2025-09-03 Cole Lab Check-In (Aus) Meeting Notes

DC

• DC: Charlie’s NAGL paper reviews incredibly unhelpful. Two comments: incorrect complaint about training set being too small, question about effect of changing architecture. Could address, or complain to the editor.

• LW: Complaint about lack of description of model not justified as well described in the text. Could add a diagram.

• LW: Could also point to ESPALOMA SI/ white paper where hyperparameter optimisation/ investigation of models done.

• DC: Could also just point out previous results with ESPALOMA charge using architecture, which provided robust charge assignment.

DC: MACE people still interested in physical property datasets. Could make condensed phase benchmarks a general benchmark for the ML community.

LW: Specific format required?

DC: No, just pass on to me in current format.

LW: Dataset is similar to Sage, but with a few tweaks.

Discussion of systematic torsion type fit

FC

• FC: How to address alkene issue?

  • DC: Add in single points for subset of simple alkenes?

  • LW: Could run a quick simulation and grab snap shots – we could compute.

• LW: If you want MSM starting point from more data see:

  • https://github.com/lilyminium/ash-sage-rc2/blob/main/03_fit-valence/03_generate-msm/output/msm-force-field-v0-mean.offxml is the FF with MSM values. I've just taken the mean here.

  • https://github.com/lilyminium/ash-sage-rc2/tree/main/03_fit-valence/03_generate-msm/images/distributions/ff-v0 contains images of the distributions for each parameter.

• LW: Have you looked at the distribution of a28 angles in SPICE2?

  • FC: No, but should

• LW: Could use the OMOL models as a quick way to get near SPICE-LOT results to test adding more data.

• LW: JW has code for using QCA TorsionDrive data with SMEE.

• DC: On overall plan – probably stick with initial expanded torsion type set as it performed well, keep ByteDance data as test set, experiment with regularisation.

Updated results

LW

https://docs.google.com/presentation/d/1XAYz52toOHY5HH27EyjLZEE0etRy9d_E8RpPWYk3Hk4/edit?slide=id.g2f3a95763d8_0_883#slide=id.g2f3a95763d8_0_883

• LW: Condensed-phase benchmarking for previous FF which did well on QM data. Previous issues with lack of constraints and linear angles with non-zero force conmstants on rc1 dataset. Current results on rc2 – slightly smaller dataset.

• Force field performs well and no blow-ups. Improves on densities, but worsen with amine densities.

• Break down by functional groups and parameters. Problematic angles - a28 seems to be doing alright.

• Issues with some torsions – amines and diether type.

• Missing results for one validation property.

• Previously flagged formic acid properties as problematic

• LW: Looked at densities first as quite affected by torsions.

• LW: JC looking at splitting out bonds and angles – good to discuss with her.