2025-08-05 FF Fitting Meeting

Sage 2.3 updates – LW

• Common thread of phys prop issues in 2.3.0RC1 is amines

  • CB - release candidate is worse with neural net charges than AM1-BCC?

  • LW - (missed something) … this is only a release candidate for now

• Last meeting: benchmarking other candidate to see if some valence were overfit

• looking at “v1” and “v6” in fitting iterations

  • v2 is relatively simple, v6 has more added terms

  • differences in benchmarks vs QM is relatively small

  • v2 is better than released sage in density, v6 is a little worse than v2

  • similar story with (?) hvap

  • … might have been overfitting to gas-phase data with amines, also a valence refit improved density

• CB: Note that I hand-tuned amine BCCs to get solvation energies correct. Many comments on this over the years

  • LW - NAGL should have reproduced the hand-tuning since it was fit to AM1-BCC and not underlying QM. Worth looking into / verifying

• CB: Before hand-tuning, secondary and tertiary amines always had bad solvation energies. Hand-tuning was needed to get SFE correct

  • LW: Don’t know if there are enough amines in FreeSolv. But the tooling should be there from OpenFE collab

• LW: Skipping FreeSolv results, very comparable across different force fields

• MNSol (lots of non-water solvents) slight regression compared to released FF (2.2.1) across entire data set

  • CB: Are the errors systematic, just with a few functional groups, … ?

  • LW: Every chemistry is a little different. Generally a slight shift in getting worse

• LW: Suggesting using “v2” as RC1 instead (currently “v6” is RC1) because

  • Lower error in amine density

  • Lower error, in general, on density and Hmix

  • Similar MNSol performance

• JW: v2 vs v6, not sure which should be the RC

  • LW: Would want to go back to 2.2.1 with NAGL charges, which could be better than any of the refits

• JR: Amides in training set?

  • LW: Based on curation, need 5+ for every functional group. Probably 20 or so, not sure

• LW: Breaking down benchmarking results by functional group. Re-curating training set to better represent what was in 2.0 training set, previously some were trimmed out

  • Added some esters, …

  • Keeping high-viscosity molecules (glycerol and some amines) out of training set because very hard to work with in Evaluator

• Current patches compared to canonical Evaluator + ForceBalance

  • Equilibraiton time bumped from 12 ns to 200 ns

  • Hydroxyl hydrogen likes to go to 0 (but this causes numerical instability) using a softplus at 1e-5 to prevent this

• CB: When you find outliers in benchmarks of a ff fit, are they because functional groups are not present / heavily underrepresented in training data?

  • LW: Some common throughs - often nitrogen-containing, particularly amides, and sulfur. Have not looked super closely yet

New charge model – Shehan Parmar, Jesse McDaniels @ Ga Tech

• Developing force fields for ionic liquids, battery/desalination applications

  • Integrates with Atomate2, paper just published

• Trying to get high-quality (compared to DFT/many-body methods) electrostaic energies with point charges

  • Compared to SAFT, looked at RESP and was good on some molecules but not others

  • Looked also at MPFIT (multipole-fitted charges)

  • some math - A.J. Stone - to get multipole moments, which you can get with Psi4+GDMA

  • another step to get point charges, works similarly to RESP

  • in general, polarizable FF developers focus on working well with polarizable MD codes. Not much work on getting multipoles working with point charge MD.

• Working on a fork of Recharge, modified it to use GDMA instead of ESPs from Psi4

• Looked at 5 molecules in ZINC20 dataset

• MPFIT comparable in computational cost to RESP, similar (but hard to say for sure) accuracy vs. QM ESPs compared to other models

• Not sure where to go next in terms of training, benchmarking data

• BSwope: Thought AJ Stone had a couple of methods, one which was less sensitive to basis sets

  • JR: Lots of history and work on this, Ponder with AMOEBA

• CB: Looking to optimize for small or large molecules? Conformer dependence desired or not?

  • SP: Not sure

• BS: Can SAPT-based fitting include implicit solvent? Want to get condensed phase right too if possible

  • SP: Don’t think people have done implicit solvent SAPT

  • CC (in chat): Not really compatible at the theory level

• JW: Really cool to see people using OpenFF software without our direct involvement

• JR: Question about better SAPT theory. Might be worth looking at SAPT-DFT, which might be better than SAPT0

• PB: Have you looked at relative energies of each component?

  • SP: Goal is to get something for a force field, so electrostatics are the focus

• LW: Where are vdW terms from?

  • SP: Using a buckingham-type potential here, using methods to get parameters from SAPT (?) paper from 2013, Jesse McDaniel

• LW: Would buck parameters need to be adjusted with different charges?

  • SP: In general yes