SLIDES HERE
TG: I am starting from a simple set of alkanes and trying to automatically split parameters
CB: So, what you’re showing in the previous slide is that if two bonds have the same parameter and different objectives they are ranked differently?
TG: Yeah.
CB: And you calculate the tier2 scores after optimization and update the parameters in a conjugate gradient fashion?
TG: Yeah.
TG: Splitting and deleting oscillates.
CB: What if the tier1 indication you got for the parameter which really wants to split but cannot because of the small contribution to the objective function?
TG: May be I shouldn’t be looking at the max delta. Some property inside tier1 score is driving the splits and I am still looking for it.
CB: Looking at parameters in isolation might help?
TG: Will think about it.
TG: I am thinking of making this process faster since I have lot of openmm opts to do and someitmes it takes 7 days to get the first split and even with a greedy approach 3 days. I am trying to size down the number of molecules by constructing an adjacency matrix.
For alkanes this approach brought down the total number of relevant molecules to 11 as a part of minimal training set and I am using the rest as a test set.
Adding H decorators I can make these 11 molecules more diverse to pickup all the smarts I want.
DLM: Last time we talked about a multiple minima problem and starting from a varied set of initial values to see how your initial optimization graphs would do in dropping the objective with a better guess? May be the seminario method helps in initial values?
TG: Yeah, I am working on that. And this is why I chose the valence terms but there are no good guess for torsions with seminario.
DLM: Yeah, even then I think it should be a great starting point since torsions are really hard.
CB: Torsions are hard because they are soft. Torsions are difficult because they have much lower gradients and make much bigger changes to equilibrium structures.
We make the bond lengths longer between the terminal atoms it would then make torsions 1-4 terms better but that’s not what we want.
Instead of simplifying the chemical space may be another way of simplifying the problem is to keep the torsions fixed and let the bond lengths change and play with the changes to the objective function.
Then fix those better bond lengths which match closely the chemical reality and change the soft torsions, change barriers, constants, etc., and this simplification may help. The overall objective function mixing everything might be the overall confounding complexity.
TG: Yeah, I agree. That’s how we are kind of doing for 1.3 with bond and angles first and then torsions.
DLM: So, do you mean bond parameters? Also, angles?
CB: Kind of, I was alluding to TG’s previous comments. It is easier to understand the bond lengths and understanding force constants would be the next thing.
I am still wondering how the stiffest degree of freedom, the bonds, which should contribute more to the objective function, and the soft ones, torsions, are connected.
DLM: Yeah, ignore torsions and try it out.
TG: The chemical space is small here I guess.
CB: You have the rings here
TG: Yeah, I can iterate.
