Co-optimisation of VdW and Electrostatics

Problem Statement

We know that the electrostatic and VdW force field parameters are strongly coupled - only optimising the VdW parameters will likely only improve things so much, while also running the risk of encoded electrostatic phenomena into the VdW parameters. To produce the largest improvement / impact, the two should likely be co-optimised.

There are several approaches in which to refit the electrostatics -

• optimise templates of charges directly

• optimise bond-charge corrections

• optimize atomic hardnesses and electronegativities in a vcharge-type model

Possible Approaches

Charge Templating

This refers to an approach where charges are assigned based on functional groups, which could be defined with SMIRKS strings. For example, every ether group looking like C-O-C might get charges 0.1/-0.2/0.1, and so one. Thus, one would look for groupings of atoms that fit a template and then assign those templated charges. This is how Momany-Rone charges worked in the commercial version of CHARMm in Quanta. On reflection, however, I think this may add complexity without much benefit.

BCC Refit

BCC Refit + dimensional reduction

From MKG: “Here is a suggestion for a relatively straightforward study. Feedback invited!”

Maybe the lowest-hanging fruit would be to simultaneously optimize LJ parameters with the H2CON and H2CO3N typing models, along simultaneously with a similarly compact set of BCCs used in conjunction with regular AM1 charges or else a better but still fast QM alternative. The simplest BCC sets would have a single BCC type for each LJ-type pair. Thus, for H2CON, with five LJ types, we’d have 4+3+2+1= 10 BCCs (one for each pair); and for H2CO3N, with seven LJ types, we’d have 6+5+4+3+2+1 BCCs. If these BCC typings seemed insufficiently fine-grained, we could do something like what Michael Schauperl did in his LJ paper, allowing BCCs for various simple typing models (which could be decoupled from the LJ typing.) For example, we might have the H2CON LJ types, but H2CO3N2 BCCs. The whole thing could start simple and probe out to more complex schemes. By keeping GAFF v?.? and OpenFF v?.? as baselines, we could quickly get a sense for whether a given model is deficient vs promising.

Training Set Composition

Any optimisations will be done upon small training / test sets of C, O, H and limited functionalities to keep the scope of the study tight and attainable within a reasonable time frame. Limiting the number of functionalities should reduce the difficulty of the optimisations due to the reduction in the number of parameters (DoF) which need to be optimised.

Lorem ipsum.

Parameters to Retrain

• One possibility to simplify the study would be to start with the simpler LJ typing schemes investigated by @Michael Schauperl (Deactivated) and @Michael Gilson, in order to reduce the number of parameters required to be optimized.

  • This would likely involve the H2CON (2 hydrogen types, single types for C, O, N) or the H2CO3N (2 hydrogen types, 1 carbon type, 3 oxygen types, 1 nitrogen type).

  • As mentioned before, in order to keep the scope tight would not optimise the nitrogen types.

Required Infrastructure / Software

References

• https://onlinelibrary.wiley.com/doi/full/10.1002/jcc.10128 (Bayly BCC parameterization paper)

• Michael K. Gilson, Hillary S. R. Gilson, and Michael J. Potter, Fast Assignment of Accurate Partial Atomic Charges: An Electronegativity Equalization Method that Accounts for Alternate Resonance Forms J. Chem. Inf. Comput. Sci. 2003, 43, 1982-1997