new fitting we larger set, ~170 molecules, mainly AlkEtOH
Gradients are now an objective, no hessians(?)
Rest of procedure is similar as before
CBn: Why one conformer per molecule?
TG: Molecules are small, that is why only one conformer. Standard settings for conformer generation
CBy: What is the purpose of 3-4 Ang cutoff in the conformers?
TG: For the simple molecules only small number of molecules came out.
CBy: Relative energy differences between conformers are informative. Why only few alkenes, but many alkanes
TG: This is what is in the dataset (Alketoh)
CBy: The parameters (slide 3) are two categories. One that have enough data, and the ones that don’t.
Splitting of a1 is important, same result as last time. It is split multiple times. See slide 4
four membered rings seem to drive the splitting here, see splits 4,5,6
CBn: Can your method recover something where we exactly know the result?
TG: Something I’m working on. Will do later, needs lots of compute.
CBy: SMARTS strings representation seems to be very specific for dataset. They will fail badly with other molecules.
Looking at “variances” of the parameters.
Splitting of a1 reduces number of bits-represented and No. of primitives
CBn: Additional parameters will almost always make ff better in the training set. Will be specific for the training set
DM: Might try to evaluate FF on validation set during the fit
slide 6: not completely clear how the scoring works right now. Needs some more clarification
DM: Maybe a small write-up would be helpful
TG: Yes, good idea. Was planning to do so.
CBy: The amount of possible splits and chemistry is very huge. Can you find common chemistry in splits. Need a way to divide space in an informative way.
Next steps: only alkane dataset, make a case where we know the answer and see if we arrive there. Start with one bond and one angle. generally improve experimental design.