I have been having significant trouble eliminating residual imaginary frequencies during geometry minimisations of fairly typical organic molecules in Dmol3 (PBE0+TS/DNP/COSMO). 

I have tried the typical remedies on various structures  - tightened geometry convergence criteria beyond the pre-set "Fine"; changed the SCF convergence, orbital cutoff and integration accuracy to "Fine"; perturbed structures along the corresponding normal modes etc. - but repeated optimisations still often result in one, or even several imaginary, frequencies. 

I'd be grateful if anyone had some DMol3-specific advice (e.g., there appears to be an xfine setting for integration_grid, but this setting appears unavailable in the Dmol3 Calculation tab). Thanks