2025.12.17 Cole Lab Check-In Meeting Notes

FC – Bespokefit-SMEE Progress

• FC will upload slides here

• CC –

  • FC –

• JW – How many bond parameters would be produces for mol on slide 6ish?

  • FC – Roughly one per chemical bond in mol, but considering symmetry it might be about half the number

• JW (14): It’s interesting that Bespoke fit 1 is the second best in RMSE, and second worst in RMSD. Wonder if there’s necessarily a tradeoff between performance on one metric and the other.

  • DC: This was somewhat shown in the paper as well.

  • FC – Right, will talk about this later.

• DC (18) – (comprehension Q, see recording ~14 min)

• CC (18) – I could imagine maybe there’s a quadrupole on the aromatic ring, and that’s interacting with the Hs.

  • FC – I was looking at different contributions to energy/force, will check back into this.

• CC (21) – Possible that Egret just has bad performance on this molecule?

  • FC – I’ll check into that

• DC – I wouldn’t go too deep on the pathological case, and Bespoke vdW is interesting but not practical.

• CC – Did you look at what happens with more MLP opt steps?

  • FC – I tried on a couple examples but didn’t see improvement.

  • CC – This could be exposed to the user as an option

• JAC – Wonder if there’s a way to encode the “this looks weird” human intuition on slide 23 into something automated.

  • FC – I’m trying a few things for this.

  • DM – The problem is that complex torsions may show discontinuity fortuitously.

  • JW – Wonder if it’d be possible to make a “can torsions fix this molecule at all” metric - Like, do an unrestrained torsion parameter optimization to try to overfit to data, and if even that can’t match landscape, then acknowledge that torsions won’t fit this mol.

  • JAC – Collecting more cases of these jumps may help us reason about how to handle tricky cases

  • FC – This is the only major outlier from the ~100 mols in BF1 paper dataset

• JAC (23) – Considering an alternative to differences in slope of energy profile, there may be a way to use this - Something like a Z score of RMSD across torsion scan could be useful while leaving the fortuitous discontinuities intact.

  • FC – That could be cool, was thinking of similar things so may try these out.

  • DC – If the LJ opt for this mol fixes it, that could be worth mentioning at end of manuscript. But not trying to implement general method.

• JAC – In my back to school project, I’m experimenting wiht hyper-specific bonds and angles, and I’ve noticed that bonds don’t improve much, so I was thinking of having bonds have a different learning rate than angles. But hadn’t thought much about interaction with torsions.

  • FC – I did find that things were sensitive WRT projection onto linear harmonmics, had to change scaling.

  • DC – How have linear harmonics progressed in terms of inclusion into code?

  • FC – I have it prepared in a branch, will make a minimal branch to prepare for a PR

  • JW – We will likely soon end up with a fork or transfer of smee+descent in OpenFF GH repo, I’ll keep you in the loop as this may affect PR reviews.