Force Field Optimization

Main objective: Refine and improve FF optimization process.

Project driver: @Lee-Ping Wang @David Mobley @Michael Shirts

Objectives for 2020:

• Identify and improve poorly performing parameters in Parsley

• Perform feasibility study for property calculations in FF optimization and benchmarking

• Design a systematic procedure for selecting molecules for QM optimization/scanning from potentially available datasets (and possibly in some cases finding unusual chemistries outside datasets we have on hand)

Team members: @Hyesu Jang @Jeffrey Wagner @Simon Boothroyd @Jessica Maat (Deactivated)

Driver

Project / Study

@Simon Boothroyd

Non-bonded optimization Binary Mixture Data Feasibility Study

@Hyesu Jang @David Mobley @Jessica Maat (Deactivated)

QM Data Selection Parsley minor release(s)archived Training Dataset Selection

• ForceBalance:

  • Cross-validation: Multiple fitting stages using the same training set and computing parameter uncertainties based on the observed changes?

  • Stochastic gradient optimization techniques? 

• Have we selected right QM level of theory?

  • Do we have to be consistent across the use cases / types?

  • Can we use it to inform LJ?

    • Can we focus on the subset of molecules with strong steric interactions? Do those inform LJ effectively (more so than full set of molecules?)

  • Can we use ANI / Where can we use ANI?

• What things should be co-optimized, when, and how much does it help?

  • Co-optimization of torsions and LJ parameters?

• Should water model be co-optimized with FFs?

• Does optimization get harder with more parameters?

• Use of Dimensional reduction (PCA) in optimizations? 

  • An alternative to this is to test a scheme like GAFF uses to compute valence terms from a small number of fundamental parameters (See section 6 in http://ambermd.org/antechamber/gaff.html)

• Constrained parameterization for coupled FF parameters (for example, angles in benzene etc)?

• Can Hessians identify soft degrees of freedom that should be driven?

• How do we derive GBSA parameters?

• H-bonds in fitting? Two versions -- one with and one without? What is the proper way to use constrained hydrogen bonds in QM fitting?

• How do alternate RESP models impact accuracy?

• How much do virtual sites help?

• How much does polarization help? Which model (Drude, point polarizable, fluc-q) is best, or how do they compare?

• What data types to get better results?

• How do we use optimization trajectories to help?

Desired features:

• Cross-validation capability

• Scaling ForceBalance

• Internal coordinate Hessian for fitting

Software:

• ForceBalance

• OFF Evaluator