input in ORCA, I have got very smooth PES without any issue.

So, any possibility do these PES job further?

With the new ORCA 5.0 and it’s new SCF converger? Same requested accuracy?

The issues I see for psi4 is that the SCF jumps between occupations/states.


The necessary terms are unfortunately not implemented.

If Psi is jumping between states, why not specify the symmetry of the occupation with DOCC?

yes, I use the very same/similar input, below is the ORCA input:

! RHF NoRI DefGrid3 aug-pcseg-3
! PAL6 MiniPrint SCFCONV10
%method intacc 16 end

# Scanning from the outside to the inside
%paras R=0.5,6,56

* xyz 0 1
h 0 0 0
f 0 0 {R}

Almost the same converge criteria, almost the same DFT grid setup(ORCA 110322; PSI4 110786)

Thanks @jmisiewicz
yes, fix the occupation with DOCC is the solusion. And at the same time, I am here to report the converge performance of PSI4.

Thanks. It’s an interesting test case.

ORCA’s KDIIS is surprisingly strong as always. The new TRAH-SCF is certainly struggling but manages as well in the end.

My group doesn’t have access to ORCA, so I can’t do much more digging on this. But, some quick notes:

  1. It isn’t obvious to me this is even a fair comparison. If ORCA’s initial guess is the orbitals from the last geometry, of course it will have an easier time converging than with your SAD guess. (Incidentally, you can enable something similar in Psi with set guess read, though there is a bug in it that I need to fix.)
  2. Yes, I am interested in improving Psi4’s SCF capabilities. ADIIS/EDIIS will be in the next Psi release, although that doesn’t solve this particular example.

I tried this, and also a PBE0 and HF guess. It doesn’t help.

Mind that a GGA is expected to struggle here. For bond breaking it is also usually recommended to go from long to short distances when reading previous orbitals.

No, every single point calculation in ORCA is proceed alone without any previous orbitals.