2024-02-29 Force Field Release Meeting notes

Sage 2.2 update

AMI

• Slides

• CB: Why is the optimizer so bad for epoxides in Sage 2.1.0? Why is this parameter so bad? Is our optimizer going to keep going into pathological regions?

  • AMI: the epoxide angle is not due to the tighter priors. Epoxides were being parameterized with the generic X-O-X angle. This has a value of 120 degrees. The O-C-C angle is treated with the proper r3 angle.

  • CB: I see. My takeaway is that either force constants should be approximately equal, or that equilibrium values should be around 60 (?).

  • CB – As we introduce parameters for marginal benefit, are we consdiering any penalty for adding in a bunch of parameters/potentially overfitting.

  • AMI – Penalty is largely applied by judgement. FF expts take about a week so we’re limited in what we can do. When we optimized 3- and 4- membered ring params without including 5, it was messy. That’s why we did the 5-membered ring fit.

  • CB: in ab initio world, if you add more basis sets / complexity, fits improve. There’s a penalty for adding another degree of freedom that we should consider adding.

    • CC: For the record, I think this is the standard approach to address Chris's concern: Akaike information criterion

  • DLM: in this call we’re focusing on what we can release in the near term, but we should take these considerations into account in the future in the context of automated improvements

  • MS – Would be good to focus on in the next round.

• (End of slides)

  • PB – Re internal cooridnate RMSDs - the bond error looks quite low. All of these errors look pretty low

    • LW – Yeah, looking at these Y axes, the dihedral error does seem to be the one significant one. So I think they’re more important. So my vote it for the a7C.

    •  

  • JW: (on slide with 4 membered ring histogram and kdeplots) – the RMSD plot seems shifted more to the left, so it’s better, but the ddEs are counterintuitively worse for the red FF.

    • AMI: I think change in dihedral angle is the major difference between these FFs and could be why the red FF looks much better

    • BW: Is it possible the KDE plot is smoothing things out?

    • AMI: tried visualising as box plot, but the box region was very small compared to outliers. Visualising as a histogram was also not helpful. There are only 1000 data points in this plot, compared to normal dataset being 70k, so that could be a factor.

    • JW: let’s discuss this in internal benchmarking meetings

• Last slide

  • CB: ddEs are important, dihedrals are important. If I can get “less worse” ddEs and “as bad” dihedrals, I would take that at the cost of angles and bonds (LW agrees).

  • AMI: I agree, downside of angles and bonds being worse is also not a change from 2.1, whereas dihedrals are actually a change for the worse. i.e. we’re not introducing issues newly to Sage 2.1.

  • PB: I think the changes between bonds and angles between two 2.2 candidates is pretty much the same when you look at y-axis numbers.

  • AMI: IMO there’s clearly a change, although the scale of it is much lower than dihedrals.

  • DLM: No strong preference for either, but will go tentatively with a7c FF. But we should follow up on this because we would like the best of both. But in the short term, a7c is the safest.

  • JW: I agree with a7c

•  

•  

•  

Finishing JCP review

DLM/MS

• MS: there’s a few things left to do. DLM, how hard is the deadline?

• DLM: hoped to submit it before CUP. How about end of next week?

• MS: we should divide work. Some things we can resolve now

  • Section organisation and flow can be resolved during proofreading/review, we can do that next week

  • History of SMIRNOFF to Sage, very very broad overview of papers and summary of history

  • Software:

    • do we commit to FAIR? If so, how do we? Especially highlight Reproducibility

    • Philosophy on licensing

  •  

•  

•  

•  

Meeting times

Does next meeting time work? LW will send out poll for March/April times onwards

• Pinned to 3 PM US time

• (General) this time is good.