Automated outlier detection | MT | Prepped notes: YAMMBS currently outputs a highly nest data model that stores, for each force field and for each QM molecule in the reference dataset, the “ICRMSD” a.k.a. bond, angle, proper and improper torsion geometries from MM-minimized structures. Could filter through for a given criteria and return molecule (mapped smiles) and MM & QM structures i.e. these QM molecules optimized from force field foo.xml have these bond lengths off by more than 0.05 A: (list of molecules, QM & MM geometries, and atom pairs defining problematic bonds these QM molecules optimized from force field foo.xml have these torsion angles off by more than 30 deg: (list of molecules, QM & MM geometries, and atom quartets defining problematic torsions
These analysis are automated on OpenFF infrastructure (using a fairly large AWS runner) by anybody with OpenFF credentials, but can could be run on any hardware with marginal startup cost.(All of the above already exists)Could build on top of this a small semi-automated tool that, given ICRMSD criteria a short report/JSON dump about which molecules are “outliers” by that criteria visualization(s) of some molecules, highlighting problematic valence failures (i.e. highlight four atoms in a torsion)
Any other details I’m missing?To run this on fitting data (Mark Mackey’s suggestion), I would need help consolidating training data into the right QCSubmit and/or YAMMBS data model(s). Anybody who has recently done a fit with “all” of this data would be a good person to coordinate with. (I think internally that’s only Lily right now?) Don’t think I need any other to get started.What priority/urgency should this be? (end of prepped notes) MT – Does this cover what folks would want/is the feature scope enough? LW – Would this be built into YDS, or would it be a separate functionality? BS – I see this as focused on internal coordinates. Could it also be generalized to include energies/barrier heights between conformers? MT – Yes, I’m focusing on internal coordinates since that was MMackey’s original suggestion. But I think the same process could be applied to find ddE outliers and such. BS – Would be a slightly different kind of properties since it involves pairs of conformers instead of individual sturctures. MT – Good point, but generally still looks at pairings between QM and MM structures so this should be tractable.
MT – Great, so moving forward I could use some scientist-time to identify good dataset(s) for initial testing. LW – JClark or I could help here. Though your existing datasets should work well for this as well. MT – Oh, right, I can just use industry set.
TG – ForceBalance’s objective function is essentially the internal coordinate RMSD. Is there a need that’s not being met by that? Also in FB you can scale internal coordinates by a “denominator”, which is nice. MT – My incomplete view of things is that our goal here is to save human time in running standard FF benchmarks between many people. So this isn’t new functionality that we’re looking to make. And denominator could be user-inputtable. LW – Big benefit of this is that it’ll be easier to interpret - it’d be a bunch of work to do the outlier detection style stuff with ForceBalance
|
vdW updates | TG | JW (chat) – This is so interesting because Andrea Bortolato was just reminding us of the talk Sandbox gave at the 2024 OSMF symposium about how they needed to change vdW parameters to reproduce stacking (would involve that #6X3 vdW param) DM: why are some of these parameter differences very tiny, e.g. n20 vs N48? TG: BESMARTS takes the best candidate out of a large pool. Sometimes the best improvement is very small. DM: could some of this be noise? TG: I picked a very low sensitivity threshold
LW: … a technical question – would BeSMARTS automatically remove these parameters itself given more fitting time? BS – The nitrile is weird since it has a lone pair sticking out. I wonder if there’s a sigma hole and a different charge model is needed. TG: Yeah, maybe (outside the scope of this work, which is to test BeSMARTS). If you look at the energy scan, the potentials don’t look very vdW-like, so something else is needed.
LW – Regarding the difference between the splits, the first includes all the same datapoints as the second, and more. So why is the methyl split only appearing in one? TG – The difference between the datasets is that I included fewer structures at larger dimer distances. With more data it gets harder to find splits, and in this case the additional structures weren't adding information. JC – So you’re saying that, if a certain moiety isn’t broadly expressed in the dataset, the algorithm will have trouble splitting to improve that since it will have such a relatively small effect on the overall objective? TG – Yes, basically.
CC – In plots of vdW energy vs. SAPT terms - Are you trying to fit to SAPT terms during fit, or just benchmarking? I don’t expect that they’ll map well to each other. I’d suggest just fitting total energies and not trying to use decomposition. TG – This is something we’d talked about last meeting. One thing I have to be careful with is that the charge model will add noise, so what I don’t want to do is end up compensating for charge model terms using vdW… I’m hesitant because for a lot of molecules, you have a large amount of error … CC – I don’t think that’s a good comparison either. I wouldn’t decompose further than “total SAPT energy = total nonbonded energy”. The SAPT devs themselves tell people not to try and fit to SAPT decomposotions. DM – Somewhat agree, I’d like to check whether these splits really improve things with outside benchmarks. BS+CC – SAPT energy decomp includes terms that don’t neatly fit into FF. TG – I don’t disagree, but the alternative is going to end up with a vdW model fixing errors from charge model. CC – But that will happen anyway in condensed phase sims.
|
Torsion fitting | Arvind Ramachandran | AR – Seems like there are torsion parameters for ex. butane - a bespoke term for butane would be HCCH, HCCC, and CCCC, and you’d fit them all together. But it seems like OpenFF parameters are those individual torsions. AR – For a given periodicity, you sometimes have multiple phases. Where do they come from? Bespokefit will expand to multiple periodicities. LW – We don’t normally have torsions that complex. Torsion periodicities and phases are largely determined by history. JW – Unsure about bespokefit science DM – I think there’s some awareness among bespokefit folks that you can easily overfit by adding lots of periodicities. And agree with LW that a lot of the periodicity/phases in our current FFs are historical…
AR – We’re seeing that, when you have multiple torsions and you’re fitting to their sum, and the total energy profile isn’t that high, you can sometimes get parameters that are big in opposite directions, and regularization helps with this.
|
