If spin-unrestricted calculations fail to converge, you may try to run first a restricted calculation. Then perform the unrestricted while using the TAPE21 from the earlier run as a restart file. Now, in the restart run you have two alternatives that may stand a better chance to give you the desired result than your original failing calculation:
(1) Apply the MODIFYSTARTPOTENTIAL key to steer the (at the least the initial) spin-density on a per-atom (fragment) basis towards what you expect should be the final self-consistent situation. It may be necessary to apply stronger damping, or do other SCF convergence tricks like level shifting, in order to preserve your initial construction long enough for the system to get near the minimum it should converge to. It is recommended to supply occupation numbers (in case the system has symmetry) in input.
(2) If all SCF tricks fail (but you did get the restricted converged result), you may simply run a one-cycle unrestricted run, specifying the occupation numbers you consider appropriate and so obtain the unrestricted occupation state for the restricted-self-consistent MOs. The difference with the 'true' converged unrestricted MOs is often quite small. See the Examples document for the discussion of a similar case (H2).
