I have tried running your case myself and for me it does not converge either. Now COSMO in NWChem does not have analytical gradients and the gradients for the geometry optimization are obtained from finite differences. This can introduce some numerical issues. So I want to test a few things to see if I can come up with a recipe that works.