This is a WIP repo for working on an experimental force field split and re-fit using BeSMARTS and Smee.
After going through Sage 2.2.1's performance, several torsions have been identified with our existing torsions that could potentially be fixed with some combination of splitting torsions, modifying the torsional form, and re-fitting.
- t48a: this torsion is only fit to one molecule where the complementary torsion drives most of the energy. The term may benefit from splitting to cover aromatic and non-aromatic systems, as well as a change in the torsional term
- t17: Across a symmetric ring where the carbon adjacent to the ring has one additional substituent (e.g. ethyl benzene) the Sage energy is very high (mostly contributed by nonbonded and angular terms). The torsional potential should possibly have a different functional form that mitigates the high MM energy. The term is also maybe too generic, as the profiles of higher order substitution yield a better match.
- t19, which is mostly trained to terminal methyls, has an odd functional form where the periodicity=1 term dominates and gives each single hydrogen a strong central peak instead of weak n=3 peaks
- t18 performs poorly for amide-adjacent torsions
- t105 covers an O linker with an sp2 or sp3 terminus. While the sp3 profiles match the QM well, the sp2 profiles look too stiff
More detail can be found in the slides and recording here.
Using smee as the force field fitting and evaluation backend and BeSmarts to generate potential clusters or splits ("split/merge") and torsion functional forms ("modify").
Output product: a SMIRNOFF force field (MVP) that (nice-to-haves):
- ideally reproduces the torsion drive profiles highlighted in the problem cases better
- ideally with torsions having been split/modified/merged in some way
- ideally is able to easily label which torsions have changed, and (if there's a clear inheritance) their parent parameter
- ideally is fit to optimization and torsiondrive data
MVP: for a smaller toy example, focusing on one particular torsion (e.g. t17 or t48a) might be easier. It may also be easier to focus on optimizations instead of torsiondrives, as the machinery in smee/descent needs no modification.
Data is curated from both existing training data (specifically the Sage 2.2.1 set) as well as the additional molecules suggested by Cresset.
Datasets of interest:
- original-optimizations: singlepoints from unrestrained optimizations from Sage 2.2.1 fit
- original-torsiondrives: singlepoints from restrained optimizations from Sage 2.2.1 fit
- additional-optimizations: singlepoints from additional unrestrained optimizations of molecules suggested by Cresset
- additional-torsiondrives: singlepoints from additional restrained optimizations from torsiondrive scans suggested by Cresset
- combined-optimizations: combined original-optimizations and additional-optimizations
- combined-torsiondrives: combined original-torsiondrives and additional-torsiondrives
Ideally this experiment would use the combined- datasets, but it may be cleaner and a smaller toy example to start from the additional- datasets.
We expect to use the functionality and data types in descent and smee respectively to do the force field fit.
Cresset included some ideas as to what may be the correct answer based on their analysis.
- t48a: remove the periodicity=3 term and possibly split between aromatic and non-aromatic systems.
- potentially: split out t47 with an O-specific type for HCCO, as currently t47 may be too strong
- t17: either split out the term or have an additional 2-fold term
- t19: a good end state would result in a stronger n=3 term and maybe even removal of the n=1
- t105 may need splitting