2024-07-31 FF fitting meeting

Participants

  • @Trevor Gokey

  • @Alexandra McIsaac

  • @Pavan Behara

  • @Lily Wang

  • @Jeffrey Wagner

  • Chris Bayly

Goals

  •  

Discussion topics

Item

Presenter

Notes

Item

Presenter

Notes

Follow up on Hessian work

@Trevor Gokey

  • Recording here . Passcode: b^$%7W7t

  • JW: What range is H stretching? (re frequency data)

    • TG: around 3000-3500 cm-1

    • JW: Most people use constrained FF, so H stretch isn’t too important. Issue is with lower frequencies

    • TG: Bonds in 3000 range aren’t only C-H bonds, it’s all bonds. Those errors aren’t unimportant

  • CB: Resonance frequency issue with C#C-CH3 triple bond that caused explosions. Other triple bonds are C#N, maybe one other….should be around 2500. Generally worried about resonance frequencies, plotting like this would help identify

    • TG: What do you mean by resonance

    • CB: Reason that H bond are constrained is because MD timestep is close to frequency of CH bond, leads to explosion in energies during MD. Depending on timestep, could happen with other bonds

    • JW: Wouldn’t see same issue with triple bond, because in those simulations we used HMR which lowers mass of C. So wouldn’t show up here, with genuine masses.

      • CB Could recalculate these frequencies using scaled down HMR mass

      • JW--could be good topic for FF release meeting, could be an addition into release process

    • CB what other bonds besides H have frequency over 3000?

      • TG rings with two N’s triple bonded or other N#N, force constant 3300.

      • CB also cyclopropyls, still looking into the cyclopropyl resonance

    • TG good point that this could be useful for identifying resonance frequencies

  • PB--Last time you also showed difference from MSM, did you check that?

    • TG--no, too many molecules, didn’t check. But Sage is very close to MSM. Sage underestimates bonds but does angles well.

  • TG--using vibrational modes of QM, not minimizing MM structure beforehand. There is a white blank space line in frequency plots, anything closer to y axis is a negative frequency (MM isn’t at minimum)

    • CB--is that a meaningful thing to calculate? Second derivative wouldn’t necessarily be comparable at minimum vs not minimum?

    • TG--If you minimized, just because you got same FC wouldn’t mean same vibrational mode, could be same eigenvalue but not same eigenvector

    • CB--but now you have same eigenvector but not necessarily meaningful eigenvalue

    • TG--have to make tradeoffs and decide what is better

  • LM--you suggested incorporating this into fitting pipeline, what aspect would you incorporate?

    • TG--idea would be to replace MSM with something like this

  • CB--Hessians come from QM min, right?

    • TG--yes

    • CB-- That would inform the fit, especially the torsions. Would inform FC of bonds and angles (single minimum per conformer). However, for torsions, a much better description of minimum would change description of barrier in addition to FC. Would it make it better or worse?

    • TG--good point, I was first looking at description of minimum, before looking at other points on PES like barrier.

    • CB--Fitting the Hessian at any point would also affect all targets

  • CB--are you using torsion drives?

    • TG--no, want to see how well we can do looking only at minima before using exotic properties/conformers

  • TG--Chris, what’s your opinion on Sage frequencies being overestimated for torsions? Got same result for Frosst

    • CB--Definitely not a planned feature. I didn’t know it, but not necessarily surprised

    • CB--Just speculating here. All our FFs do torsion parameterization as though torsions are independent, but know that there is a strong correlation between adjacent torsions. Could be outcome of not considering correlation/dependence between terms (“diagonal” FF). This is where that would show up

    • CB--Makes me nervous that you aren’t using MM minimum for frequencies, QM hessian includes anharmonic features which we can’t capture with a “diagonal” FF

    • TG--In general, there are a few negative frequencies, usually except for pathological cases they are low eg -100/-200 cm-1

    • CB--Could be good to check using MM minimum for small subset of dataset

    • TG--problem would be matching the vibrational mode for comparison

  • LW--Diagonal lines for bonds, looks like Sage is predicting two averaged values so half are below and half are above. Wonder what the reason is.

  • JW--is your FF coming up with a bunch of H bond parameters? Or just 3?

    • TG--Using Sage SMARTS, not changing the patterns

  • CB--See recording around 47 min, shares some details about which bonds are in which frequency ranges

    • CB--not seeing alkane CH on this plot, but might be constrained

  • TG--can pass on re-set Sage parameters for fitting team to re-fit, and am also trying to re-fit myself

  • CB--is MSM using hessian to get FC?

    • TG--Yes

    • CB--So, a similar method. Do you think including this info would make result closer to MSM?

    • TG--No, they land in different local minima. MSM is very local, e.g. for a bond it only sees the two atoms that are bonded. This uses pseudo-inverse of B, which means it’s aware of all the internal coordinates







Action items

Decisions