[Sage Discussion]: Data to use for the next round of force field fitting

[ocmadin]

Discussion

This issue is to facilitate discussion of physical property data for use in the next round of OpenFF fitting. The idea here is to have overarching discussion in this thread, and have specific discussions that get more into specifics in separate issues, which will then be linked back to this master thread in order to maintain continuity

Decisions

When final decisions are made on a particular dataset, they will be recorded in PRs (in text form, as well as nistdataselection scripts that select that particular datasets, if applicable)

Starting point for discussion: here are the notes from the discussion we had at the OpenFF meeting in August:

https://docs.google.com/document/d/1z7hZDklgZHPxiR1JfpktGsM8sa0ZS_CW_1bGqfFU0SI/edit#heading=h.993v2np6jmi8

[davidlmobley]

@ocmadin maybe it'd help if you pull some specific points we need to decide and stick them here as questions? The document you linked to has 20 pages of notes.

How does this interface with the DIPPR discussion currently going on in Slack and with Simon's current work?

[ocmadin]

@davidlmobley sounds good. I'll add some of the main questions that come to mind here (tagging @SimonBoothroyd for additional input).

Here are some broad questions/areas of discussion (with github issues linked):

• 13 Pure Compound Properties: Currently we are using density, hvap, static dielectric constant for fitting. For the sage release, are we going to stick with these properties or do we want to add/subtract properties (e.g. removing hvap, adding surface tension)?

• 14 Pure Compound Data Selection: Our sources for pure data include ThermoML, the DIPPR free set, and (hopefully) the DIPPR subset that we are currently talking to them about. Broadly, which data points should we use from these sources, and how should we divide them into test/train/validation sets?

• 15 Mixture Properties: Which mixture properties do we intend to use in fitting for Sage? For Parsley, we used enthalpy of mixing and excess molar volume for benchmarking; should we use these for training for Sage?

• 16 Mixture Data Selection: We currently have ThermoML as our source of mixture data. What mixtures should be emphasized in the training data? What solvents other than water that we should be focusing on?

• 17 Pure/Mixture data split: In fitting, how should we prioritize pure vs. mixture data? In each set, what do we emphasize to cover the deficiencies of the other sets?

  • 18 Host-guest/protein ligand binding data: Do we use host-guest/protein-ligand data in our fitting? Or just for benchmarking? If so, which systems should we look at?

• 19 Crystal data: Mentioned in the August discussion, but probably out of scope for this release. However, we might want to get the ball rolling on this for further down the road (if its even something we want to do).

• 20 Water fitting: What water model do we use in this round of fitting? Do we attempt to do any water model parameterization?