Overview methods: Optgeo and Vib Freq targets, overall same weights/denominators, no regularization
Same procedure as previously explained (see previous notes)
Application to butane:
a1 split shows largest drop for obj func
after split, force constant of a1a (the child parameter) get stiffer and bond gets shorter
total improvement for obj func is about 82% compared to 1.2 release
Application to aliphatic molecules:
CBy: Five molecules all that you used?
TG: yes
CBy: Are the only ones with unbranched(? ) chains?
TG: yes, were the only one with hessian data
CBy: Hessians are very local information. Give only info about particular minimum. Should use broader set of molecules and not restrict too much to hessians
CBn: Cis/trans diastereomeric molecules should not be removed from set only because they are too similiar
TG: Agreed
Hessian target leads to very similar splitting as in the butane target
b83 had consistent gradient difference, but actual obj function drop is low
CBn: Were params fitted before splitting?
TG: Yes
Looking closer at splittings.
a1a splits out C and H
a10a splits out H1 bonded carbons and H0 bonded carbons
Intersting, second best split was only because of one molecule CC/C(=C/C)/C
CBy: Splitting and increase of angle is due to cis pattern on the double bond. This split might be due to small size of dataset, but still makes sense. Suggest to use larger dataset, maybe just alkanes (e.g. c1cccc1 c1ccccc1, c1cccccc1, CC(C)(C)C(C)CCC, CC(C)(C)C(C)(C) ). Also stitch hessians.
CBn: Were you looking for one conformer?
TG: yes, just one conformer.
CBy: Suggestion to have multiple conformers for each molecule.
Outlook: Scoring for the splits
[…]
CBy: What you describe sounds like decision tree.
CBy: Could you show us equations for gradients and obj func look like next time?
Add Comment