2025-10-08 Cole Lab Check-In Meeting Notes

DC – Possibility of at-risk compute

• Yesterday, I heard that someone lost GPU access for biomolecular simulation testing. We currently have alpha access to Newcastle cluster. It’s possible that we’ll get exclusive access to 30 GPUs.

•  

FC – SMEE-SPICE2 transferable valence parameter fitting: torsion-angle coupling and other experiments

• DC (6) – … SPICE2+EDKTG has more 400 more torsions that SPICE2. … Pink line has lots of torsions…

  • FC – Algorithm can walk in/out and increase/decrease specificity.

• DC – Were theta_0s set to equilibrium vals?

  • FC – No

• JW – Could you explain this functional form?

  • FC – (walk through functional form, see paper)

  • AA – … Suitability for alkenes? …

  • DC – I’d envisioned this making more sense for protein backbones

  • …

• JW – Balance of trainable parameters to number of training data points? These results look like they could be

  • FC – 20k mols in training set with 50 confs each. Could reduce parameters to something more like Sage types

• DC –

  • FC – Training improved a lot on forces, but these metrics are all about energies.

  • DC – Could happen if you have a slightly implementation in SMEE vs. YAMMBS.

  • FC – Yes, should double check. I’m more confident in YAMMBS implementation than SMEE. Did some energy testing but I could do more.

  • AA – Checked force equivalence or just energies?

  • FC – Just energies.

• DC – Shared theta_0

  • FC – Parameters in smee get stored in a broken up way, each angle parameter… becomes its own potential, and there’s not a straightforward way to couple them… smee parameter hierarchy is complicated. Could have done it a different but more brittle way, but we should do more testing.

• JC – Applied to FF where torsions were enumerated?

  • FC – Yes, applied to one FF with 2000 torsions.

  • JC – Generated this coupling term for all possible torsions+angles, or just the ones that looks bad?

  • FC – I did it very naively, just enumerated the angle coupling terms from the torsion smirks directly.

• DC –

  • FC – I think there may be overfitting here, some of the test set is far (geometrically) from the training set.

• DC – So let’s run with the SMIRNOFF types and pin angles to a reasonable physical value

•  

  •  

Bespokefit-smee update

• AA (19) – Have you looked at how ML potential does compared to MM?

  • FC – Yeah, you see hydrogens in the amides pucking away from planar. However nothing jumps out as being massively different.

  • AA – If you’re fitting to the energies and forces from the ML potentials, what is the ML potentials themselves are getting it wrong?

  • FC – Someone else had done a benchmark of EGRET1 on amides and found good agreement.

  • DC – But could be good to sanity-check.

• DC – Re puckering - Do impropers come intoplay here, what do we do with those?

  • FC – I think we keep the types from Sage, but differentiate them by SMIRKS(?) and train them(?). Also worth mentioning that the geometry optimization routine is different - I don’t restrain/constrain torsion angle the same way.

  • DC –

• AA – It’d be good to also check the distribution of torsion angles sampled.

• PB – Aimnet2 may give some wacky torsion profiles for simple molecules

• Open

  

   

• PB – GPU performance of SMEE training for full transferrable FF. When you train to SPICE2, how long does 1 epoch take on a GPU?

  • FC – I vaguely remember that a full run takes ~4 hours on a A4000 with 20GB of vRAM. Maybe ~100 epochs?

•  

•  

•  

Plan scheduling for meeting between European/ US daylight saving

FC: Not required as no meetings scheduled for this week.