16.10 Handling difficult cases: When SCF does not converge

General suggestions:

- Carry out convergence experiments with a small but reasonable basis set (e.g., cc-pVDZ, def2-SVP, aug-cc-pVDZ, ASVP). STO-3G is not a reasonable basis set.

Before you start you should check:

- Whether your geometry is sensible (e.g., look for Angstrom/Bohr conversion issues).
Note that Molpro prints bond distances in both Angstroms and atomic units at the top
of an output.
- Whether you have selected the correct electronic state (spin and symmetry).
Molpro tries to guess spatial symmetries of open-shell compounds
automatically if none are provided. However, the guess is not always right.
In such a case you need to give the symmetry manually
(in the simplest case as
`wf,sym=N,spin=M`. See section 16.2). Molpro does not attempt to guess the spin state of the input compound automatically; it defaults to spin=0 for systems with even numbers of electrons and spin=1 for odd-numbered species.

If convergence problems persist, the following techniques can be attempted:

Hartree-Fock options & Small-basis initial guess:

**Level shifts:**Adding a level shift like`{rhf; shift,-1.0,-0.5}`stabilizes the current RHF solution against changes and leads to smoother (but slower) convergence. That should be your first try; it is often sufficient.**Occupation freezing:**The option`{rhf,nitord=N}`can be used to freeze orbital occupations at iteration N. When the programs emits warnings about reassigned orbital occupations, you could try to freeze the occupations only later (give a higher N) or earlier (give a smaller N).By freezing the occupation pattern you tell the RHF program to try to lock on whatever solution it currently is pursuing. This often helps if multiple RHF solutions with similar energies are present and otherwise the program would oscillate between some of them.

Note that

`{rhf,nitord=1}`will tell RHF to lock onto the initial occupation; if combined with orbital rotation or advanced initial guesses this can often be used to converge to specific solutions (e.g., some excited states).**Minimal-basis SCF guess:**Try to obtain a Hartree-Fock solution with a minimal-basis AO set first and to use this as initial guess for the actual Hartree-Fock calculation. For this purpose we provided the basis set definition "MINAO" (to complement cc-pVnZ basis sets) and "MINAO-PP" (to complement cc-pVnZ-PP sets with ECPs). See section 16.4.1 for an example. These sets simply consist of the AO part of the cc-pVTZ or cc-pVTZ-PP basis sets, stripped of all their polarization functions. Since a minimal basis has fewer degrees of freedom than a real basis set, convergence is often easier, and it can still provide reasonable guess for the valence electronic structure.Note: The MINAO basis sets are very small, so conventional Hartree-Fock (in integral-direct mode if necessary) is typically much faster than density-fitting Hartree-Fock.

**Increasing the DIIS dimension:**In rare cases`{rhf,maxdis=30,iptyp='DIIS',nitord=20; shift,-1.0,-0.5}`can find solutions which are not found in the standard settings. Usually increasing the DIIS dimension beyond 10 (the default) just slows down convergence. It is also worthwile to try a variation of DIIS known as KAIN (Krylov-subspace accelerated inexact Newton):

`{rhf,maxdis=10,iptyp='KAIN',nitord=10; shift,-1.0,-0.5}`which sometimes shows different convergence behavior than straight DIIS.

Cationic or Anionic initial guess:

- If molecule does not converge, it might still be possible to converge (, ..) and use this as initial guess for the actual computation. Particularly if is a closed-shell compound this will often work. If using this technique, you need to be careful about the final state you arrive in. Because your initial guess is biased, the calculation might converge to an excited state.

Density-functional initial guess:

- For transition metals and transition states sometimes DFT methods show better
convergence behavior than RHF. You might perform a DFT calculation (possibly
with a smaller basis set) and use it as initial guess for the SCF. E.g.,
`{df-rks,b-lyp; coarsegrid; save,2100.2}``{df-rhf,nitord=1; orbital,2100.2}`

If all else fails: Use the MCSCF program. The MCSCF program uses an advanced orbital optimization algorithm which is much more robust than the SCF method, and which can converge almost everything you give to it (but it is often slower and sometimes locks on an excited state if started from an atomic density guess). MCSCF can also calculate Hartree-Fock solutions if used with suitable input cards.

molpro@molpro.net 2018-12-15