CB: When pickjing parameters based on eq. values. You seem to say that this does not work that well.
TG: Yeah, it is not working as well as expected.
CB: Agreed, understand better now. Agrees well with my intutiion.
CB: Whatever you are doing from scratch seems to do as well as sage (looking at lines 4,5 slide 2).
TG: yeah
CB: From scratch means not from eq. values?
TG: Yeah.
CB: first three lines say, working from eq. values does not work well. line 4,5 are interest.
TG: Could you tell us why you think it matches your intution?
CB: Looking at cyclopropyl, molecule is forced in triangle. Doesn’t matter what parameters are. If you look at angles, steric interactions can affect angle.
TG: Tried to split on frc constants as well. Did not work, and angles were essentially splitting force constants.
CB: The real splitting has to happen on frc constant and eq. angle. Like the idea of using derivatives to split parameters. Your current approach is very exhaustive. And I think it will find the correct answer. At some of point at the end of your runs look at the derivatives and try to find populations based on that.
DM: Are the periodicity taken from TorsionDrive?
TG: No TorsionDrives here.
CB: Are you optimising these structures as you go?
TG: These are based purley on QM.
CB: Your obj func is only energies?
TG: These are Torsion eq. data. I am not using energy per se.
CB: Are you relaxing structure on MM
TG: There are no MM data. Splitting based on geometries. Targets are only forces and geometry.
CB: You’re starting with QM geometries. Are you allowing to let the MM geometry change?
TG: Yes.
CB: On the torsion side even for a smaller set of CHO we see 35 torsions, kind of like a bespoke fit, can we figure out general parameters from these torsion splits even if we see more bespoke parameters on the side?
TG: Yes.
Slide 6:
CB: regarding b2, b3 consider five membered heteroatomic rings, like furan, etc., there would be difference in double-ish (short) and single-ish (longer) aromatic bonds, if we look at molecules then we may make a distinction. From your table the force constants and bond lengths seem to be in line with the single, double, triple bond expectations.
TG: Still working on human-readable angle definitions, for now there are too many wildcards
CB: This looks good and seems to follow the hierarchy.
DM: Let’s look at the final results after the fit from scratch ends, then we can test it.
CB: TG triggered a thought, given a forcefield we want to improve the obj. function by creating more bond/angle params or more torsion params, seems like a choice we have to make at some point.
DM: Looking at the multiplicities of torsions that are being used in different contexts.
CB: From numerical standpoint trying to make a six-fold torsion work in an eight-fold or other context is concerning. Even if the chemistry overlaps this is still a numerical problem.
DLM: There are 25+ torsions that have this problem of mixing multiplicities.
CB: We have built-in multiplicities for the chemistries, and we need to make a fix for the numerical problem.
Add Comment