CBa: Are you refitting smarts you found before and refitting?
TG: I ook the whole release fit from 1.2 and my two parametersn (i.e. my alkane parameters) and started from there.
CBa: ok . what are the 3 different plots? (slide 4)
TG: left, percentage improvement of obj function. middle is penalty function, right raw objective units
TH: What about references for splitted parameters?
TG: References are inhereted from previous parameters.
TG: slide 5 shows some important parameters. Angle usually showed larger improvements than bonds,, that also shows up in the force contants
TG: (slide 6), applying my clustering technique to 1.2
It seems that vehicle is rather redundant, other datasets (1.2 sets) seem to be more diverse
DM: Wanted to avoid vehicle set, runs risk of being selectively good for heterocyclic rings
CBa: Do you have a measure of how many smarts are selected by the ff?
TG: currently not
DM: we probably need to split a lot of parameters in the vehicle set
CBa: maybe torsion interpolation can help at some point
DM: vehicle has not many rotatable bonds
https://github.com/openforcefield/qca-dataset-submission/blob/c1fb524baac8dcf16c7d88e168e9918856a1f8e9/submissions/2019-07-02%20VEHICLe%20optimization%20dataset/optimization_inputs.pdf
This dataset is mostly for heterocyclic rings.
CBa: We would have lot of trouble with inferring aromaticity.
DM: How much time does it take for a run?
TG: 2-3 weeks minimum.
DLM: Doing a diverse chemistry, or doing chunks of targeted improvements. You already showed improvements with alkanes.
CBa: Yeah, we thought you may arrive back at the alkane parameters we already have but your work showed improvement over current set.
CBa: Add oxygen, and allow aromatic and normal carbons, from the PhAlkEtOH dataset, like three atom types.
DM: You could also pull out from a larger dataset but for now I think extending to Oxygen is good and let’s not open a can of worms by going all the way.
CBa: AlkEtOH was used in my smirky work and we needed more bonds and added more molecules to account for that.
TG: Yeah, I will checkout molecules from QCA.
DM: If you have too much data you may miss some drawbacks of our work and so focusing on a small subset would help us workout any kinks we may not yet know.
TG: In terms of targets I can be selective but combinations of parameters is difficult. I don’t want this chemistry in the new parameters is a bit contrived.
CBa: My worry is that we scale this up, run for a month and we get bad results. So, a careful balance with a small set is good.
TG: Yeah, study certain chemical spaces and go forward is the best way.
CBa: I wonder if we should include r3, r4 together with your current set
DM: CBy’s thought was to keep these legacy parameters. We had a bad puckering with r4 and we brought in more r4 parameters.
CBa: Yeah, there was a lot of manual work while transferring from atom-type land to smirnoff in the early days, so I am not sure if we really need some of these parameters. It’d be an interesting experiment to see whether this messes up the entire FF or not if we remove r3 parameters.
