From: JC Gumbart (
Date: Sun Oct 15 2017 - 07:24:45 CDT

(bringing this back on list so that others can benefit from the discussion)

Did you run the molecule through the cgenff/paramchem webserver? I just did. Only a couple atoms’ charges have penalties above 50 (C6 and C12). The problem is that these are “buried”, in the sense that they have no Hs for the gaussian calculations. In our experience, it’s difficult to improve upon these charges without significant effort and diverging from standard charmm procedures, e.g., using NPA charges as a starting point. To be clear, Mulliken charges are not expected to be accurate; they are just used because you have to start with something.

Some of the parameters have much higher penalties and will be worth further optimization. Even then, I would only optimize ones close to or above 50.

The difficulty (one that we’ve all run into!) is that force fields are built through a combination of fitting data, chemical intuition, and a desire for transferability. FFTK addresses the first only. CGenFF helps with the third and limits the need for the first and second.


> On Oct 15, 2017, at 5:18 AM, Luba Simhaev <> wrote:
> Hi JC,
> I checked all the gaussian output files and removed from the optimization those that produce water states with high energy.
> Also I saw that it is recommended to set the distance weight at a value lower than 1.0.
> But even with these changes, the optimization produces weird results. There is a strong disagreement between the Mulliken charges from the SP calculation and the charges from the optimization process.
> Can you please advise me how to proceed?
> Also, how can I learn about the convergence of the optimization from COLP?
> Thanks,
> Luba