Property selection: L.-P. Wang suggested using surface tension for parameter fitting, which isn’t a traditional property used in force field parameterization. A discussion ensued about the merits of using surface tension as one of the target properties, and whether precedents in literature exists. L.-P. Wang described some previous work looking at the simulations of a slab of liquid showing moderate agreement with experiment, and highlighted that, according to the study, the results are relatively consistent as long as the film has no holes. This was to answer M. Shirts’ question whether the calculated surface tension depends on the thickness of the slab. L.-P. Wang thinks that surface tension could be a decent stand-in property where enthalpy of vaporization data is lacking. O. Madin and R. Messerly were looking at the relationship between surface tension and vapor pressure, finding that is difficult to fit parameters for both properties. This suggests that these two properties contain distinct information for the force field parameters. They were using surrogate models for a four parameter model (a two-site Lennard-Jones adjustable bond-length with point quadrupole) and only investigated ethane, but the study they were using as a template did investigate several small molecules. In each case, including surface tension in the optimization leads to significantly worse prediction of vapor pressure and/or liquid density. This suggests that including this property could help develop a more sophisticated force field by providing unique molecular insight. M. Shirts thinks it might be worthwhile to include adhesion properties, as they are somewhat orthogonal to the other properties that will probably be used in parameterization and could help us determine the value of, eg. polarizability terms, but later found a paper suggesting that there is a moderate correlation between two properties. A fundamental issue for M. Shirts is how do we select one representative of “cohesion” properties that covers that whole family of intermolecular interactions, but answering this type of question is beyond the scope of first generation OFF force fields. J. Chodera suggested identifying a basic set of requirements/tasks for a surface tension property calculation, assessing the amount of work associated with implementation and assigning priority to this task with respect to the existing fitting plans. L.-P. Wang doesn’t see this as a part of the first generation force field fitting and is not planning to commit a person to work on this further for now.

Another important question was raised by J. Chodera, asking about the number of measurements present in ThermoML, which will be the main source of experimental data for force field parameter fitting. R. Messerly stated that ThermoML is likely to contain a high number of surface density data points, but their quality might vary more compared to some other properties of interest and the compounds in the database should be screened to match with our chemical space of interest. Some preliminary surface tension data and accompanying scripts should exist in open-forcefield-data repository.

Equilibration: O. Madin raised a previously reported Issue 283 on equilibration detection of corner cases, in particular where a timeseries gets tricked by a few bad initial inputs. J. Chodera responded that he has posted a potential fix on that issue already. M. Gilson believes that there’s nothing wrong with removing the initial part of a timeseries, as any solution will be heuristic to a certain degree, and L.-P. Wang agrees that dropping the first 5-10% of the timeseries might be reasonable. O. Madin further highlights a problem with properties that equilibrate slowly, for example, there’s a long equilibration time on glycerol energies. How to best approach equilibration problem remains to be determined and this task has been assigned to O. Madin and S. Boothroyd, as a part of openforcefield/timeseries repo.

Reweighting: S. Boothroyd asked who is in charge of implementing reweighting, which should be resolved around the same time as the property estimator launches. M. Shirts should look into property estimator documentation to understand which variables will be accessible directly and which through queries in a calculation layers API, followed by another call to discuss this in more details between S. Boothroyd, O. Madin, J. Chodera and M. Shirts.

API: Currently the highest-priority task is the public API and the pluggable API, where people can add/experiment with new property calculations. At the moment, density and dielectric constants are being implemented.