2024-10-16 FF fitting meeting

Fitting updates

Recording: Video Conferencing, Web Conferencing, Webinars, Screen Sharing
Passcode: f1wA4&3f

TG

• Updates on fitting using projected Hessians

• JR: What do you define as a low frequency mode?

  • TG: 1500 cm-1

  • JR: How many between 100-1500 cm-1?

  • TG: Many, map to torsional and angle displacements in the molecule

  • JR: Shouldn’t care about frequencies under 100 cm-1, typically not meaningful as they start to become anharmonic

• JR: You said you want to use it for optimizations, do you mean geometry optimizations or optimizations in parameter space?

  • TG: Opt in parameter space, to minimize the difference in frequency

  • JR: For geometry optimizations, Hessian will change as optimization proceeds, and you need to update it based on what you know from points along the way

  • TG: That’s true, it’s very useful, but I’ve been having trouble fitting it correctly

  • TG: One thing that I didn’t mention is that in all these plots, there is no effects from the FF. In a FF you’d have to calculate all the parameters and then average over all the force constants; but here, each projection gets its own unique value, so any FF fit to this data will follow

• TG: Discussion topic--how should I be defining objective function for these?

  • TG: Trade-off between relative energy of frequencies and matching the direction of vibration

  •  

• PB: Can you explain the Sage QM Modes again?

  • TG: You calculate QM vib modes, save them, then calculate Sage hessian and project it on to QM modes, e.g. what are frequencies of Sage Hessian, according to QM modes

  • PB: By QM modes you mean like bond stretching, etc?

  • TG: No, calculated in Cartesians. I’m discarding MM modes and just using QM modes on both instances

• JR: Have you tried looking at this using internal coordinate Hessian? Would more naturally translate into bond stretches, etc. Maybe from that, we could look directly at dihedrals, and ignore bonds/angles?

  • TG: I’m getting internal coordinate Hessian, then using that to define MM FF. problem is that MM functional form is diagonal for Hessian, but QM Hessian has cross-terms which are not included in FF, leading to a discrepency

  • JR: So what happens if you zero out the coupling blocks in QM?

  • TG: Then you get the MM FF

  • JR: Let’s discuss offline

• BW: Might be related, in the past, I’ve used eignevalues of QM hessian as coordinates for displacement. Could compute a normal coordinate Hessian that way

  • TG: Sure, but then how would you get a FF out? WOuld have lots of weird linear combinations

  • BW: True

• LW: Do you mind expanding on the point that says optimizing the FF after the split isn’t working?

  • TG: Generated these splits but when I try to fit it, objective increases

  • TG: I’m not exactly trying to fit the whole FF, just running it through one step to see if it’s a “good” move, and the objective always goes up in step 1

• PB: Previously you tried this and it seemed to work, was the difference that you’re trying to do hessian matching now?

  • TG: maybe

  • PB: Do you think the hessians are detrimental?

  • TG: I think it’s overfitting

• JR: You mention gradients, are those also from QM results and projected onto MM, or what are you comparing?

  • TG: Taking QM gradients (cartesian) and comparing to gradients in MM calculation (internal, converting to cartesian)

  • JR: So that works for gradients but not hessians?

  • TG: Very sensitive to hyperparameters, here I’ve weighted the geometries quite highly, which improved gradients but not freqs

  • JR: Are you force matching?

  • BS: Hessian should give gradients when away from min energy structure, could be able to use a modification of this approach to do a force-matching type approach, and ignore the frequencies

  • TG: What is force matching? Just matching the QM/MM forces during parameter fitting?

  • PB: Yeah

  • TG: Ideally could do that, but there are issues with linear dependencies etc, I’ve found gradients and frequencies give quite different info

  • BS: We were looking at force matching, but weren’t based on hessian, was wondering if H could provide more info for that

  • TG: doing force matching with grads, but freqs tell you coupling between grads

  • TG: Only using minimum energy structures, so gradients basically just tell you how far it is from 0, since QM grads are all 0

  • BS: are you minimizing with MM or just using QM min structure?

  • TG: Just using QM

  • JR: Have you tried using the MM-optimized properties?

  • TG: Then run into a third issue where geometries could be different

  • JR: Well gradients would be the same

  • TG: Sure but that’s not meaningful if they’re different geometries

  •