Internal coordinate Hessian fit status | LW: under the impression this was shelved due to suboptimal results CB + DM: Surprised – it seemed quite promising PB: didn’t perform better on industry benchmarks for ddE, RMSD, TFD CB + DM: may not be best targets to evaluate ICH results – would expect improvements to show up in less-represented chemistries that may not be in standard set
|
PB – why harmonics for angles and bonds? Why not Morse, etc?
| CB: Old guy history: agreed as a field to prioritise generality and a simple functional form Modern history: currently only OpenFF is simplifying and exploring the science and cheminformatics to the point that we may be reaching the theoretical limit of the current functional form. At some point, the functional form will be the limiting factor. Not sure if we are yet. When we should, we should have anharmonic bonds and angles. Morse is good for bonds, not sure if it’s good for angles. I proposed to LPW some time ago a 2-parameter form – this is preferred over 3+-parameter forms for simplicity
|
LW – any other topics for the annual meeting? | PB: WBOs had decent results for biphenyls but not outside a specific dataset. Also there were issues with the backends and protonation states CB: is this a negative result or just not yet good? (paraphrased) PB: outside biphenyls the backends are a bit confounding, are we ok with different force constants with different backends? DM: PB has found some series where it works reasonably well, but there’s a large swathe where it’s challenging because WBOs don’t change that much so it might not actually do much. And also backend problem CB: the first problem is more of a training/test set problem PB: also xtb adds another dimension, if we use it CB: maybe we should move from AM1 onto a new efficient method
DM: Torsion multiplicity problem (JM’s work) The force field has some torsions where central atoms do not have explicit valence, or have multiple possible valences, the multiplicity becomes muddled So the barrier heights can be wrong because the barrier heights are applied at the wrong numbers JM has been enumerating cases where we mix multiplicities, and making them explicit, then re-fitting. However, we don’t have TorsionDrive data for many of these new torsions First pass with Optimization data shows clear improvement. Should probably show this. JM: now generating new torsion data
CB: JM’s impropers status? Advancing from planar to out-of-plane? DM: there’s still more to do – PB is re-fitting them Still in progress PB: interaction of other parameters? CB: running into problems with sulfoxides CB: PB found that there’s a significant steric interaction with torsion term, e.g. 1-4 torsions around ring
|