I would perform the sampling simulations using the mutual AMOEBA water / ion models at room temperature, but also a slightly higher temperature (330K) to sample a broader conformational space. Also, I would run some sampling simulations with a range of ion vdW radii (-0.4 through +0.4 Angstrom) because parameterization requires sampling from the forbidden region. From each of these simulations, I would extract on the order of 10-100 snapshots for each of several cluster sizes (1 water, 6 waters, 18 waters) and run QM calculations on them.
You should try interaction energy calculations because this would not contaminate the objective function with the errors in the water model. Another option is to do gradient calculations but only for fitting the force vector on the ion.
I'm inclined to go with the first option, and you can do these calculations in Psi4 which is an open-source quantum chemistry package with good facilities for interaction energies. I have some code for parsing Psi4 output which I can include as part of ForceBalance for you to use.
My guess is they should all be different targets with equal weight, because if you include the repulsive targets with everything else, they might dominate the objective function without you knowing.
Later on, you can reduce the weights for the repulsive targets in order to improve agreement for the attractive region. After all, it doesn't have to be so accurate in the highly repulsive region, it just needs to be repulsive.
Thanks,
- Lee-Ping
