2025-04-23 FF fitting meeting

Current NAGL FF status

LW

• LW will link slides here

• (slide 5) BM – LW, you’d mentioned this earlier, but I haven’t had time to look into this. Should I report this to NIST folks?

  • LW – Yes, let’s coordinate, and if this is real then let’s let them know

• MT – Was this also in the Sage 2.0 training?

  • LW – No, just in the most recent refit. BM and I have tried to include some more data in our runs.

  • MT – This torsion (Br-C-C-Br) seems fairly rare and is likely underrepresented in our data. Might have some confusion with aromatic rings.

  • LW – Right, and aromatic rings would get a different torsion.

• (slide 7) DC – Which fitting software are you using for MSMs?

  • LW – Qubekit

  • DC – I might ask you for help on how to install, we have a user writing in about that.

• (slide 10) JW – This is great, looks like everything is substantially better than 2.2.1

• JW – so, should we go ahead with a possibly problematic training set, or go back to the drawing board?

  • LW – Want to look at phys prop validation to decide. If performance gets WORSE than 2.2.1, then we go back to the drawing board and restart from the phys prop fit. Otherwise we move forward with the valence refits if it doesn’t look too bad. BM, what do you think?

  • BM – I think this is a good plan. But could you clarify what we’d do with the brominated compounds? Exclude them and rerun?

  • LW – Thinking is that, since it’d take so long to refit to phys props and we’d miss release target in May, that we’d move forward with potentially problematic fitting data as long as things don’t get worse. But we’d plan for a refit immediately after either correcting or excluding these data points.

  • BM – Ok, on my side I’ll bring this up with NIST folks. I’ll run a draft of this by you before I send it.

  • BM – I’m looking forward to evaluator release as well with pre-eq functionality. I’ll let MT know on slack if there’s anything else I need from evaluator

• DC – re slide 12 - what is meant by “optimization”? Trajectory or last snapshot?

  • LW – The latter, snapshots with energy and force.

  • DC – And you can’t grab snapshots from SPICE?

  • LW – I could grab them from SPICE, just used this dataset since it’s familiar.

  • DC – We’re starting to see a gap opening up in the benchmarking curves using smee, so may be worth considering. Also TorsionNet500.

  •  

  •  

•  

Reschedule this meeting?

LW

• timing works poorly with OpenFF Advisory board meeting

• JW – Would suggest moving two hours later and one day earlier (tuesday 3 PM US) to avoid mondays/fridays (different natl calendards and long weekends)

• LW – I’ll start a poll

•  

•  

•