According to an amber-l post, here's the best practice to avoid inconsistency:
In my humble opinion, the most important consideration for correct charge
derivation is that the conformation of (especially) flexible molecules is
forced to be as extended as possible to prevent any intramolecular
interactions (H-bonding) that would corrupt the charge fitting method (the
documentation to Openeye's Quacpak clearly quotes Bayly in this
regard...). The simplest way to achieve this is use a program like VegaZZ
to assign charges using their SP4 FF then simply change the signs to be
all positive or all negative. Then use the SP4 FF to minimize the
structure. Since all atoms will repel each other, you will (always?)
achieve the most extended conformation. Then use MOPAC with AM1 to
minimize that structure, and then send it to antechamber. My gut feeling
is that the inconsistent use of the most extended conformations is much
more detrimental to accurate charge derivations than any advantage sqm
might have over MOPAC for assigning the AM1BCC charge.
Without knowing if your structure is highly flexible, it's hard to know if
what I discussed is relevant to your particular situation or not.
At the end of the Ambertools 1.5 it states that for AM1-BCC charges for
585 drug molecules, MOPAC and sqm give essentially similar charges for all
the cases and the average charge difference is 0.005. So what are the
advantages to sqm over MOPAC for AM1-BCC, since there appears to be tight
convergence issues with sqm?
It was only after sqm was run "in the wild" on many more compounds than
we ever initially tested that we became aware of occasional convergence
By default, sqm demands tight convergence for both SCF and geometry
optimizations, whereas MOPAC has much less stringent criteria. We may not
have the most appropriate defaults, especially for am1-bcc calcs. This is
discussed in the antechamber chapter of the AmberTools Users' Manual.
So the punch line is sqm tolerance is too tight, maybe try mopac in amber9 or (10?)