DC: Extracting bonded parameters from seminario method in the context of triple bonds we discussed.
Average error in predicting frequencies is not so good with seminario method but we modified the method and we found the errors are muhc lower. Since it is all coded by in Qubekit Josh used QCA molecules and formed parameters.
JH: Here is the jupyter notebook (link?) in which I used the Gen 2 optimization datasets with hessians and used the new schema workflow Simon & I created to make the optimization_set and used ModSeminario() method from Qubekit and calculated the parameters.
From the plots we can see some cases where there is huge disagreement between 1.3.0 parameters and predicted from modified seminario.
DLM: This is great, I have lot of questions but wait till SB goes over his results.
DC: We didn’t have a solid plan yet on how ot split parameters based on these differences.
DLM: TG is working on it, one method is using gaussian mixtures.
DC: Yeah, we can handover the data to TG then he can look at it.
CBy: Is there an overlap with Hyesu’s work on internal coordinate hessian?
LPW: It seems like sometimes the forceconstants directly from hessians would restrict the values, may be we can use those as initial values.
SB: (Link to slides)
SB: Vibfreq data:
approach1: use HJ’s ICH
approach2: use k’s from modseminario() and optimize on opt+TD with heavy priors
CB: What I like about the modified seminario method is that we are drawing out the parameters from every single molecule separately which represents an extreme and on the other extreme we have the general force field which is a global optimum.
We see the value of modified seminario and how are we generalizing it, one puzzling thing is the force constants of triple bonds, with higher force constants may result in resonance.
DLM: We may have an initial set of parameters from modified seminario and override manually if we know of edge cases.
CBy: Another thing is getting the k from equilibrium values which would be a problem in case of a mix of a four membered ring and linear chains, I would be interested how these cases are dealt?
DC: That might be a place for the human to step in and say this angle should not be 90 (as in cyclobutane)
CBy: Yeah, my worry is that modified seminario might give us a different value in the case of cyclic and acyclic molecules and we should find a way to filter out those.
DC: Yeah, TG’s work may help in transferring the parameters.
CBy: It would equally be relevant for congested molecules and restraints induced by rings.
SB: I agree, I think we can do this experiment quickly.
CBy: Yeah, that’s good.
I liked the distribution bars shown by Josh in his jupyter notebook, those results could be really helpful in splitting.
DLM: SB, what’s the next step in release?
SB: Take the modseminario and do one more test fit and benchmarks and this would become the next release candidate.
DLM: I would propose to include a manual filter and check cases where the parameters differ by a large value and we are doing right.
SB: Sure.
DLM: TG, can you try spitting based on modseminario()?
TG: Yeah, I can do splitting based on forceconstants plugging this.
JH: Here is the notebook
(missed some discussion due to patchy connection)
TG: Regarding the rmsdfilter on conformers I am doing an energy based weighting instead of pruning based on rmsd, I wonder if that’s helpful?
SB: Would consider it.
TG: Another thing is on the priors, whether we should take a look once again since we are making these huge changes.