2024-06-26 Newcastle/BespokeFit Meeting notes

Participants

  • @Daniel Cole

  • Patrick Quoika

  • @David Mobley

  • @Brent Westbrook

  • @Chapin Cavender

  • Charlie Adams

  • @Joshua Horton

  • @Pavan Behara

  • @Jeffrey Wagner

Discussion topics

Item

Notes

Item

Notes

General updates

  • JH – QCEngine now supports AIMNET and MACE

  • JH – Looking at submitting 65k single points via qca-dataset-submission. Won’t need wavefuncitons/can just reconstruct using MBIS charges

    • JW – That should be fine. 65k should be quick. Feel free to make them optimizations/submit opts as well. No objections from me.

    •  

Patrick Quoika presentation

(see recording)

  • PB – Which version of the FF?

    • PQ – Sage 2.0.1 or 2.1.0, I’d need to check

  • JW – (method details, see recording)

  • DM – Re: Nitrogens - Agree this would be super interesting. We’re looking at vsites on nitrogens including pyridine. So this could be a cool way to investigate the quality of nitrogen vsites. CBayly thinks there are deficiceis in N parameters that are baked in because there aren’t offsite charges, so adding off site charges could help. We have trouble pinning down deficencies in nitrogens because we generally do our condensed phase stuff at a single temp/in a narrow range, so we bake those into the params. So having a bigger temp range may help.

    • CC – Just looked, and we do have aniline densities and enthalpies of mixing in the Sage LJ training set.

    • DC - I think there’s only one SMIRKS type for N

    • DM – Could share test FF with additional N params

    • JH (chat) – How does the a qubekit version do with v-sites?

    • PB – Did you try a potential trained to phase equil data? eg, Trappe potential

      • PQ- - Haven’t tried anything else yet.

    • CC (chat) – I think we also train only using simulations with neutral species, but the experimental enthalpy of mixing for aniline-water will include a contribution from charged species.

      • DM – This is a good point. Also, Ns in different contexts can have weird geometric effects.

      • DC – My memory is that aniline N is totally flat.

      • DC – Re: neutral vs. charged - There’s a similar discussion on Slack. I thought the experimental data was corrected to account for charged species, but I don’t know what I assumed that.

      • PQ – Just to clarify, these are pure solvents, so there’s no water.

      • DM – Ah, right, so this might be more relevant for our FF training, where we’d have mixtures with water.

    • DM – Re: water errors, we’re headed toward developing our own water model.

    •  

  • JW – I don’t actually know what OpenFF’s temperature range is.

    • PQ – Yeah, I was surprised to see water models doing OK outside this range. And OpenFF did surprisingly well in a large temp range.

Project ideas

  • PQ – re: OpenFF making its own water model - I’m not sure this is necessary. Modern water models are quite good.

    • DM – To clarify, I think we want to start from a really good water model, and then let it change slightly to become consistent with the rest of our FF.

  • DC – Re: solvation effects on globular vs. extended polymers - CC is working on this in the context of protein folding. So this first suggestion was about doing a deep dive on INTERmolecular interactions between amino acids and water. So maybe some parameter tuning along these lines in this project.

  • DC – We’re increasingly interested in using dimer interactions for FF fitting. If we can use those to fit vsite positions and charges, that’s a fast way to get started on vsites.

    • CC (chat) – Is SPICE sufficient for the first point, or do you propose generating new dimer energies?

    • DC – For INTERmololecular interactions, I prefer DES370k dimers with CCSD handling. If those don’t have protein-like-enough molecules.

  • DC – Re: partial charge assignment for vsites using ML - We do touch on this via espaloma-like methods. I don’t think anyone has looked at using models like this to assign vsite charges.

    • JW – This is something that we have on our science roadmap, but LW won’t have time to get to it for a while.

  • PQ – Moving forward, I could work with DC a bit to make proposals more concrete. Interested to know how we could work together here.

    • DC – I’ll follow up with LW on the superheating data/the aniline issue.

    • PQ – I think there’s already enough data on the superheating to make a good project, and I’ll expect to hear from you about future steps once you’ve discussed a bit.

  • JW – I’m interested to see what the statistical physics people at OpenFF think of the superheat limit - We kinda hand-picked the physical props that we fit + benchmark to, so it’d be exciting if this was a new source of data that we could use

    • PQ – Yeah, this was all from NIST. I think if you’re getting density and heat of mixing right then you should be getting superheat correct as well.

 

Action items

Decisions