Hessian/frequency fitting  Recording:Video Conferencing, Web Conferencing, Webinars, Screen Sharing
Passcode: 7RF7+?ua | TG | PB – How is this different from Hyesu’s? internal hessian work? TG – Unsure, it should be similar but I haven’t looked into it. LPW – My understanding of the previous slide is that you compute the MM hessian at the QM structure, and then you’re expressing that MM ehssian in the basis of the QM normal modes. Because the MM hessian isn’t diagonal in the absis of QM normal modes, you’re basically fitting the diagonal components of the MM hessian in the context of the QM normal modes. LPW – I remember hyesu did something a while ago that implemented a fitting target with IC hessians, but I don’t think we explored that extensively before she left. So I’m not worried if this ends up revisiting ideas from before since there’s a lot of work to do here, like how to rescale component contributions from different internal coordinate types. TG – I haven’t tested this out, but it’d be interesting to see if the MM matrix becomes more diagonal as you fit frequencies.
 LPW – Slide 14 is the formula for the second derivative of an internal coord wrt cartesian coords? (TG: yes) If the codes in geometric have a mistake they’d need to be corrected, but I’ve checked the correctness of that code using finite difference methods, but if you’ve found cases where it’s wrong I’d love to know. PB (chat) – link to paper, https://pubs.aip.org/aip/jcp/article/117/20/9160/464558/The-efficient-optimization-of-molecular-geometries TG – I used your implementation and noticed that some of the zeta functions are replaced. LPW – I followed the paper and found the difference between finite difference values, so I changed things to get them to agree. TG – Right, it seems like psi4’s implementation dropped some terms but it gets the same results. This shouldn’t happen, since psi4 coded one set of indices (one permutation) but they didn’t code for the other indices (swapping the sign). I don’t know how they’re getting the same result. LPW – I think the computational cost for those is small, so if the values are OK then I figure you’re fine using either one. TG – One thing with the geometric implementation is that, to calculate the V angle, there was initially a problem where the angle was measured backwards, so it came out as negative. But it cancels out(?) since it’s done consistently. LPW – Yeah, I guess as long as it’s consistent it’s fine.
BS – Does it look like, with this method, if you iterated, would it converge? PB – Regarding force constant values - Can we scale down force constant values (re: discussion of stiff force constants interacting with barostats) - does it make sense to scale down force constants from modified seminario by like 2/3rds to start a fit? TG – No, the frequencies would then be quite poor. PB – But Parsley and Sage had issues with triple bonds getting k=1600 without MSM, and >3000 with MSM. TG – I think that was an interaction between soft and stiff bonds being adjacent to each other. But currently, if you decrease bond force constant, you’ll decrease frequency, which will disagree with QM
PB – Do we really need to try periodicity=9 for ethane? TG – Probably not. One thing that getting nonzero values for periodicity=9 means that there’s some disagreement with other terms and QM that’s really hard to reconcile. PB – Would introducing higher periodicities introduce noise? TG – Yes, probably. It’s kinda an effect of overfitting. If fitting more targets there might be more of a (beneficial) effect.
BS – These high frequencies are probably due to CH stretching. Since most people using FFs will freeze bond llengths to H, these are probably not so important. LPW – In deriving initial parameters, whe you generate the IC system using BAT, it’s redundant since you have more primitive intenral coordinates thatn degrees of freedom of the molecule….. LPW – I don’t understand how, if you’re reading off force constants from hessian that’s been transformed into IC, and you think about taking on of the bond angles and deforming eg a HCH angle from 109 to 110, you’re actually changing multiple angles at the same time… I’m unsure that the force constant that you’re reading out of the IC hessian is due to JUST the changes to that angle, given that it’s impossible to change only that angle. If you used a Z matrix representation,the ordering of the atoms would matter, it would encode a path through the molecule, and the z matrix coordinates would make a nonredundant path through the molecule. But that’d also mean that some angles aren’t in the IC system by themselves… TG – I’ve been using the set that the force field defines, which are redundant. But there, even for the CH bonds, you can get wildly different hessian values. So on one side of the mol you can get a large force constant, and on the other you get a low one, and if you built a nonredundant set you might lose that. I think there is some work to be done to tease out that question more. I think my B matrix did some over/undercounting, I was maybe hoping that the pseudoinverse would compensate for that. LPW – It seems like you want the dependence on the IC system to not be large, or to have a physically justified choice of IC system.
TH – If this is ethane, I assume that, given the symmetry of the molecule there area bunch of modes that must be identical. Eg F17+18 on slide 20 are probably symmetric bonds. TG – Yes, they’re probably symmetric and antisymmetric stretching. TH – So, it seems like there are a bunch of freqs that should be identical. Can you use those symmetry arguments to figure out what the uncertainty in the QM freqs is, and then use that uncertainty to know how to set a target for the MM fits. TG – I don’t know enough about QM to speak with confidence on this BS – I don’t think those close-but-different frequencies are supposed to be identical. I think that symmetric and antisymmetric freqs SHOULD be slightly different. Â
|
