Co-optimisation of AM1BCC and VdW Parameters
DM – Part of the plan was to bring aqueous mixtures in, so that we include a specific water model in our training.
SB – There is a possibility of including aqueous data here. Only conern with having fixed water models is that we’d systematically train in compensation for deficiencies for water model.
CB – Alkynes are missing in training set
SB – IIRC, there isn’t much alkyne data in thermoml. It has lots of petroleum/refrigerants, but not much alkyne.
CB – Freesolv should have alkynes. Nitriles are also a good sort of weird chemistry to go out of our way to include
CB – on the question of water mixtures: alcohols, alkenes, alkanes, and aromatics – I wouldn’t expects them to hange a lot when in water. But esters, ketones, and ethers – those should have big effects when mixed with water. If you include these mixture components, you can capture the effects of having implicit polarization in the BCCs. O the other hand, if you fit one BCC for C-O bonds, you’ll get divergent gradients for aqueous vs. aprotic solvent mixtures.
SB – Good point
CB – Would be good to do fits WITH aqueous mixtures, WITHOUT aqueous mixtures, and with aqueous mixtures ONLY.
CB – MKG – Do you have anyone who can pursue a BCC refitting project?
DM - refit hydroxyl lj parameter and fix the issue related to hydroxyl group will be meaningful.
MG - practicality of the fitting. correlation(near-linear dependency) between lj parameters(9 for now) and bcc parameters(~40) will lead a complex problem, chance to trap in false minimum. how to reach a meaningful minima
CIB - two strategies (1) use restraint so that it only allow limited movement (prevent unphysical final parameters) (2) usage of gradient of restraints
SB - same restraining weight or different weights for each parameter?
LPW - starting with the same prior width for all
CIB - starting from simpler vdw parameter set(MG’s suggestion), data-driven decision on how to differentiate the parameters
DM- step-by-step exploration of chemical spaces like we do for valence parameter improvemnt
CIB - MG had multiple minima, have you identified two or three clear minima that they can start with?
MG - it might be better to start from where we are now, rather than starting from multiple minima
LPW - global minima searching scheme can be applied, like basin-hopping
OM, SB - how many conformers should we use?
CB – the “10” in “ELF10” was determined almost entirely by availability of computational power, not any underlying theory. More confs is always good. Use omega to generate conformers and input the conformers to elf10 will choose which conformers to use
DM - ESP grids are already stored or will be stored in QM ARCHIVE, will ask in qcarchive channel.
CIB - to save CPU time, one thing can do is to get the conformers out of there use conformers as input of elf10
SB - level-of-theory for esp calculation, how about for biopolymers?
DM - compatibility issue. some of the small molecule ff might be refitted.
CIB - these are non-bonded parameters , which describes interaction between small molecules and biopolymers.
LPW - esp fitting module is implemented in fb attempting to assign charges to molecules. the feature should be improved to fit bcc.