Forcefield parameterization plans

Christopher: How would work be distributed? Need to be very focused about time put into it

Methodology improvements to ForceBalance:

more efficient: reweighting; can we speed up from 6 weeks on 200 GPUs to 6 hours?
Bayesian sampling

Questions to address:

Desiderata:

Want a consistent set of nonbonded parameters, not special parameters for solvents.

Useful data:

a paper on validation of experimental data, "Algorithmic Framework for Quality Assessment of Phase Equilibrium Data" (http://pubs.acs.org/doi/abs/10.1021/je500327k)
excess chemical potential?
partial molar volume and heat capacity as a function of mole fraction (MRS)
Argument for partial molar volume and partial molar enthalpy as key parameters

Christopher is interested in small molecule forcefield with the following properties:

start with neat and mixed solvents
simultaneous fitting of all solvents
bond charge increment model for fast corrected semiempirical-derived charges
sigma hole on C-Cl / bond-centered charges in general, but judiciously chosen (MRS has reweighting scheme for bond length changes that may prove useful)
scaling factors for 1,4 torsions would be an adjustable parameter
bond lengths and valence angles aren't going to make a huge difference, but we may want to choose them to be close to some QM reference to allow for good QM/MM overlap
would like to allow decoupling of parameter types for nonbonded, bonds, angles, torsions (VSP: could be implemented in traditional codes with combinatorial atom types)

Data to fit:

QM calculations: torsion drives
excess densities as a function of mole fraction and temperature (to be collected by JDC)
radial distribution functions from scattering experiments (or scattering data directly); how do we measure these?
excess molar enthalpies: calorimetry can do this; heats and heat capacities of mixtures of liquids (see work by Paul Smith, Kansas state)
need database of simple compounds, but compounds we need in database will change with atom types
NMR for torsions? (KAB working on this)
organic crystals from CCDC?

Atom typing:

Comparisons:

As a baseline, we should be able to do better than established chemical engineering statistical mechanical models such as UNIFAC, which are group contribution methods that are extensively used in industry to design chemical plants, etc.
[Some information on chemical engineering models] (https://github.com/choderalab/open-forcefield-group/wiki/chemical-engineering-models)

Identify what data we need to improve forcefield

JDC will work with Christopher Bayly to get atom type sampling running

Starting set of molecules?

The path forward:

Christopher Bayly will come up with a few hundred molecules as SMILES strings for some initial QM calculations
Christopher will come up with base types and decorators, send along parm@frosst examples
We will work with Lee-Ping Wang to do some QM calculations; lots of torsion drives
- QM model for torsion scans must be doable by other people: DFT/6-311+G(d,p) (dispersion-corrected)
- GAMESS would be nice (or Qchem, or quantum espresso, etc.)
- do different quantum packages come up with same potential?
- 30-degree increments is good, but have to get to all minima
- go around one way and then reverse to avoid hysteresis
- barrier heights don't have to be that detailed, but have to get all minima
John Chodera will send Christopher a list of simple liquids we can buy to select optimal subset

Look into:

Provide feedback