Arvind Ramachandran’s BespokeFit Query | AR | AR: Seems like Bespokefit idea is to start at sage parameters then expand periodicities 1-4. JH: Periodicities alternate between phases 0 and 180 - hard coded. AR – When you do that, does it affect symmetry? JH – Found that you don’t need all 4 periodicities usually AR – But if you phase shift the even periodicities by 180, and not the odd … it’d be worth phase-shifting some by 90/270 to capture profiles that aren’t symmetric around 180. JH: Comes from OPLS function form where k alternates between -ve and +ve CC: Allowing some k’s to be negative means that it’s possible to get asymmetry DC: Paper been discussing recently which discusses simplest possible fourier series to give sufficient flexibility - will share
AR – Also saw that regularization is implemented to initial values. I think that’s a good idea. When I started I tried to use forcebalance and fit to residuals, and I saw big cancellations appearing unnecessarily. What’s your thinking on this? JH: Regularization really helped show which parameters were most important - effectively removes unnecessary parameters. Either to 0 of back to Sage (which are already good) AR – Re “sage parameters being good starting points” - People will often be applying bespokefit BECAUSE Sage isn’t sufficient for a molecule. What are your thoughts? JH – In testing, we found that Sage was generally the a good starting point. Additional gain may be from additional periodicities. DC – We also use stronger regularization than ForceBalance’s default.
AR: Term called prior width -what is it? AR: How is torsional coupling handled? You take all torsions belonging to a molecule and fit all together. One method to address coupling is to look at two D potential surface and fit to 2D potential surface. Bespokefit seems to use 1D scans individually - neglects coupling which would be properly treated in 2D.
|
BespokeFit-SMEE Initial Results
| Finlay Clark
| PB – Are you fragmenting molecule? PB (chat) – I think spiro compounds are also another class of interesting mols DC – Is restrained min the normal procedure? FC – Asked JH and he said to do unrestrained. So torsion atoms are CONstrained, and other atoms are REstrained. DC – We should do this WITH constraints but WITHOUT restraints. JH – In FB this is used during torsion fitting, may have been copied in unintentionally.
JH – What was mol 10 (with the really high error)? (slide 25, see recording 30ish min) DC – What we could do is report FB target as sort of weighted error (will focus on minima and put less emphasis on very high energy areas) JH – Bespokefit struggles when there’s an N in the bridge since it doesn’t fit improper. DC – Also, should openff profile in plot 10 be able to go negative? (general) – I think the lowest energy QM struct is set to 0 in all profiles.
CC – Re: assessing convergence of ensemble - what’s your loss funciton? FC – Take MM gen ensemble, calc energies and forces with MACE, and empirically scale up forces weight to be comparable with energies, scale down super high energies/forces… CC – … FC – The fact that dotted and solid lines on slide 10 converge… (see recording around 24 mins) CC – Would suggest talking to Joe Greener (who had some done work on getting impl solv to match expl solv). May have suggestions on this.
AR – Re: the subtlety on restraints - more description? FC – When you benchmark against torsiondrives, you constrain the torsion atoms, but relax everything else AR – Was this done in bespokefit? DC – For training/fitting, it’s good to use restraints since the FF isn’t optimized and is more likely to wander away from QM min. For testing though it’s better to NOT use restraints since that will reveal errors more clearly. … JH – There’s (something, recording 38 mins) in a branch of forcebalance
DC – Maybe pick a bigger subset of TorsionNet to find a bigger subset of mols to test hyperparameters. JH – Are you interested to see how accurate you can get it and include some coupling terms? FC – Could be good to put in some sort of automated convergence detection to stop fitting once loss levels out. DC – JW, would you be up to do a code check to see how to bring this into openff namespace? FC – Re: FE benchmarking, previous benchmarks didn’t show significant improvement. JH – Recent schrodinger paper showing that many systems didn’t show improvement, but they highlighted systems where they do. DC – And maybe ones where ex. impropers for pyramidal Ns make a difference. JH – Linear torsion stuff? FC – Not really, everything goes to 0.
|