Research Projects

Re-assess valence term fitting targets

• Currently employ ‘opt-geo’ and ‘torsion drive’ targets while fitting valence terms

• Should we fit torsions separately (before / after) fitting other valence terms?

• Can we fit to energies +? gradients +? hessians of ‘diverse’ (esp. not energy minimized) conformers instead to yield the same / better performance?

  • Will simplify creating a next iteration of fitting engine that will be used after FB

  • e.g.

Expand / re-design QC training sets and expand test set

@Pavan Behara

• Should we move away from focussing on generating QC data using minimum structures, and instead try to generate data at diverse conformers that captures the PE surface away from the minima which is more often what is sampled?

• Can we replace the current QC method with a surrogate like ANI or XTB?

• Generate more diverse TD data

  • Should we be focusing on torsions without significant steric hindrance to most inform the torsion parameters?

• Current QC test set mostly contains minimum energy structures. We should likely expand to consider other quantities such as QC ESP / EF + dipole moments to inform charge model fitting

Granular benchmark metrics

• Current QC benchmarks focus on global metrics such as ddE and TFD, but this obscures finer and critical information such as capturing the correct local geometries, e.g. regression in geometry of sulfonamide groups was not flagged by current benchmarks

• Ideally would expand openff-benchmark to

  • be general purpose for use by the fitting team

  • incorporate local performance metrics such as ability to reproduce geometry of chemistries of interest (e.g. sulfonamide)

• @Lorenzo D'Amore has somewhat started on this.

Boron parameters

https://openforcefield.atlassian.net/wiki/spaces/FF/pages/2605842482

GNN Charge Models

@Simon Boothroyd

• https://openforcefield.atlassian.net/wiki/spaces/FF/pages/2311847937

Virtual Sites

@Simon Boothroyd

• https://openforcefield.atlassian.net/wiki/spaces/FF/pages/2311061505

Improper Refits

@Pavan Behara (formerly)

@David Mobley

• Improper torsions mostly imported from GAFF? force field

• Likely low hanging fruit to refit improper to (IC?) hessian data

• Likely a good candidate for WBO interpolation as the degree of conjugation (e.g. around a pyramidal nitrogen) will greatly effect the degree of planarity

Automated Chemical Perception

@Trevor Gokey @David Mobley

Retrain alkane torsion parameters

@Trevor Gokey @Pavan Behara @David Mobley

• Trevor and Pavan have identified a number of torsion parameters in Sage that would benefit from being split into multiple parameters.

• The performance of these split parameters seems to suggest a new minor release of Sage be made with these inclusions

Alternate vdW Potentials

@Daniel Cole @Joshua Horton

• Exploring whether there are ‘better’ alternatives to LJ potential

• Would ideally:

  • have more degrees of freedom so potential can more accurately capture both short and long range interactions. Currently we focus on improving the accuracy of long range interactions and push the error in short range interactions into torsion parameters.

  • have soft-core properties built-in so as to simplify free energy workflows.

Bespoke torsions trained to XTB / ANI rather than QC methods

@Daniel Cole @Joshua Horton

• The currently level of QC theory used in OpenFF refits is slooooow to compute

• Ideally we’d perform bespoke torsion drives using a QC surrogate like ANI or XTB.

• Initial results show significant improvement to JACS binding free energy even when training bespoke torsions to XTB.

Full GNN force field

Yuanqing Wang @John Chodera

• Espaloma is showing very promising results as a pathway for training a classical molecular force field entirely as a graph neural network. See #espaloma