Participants
Notes
PB: I will look into the discrepancy with non-interpolated params fit from the feedback in ff-release call, also check performance on BM set, and chemistries of the poorly performing ones.
DLM: The non-interpolated ones should do better than 1.3.0 because it has more parameters. Start from fit0 interpolated parameters as starting point and see how it goes.
TG: I looked at fit3 output, and some of the k’s flipped to negative values (-1), and some of them didn’t change esp. the double bond and aromatic ones, most of the single bonds were being matched and they were fit.
In the regularization we kept l2 close to one and need to check on that. Also, this is the one that took 4-5 days to finish and ended up in obj value of 6000+.
PB: Reg. one of the poorly performing tids ‘t45’ torsion profile flattens out in the middle.
TG: That might be because it touched the energy threshold set by torsiondrive and still the job is marked as complete
DLM: Yeah, post a question on slack tagging HJ and LPW. It is interesting to see 1.3.0 still has a good trend with those points.
JM: Here are the k, TB Vs wbo plots I made recently and I am working on including the sterics filtering, what would be the next step after this?
DLM: I looked at your data and it is tempting to say right hand side near wbo of 1.6-1.8 have a different slope, and on the left hand side near 1-1.2 have another trend. May be looking at the molecules and after sterics filtering we can get a clear picture.
TG: The color-coded ones are the interpolated k-values?
JM: These are not interpolated, these are subsets of the generic TIG0 k-vals
DLM: Look at the torsion profile for the negative k-value at the right most point near 1.9 in substituted phenyl plot (color coded into different TIGs).
These plots are definitely better than extracting barriers from QM data.
JM: I will filter the sterics, get the torsion profiles, look at the molecules
TG: PB I have some time to help with the fits, so check fit3 and get back to me.
0 Comments