2025-10-21 FF Fitting Meeting

Splitting N vdW – MT

• (4) LW: I'm not surprised that refitting vdW without refitting valence degrades performance on QM benchmarks. Fitting vdW changes the 1-4 interactions that are coupled to torsions.

  • DM: and the ddEs are pretty comparable by eye.

  • CB: the RMSEs is about geometry change, the ddEs are about the energy changes, but these are all conformational energies in vacuum.

  • MT: yes, this is Industry benchmark.

  • CB: would expect greatest effect on SFEs/densities etc

• CB: What’s the motivation for splitting the nitrogen LJ parameter?

  • DM: One motivation is getting more team members to learn how to run parameter fits with our tools

  • LW: Split was motivated by previous work on virtual sites (5-membered heterocycles aren’t recognized as aromatic), on TGokey’s work on dimer fitting, and SRiniker’s work on dispersion coefficients (DASH)

• (Physical property benchmarks worse) LW: Why is n different between the first and second group of results?

  • MT: First group is from LW’s 2.3.0 training repo, second group is from my work

  • LW: It would be helpful to do a comparison using the same dataset

  • LW: You should also split densities and enthalpies of mixing, since they have different dynamic ranges

• (N Parameter changes) CC: it’s strange that both parameters have smaller epsilon after the split. Were there priors on these?

  • MT: I used LW’s fit settings, I think from 2.3.0 rc2

  • LW: For rc2 I did not use any priors on epsilon

• CB: (discussion about training set for Sage)

  • MT: I was surprised by magnitude of change, but they both changed in the same direction away from Sage. I expected them to go into opposite directions.

• (Discussion on SMIRKS string) – would hit guanidinium, ammonia

• DM: this isn’t ruled out as a bad idea yet. If you want to confirm it’s a good idea, a possible idea is to do the valence refit. All of these seem like reasonable directions forward

• LM: Would like to see
  * Hvap scatterplot on training data - there should be an improved tail (?) as was seen on previous vsties work

(General)

1. More analysis on existing data

   1. Make sure comparisons to Sage properties are like-like

   2. Split out HVap and density (scatter plots)

   3. Breakdown by functional groups

   4. Previous work showed bad HVap on nitrogenous compounds mixed with water

2. Retrain on larger data phys prop data set

3.