SPICE2 valence fits with torsion types generated from the training set smiles | FC will post slides here (15 and 16) (general understanding questions) (16) DM – I wonder if this means that, during fitting, we shouldn’t just look at whether the loss is still decreasing, but rather whether the parameters have stopped changing. FC – See slides 21+22 CC – Interesting that the result here is that you can double all force constants and get the same result on benchmarks DM – Did you look at energies as well as geometries? FC – Yes DM – Maybe means we need more sensitive benchmarks CC – Allowing propertorsion magnitudes to become negative? FC – Yes DM – CC – Could be that you’re adding energy in k1 and subtracting it by making k2 negative. DM – On slide 22, this is RMSE in minimum energy geometries. CC aid we may be adding too much detail. Another idea is that you’re doing a good job fitting the full torsion profile, but was getting minima right using fewer terms. So if you’re getting the barriers right … FC – This is RMSE over full scan
JW – That improvement on slides 18 and 19 isn't just "a little better" - That's a larger margin of improvement than we see between several mainline FFs DM – Agree. Wonder about next steps here. Do you have all YAMMBS plots for this? FC – Yes DM – If this were made as part of the 2.3.0 effort we’d want to release this. Do you see deficiencies in this FF? FC – Could submit this on YDS? JW – Yes, please submit on YDS.
FC – I’m going to run some more experiments with eg. ring membership in types. JW – Should torsion profile magnitudes be measured as the integral of torsion profile (with no other energy profiles)? CC – Two ideas for making the crude typing scheme better: 1) count number of molecules matched, not number of torsions matched. E.g. for the central bond in butane, you get one C-C-C-C, four C-C-C-H, and four H-C-C-H, which will underestimate the importance of the C-C-C-C torsion. 2) Include ring membership in primitives. This was the most important split from Trevor Gokey's work on BESMARTS. DM – oooh yes, I like both of these a lot. Count MOLECULES should be much superior than counting number of torsions. FC – I’ll look into this DM – … FC – That’s the philosophical goal… But right now we just look at the atoms involvedin the torsion and not much morethan that, so little consideraton of environment outside that, which could hit trouble with eg aromaticity.
|
Video Conferencing, Web Conferencing, Webinars, Screen Sharing