MKG - Willa working on these models. Does it fit with OpenFF. Can we make it fit with OpenFF if not?
Couple ideas:
Have compact set of polarisabilities.
Start off with ESP you like, then impose minimal polarisabilities, then refit.
MS - fixed charges behave better with polarisabilities added.
MKG - can be used to do a RESP like fitting.
JC - Same as AMEBA?
MKG - Similar to Coleman?
CB - AMEBA is fitting multipole as well.
Looking down the road, once you have v-sites may behave similar to also fitting multipole.
Can iteratively solve for BCCs in addition to just point charges.
MKG - RESP like method easier to do as less fitting. Start with that. Then move onto BCCs
Would want to integrate with what OFF is doing w.r.t to BCC.
Not sure how best to fit it together, want feedback from the group.
JC - Fitting BCC + polarisability simultaneously could be exciting. Could overcome conformer dependance and move away from need to include phys-prop data in fit.
MKG - could then get more meaning out of fitting to dielectric constant data.
JC - big question - how to come up with parsimonious set of point charges?
MKG - Michael and Willa thought about this
MS - You suggest to just put point polarisation to just some atoms?
JC - the question is - is that feasible? Will it still generalize?
MKG - parsimonious in terms of types or not putting it on all atoms?
JC - both?
CB - polarisability on even something like carbon not small and even response not small. 1) Couldn’t find an atom on which it wasn’t important (maybe H?) 2) In terms of parsimonious, MS came up with charge model independant charge model. Didn’t need co-optimize and they are good.
JC - How can they be optimal if they had not optimized.
MS - just needed to optimize once but then generalize.
JC - that sounds sub-optimal.
MKG - still thinking of typed polarizabilities?
JC - yes - don’t even need to elemental? Can be as elaborate as you want?
MKG - don’t necessarily need the added complexity.
MS - fitted to polarised and not polarised ESP, …
co-optimize and sequential optimization yields similar accuracy over multiple molecules.
CB - physics of polarisability is independant of physics of charge model. So why co-optimize?
JC - May depend on the type of polarizaility?
CB - MKG actually championed this. Can make polarization model charge model.
If co-optimising will have strong linear dependencies which want to avoid.
MS - Can’t make charge model which is independant of polarizability, but the inverse is true.
JC - Did you use polarisability on every atom? Still hold up if you omit atoms?
MS - None of the polarisability parameters are small (maybe only H)
CB - Can maybe go to united atom, but couldn’t omit polarizabilities on non-H atoms.
MS - Not just elemental in my case, needed specific types for aliphatic and aromatic.
MKG - Can do quite will with direct polarizability?
MS - Can’t use the same parameters.
SCF more accurate in terms of physics rather direct, but direct is much faster.
JC - issues with have SCF is implemented. Some implementations not statistically correcty. Important to get correct otherwise wrong results.
Can also do fixed iterations - solves for 99% if this and still has stat mech working.
MS - big fan of direct polarisation. Goes in direction of BCCs. Maybe some accuracy loss (not neccessarily though). Doesn’t think gain much by using SCF.
MKG - Hopefully somewhere in there something useful for OpenFF. Would appreciate feedback on priorities? What’s the low hanging fruit?
Or should we decouple?
CB - Want to commet - do ESP fitting easier than BCC fitting? ESP fitting has easier infrastructure but has more linear dependencies. More difficult to come up with consistent set.
AlkOEth has less than 10 degrees of freedom to fit for BCC. Each is very well determined as there are many cases informing.
ESP has separate degree of free for each atom. Still many DoF to fit.
Because of linear dependance, all charges tended to end up different.
BCC - fitting easier, more applicability, chemical understanding greater.
MS - not sure BCC fitting is easier? Need a lot more molecules? Need more than one molecule per BCC. With ESP fit, can just consider single molecule.
CB - right on all counts.
Got many great training sets however. But in general comments are bang on.
MKG - talking about BCCs in just QM setting.
JC - could leverage some condensed phase.
From a practical perspective, BCC scheme which works with polarisability could be very novel and immense value for OpenFF.
Could use in docking for quick predictions. Could be significant improvements for free energies.
Especially if use direct polarisability the cost is almost negligible. Very widely supported.
MKG - agreement that DirectPol route has value or should definitely include. Good interest in AM1Pol model.
CB - or at least a different level of QM theory. V-charge? Very interesting.
MKG - for Willa’s easy stabs at this will be most practical to optimize the BCCs to gas phase ESP.
MS - first step make co-optimize with phys prop with fixed model.
MKG - eventually should allow for co-optimize of phys-prop with Pol model.
MS - first pass Willa should focus on pure QM. Big fan of doing one step at a time. Otherwise too much complexity.
MKG - Just wanted to know where this is headed.
CB - like the idea. Went to BCC from ESP and RESP fitting. Ultimately want to see DirectPol with BCC which is general for organic chemistry.
Other value is, started off from doing all pharma chemistry so big use case for it.
Early on can we come up with a direct pol model even if it’s not good but covers all chemistry - including halogens and not just C, O, N, and H.
MKG - kind of the way we are headed.
CB - in your discussion you talk about ESP - in my experience electric field gives better charges.
MS - CB work makes it seem electric field more advantageous.
MS - problem is we need a lot of points.
SB - generating it on the fly.
MKG - MS most of your code is potential based? How much work to make field based?
MS - Not sure - will be more complicated but not too hard.
CB - In respyte can generate different types of grids.
MS - not that complicated as independant of the actual electric field from the point charges. Probably just need to add another vector.