DM – If I were skeptical and said “these FFs are no better than the starting FF”, what’s your best counterargument?
TG – I don’t have a starting point, so the fact that this gets close to our multi-generational FFs is impressive. This also allows us to iterate on different approaches to chemical perception.
PB – So this’d be fairer to compare to MSM values?
TG – It’s closer, but MSMs are trained on Sage torsion profiles.
DM – So the roche set one - how much chemistry does that cover? Is that potentially a general FF/could that be fit as a general FF and tested?
TG – I’m not super familiar with the composition of the Roche set. I might include a larger set/take the sage optimization sets to use as input.
DM – It’d be cool to compare a full run of this to a “real” FF
LW – Yeah, if you could output this as an OFFXML we could plug it into the other fitting machinery.
TG – That’d be cool.
LW – Any thoughts on why torsions are underesimtated?
TG – I’d love to talk to Cole group about this.
BW – I’d tried using the algorithm from the paper, ended up with really high force constants, I thinkn there was some interaction between nonbonded.
TG – Might that be because you were forcing the sage periodicities? My BAT result showed some things exploding for certain perioditicites. But this was better with lower ones. I need to move on to other work soon but I could include some new datasets if folks send me the dataset IDs.
LW – How long does it take?
TG – MM hessians take about 1 min per molecule, so the current fits took me about 30-60 mins.