LJ optimization ForceBalance consistency checks

An unanswered question in our nonbonded fitting efforts is the reproducibility of ForceBalance runs. While the process of ForceBalance fitting should be in principle reproducible, the simulations used to compare our force field output against physical properties have variability that might cause differences in the optimization process. Since we have only run single replicates in our previous optimizations and studies, we don’t have a good idea of whether this might pose significant problems to our production optimizations. This study attempts to determine whether there is significant variability in ForceBalance replicate runs, and we should be accounting for this in our production runs.

Experimental Setup

In order to understand this phenomenon, we want to explore a) whether this actually happens, and if it does, whether the b) number of physical property targets or c) the physical property types used as targets. For example, because mixture properties are calculated from a combination of several MD simulations, they have higher uncertainty and might be less reproducible.

In order to keep things consistent with the https://openforcefield.atlassian.net/wiki/spaces/FF/pages/122454022 , we can start by using the training sets/input from that specific study. This involves starting from the OpenFF-1.0.0 (Parsley) force fields, with the bonded and electrostatics left unchanged, and the LJ allowed to vary in the optimization process. (This will be identical to the optimizations run in the mixture refit study).

From those, we will select the rho_pure_h_vap (Pure density and heat of vaporization) and the rho_x_h_mix (mixture density and enthalpy of mixing) training sets to differentiate between the pure and mixture properties. We will run 5 replicates of ForceBalance optimizations (using the same settings as in the mixture feasibility study) for each of these training sets. In addition to these training sets, we will also construct expanded training sets for these types of parameters (hopefully with roughly ~2x the data points, data allowing) in order to explore the effect of the number of training data points on reproducibility.

Variability of initial conditions

We are also interested in the reproducibility of optimizations given small changes in initial parameter values. The procedure for this will be to add small amounts of Gaussian noise (probably <=5% of initial values) and run several otherwise identical optimizations. This stage will be completed after the first stage (just exploring variability from stochastic simulation outputs).

Experimental Matrix

Data Analysis

For each of the LJ parameter pairs, we will plot the epsilon and sigma values as an evolution over time. We will also calculate overall RMSD against the initial force field parameters against optimization epoch.

Benchmarking

TBD - we will do some benchmarking against similar benchmark sets as in the mixture-optimization study. This would probably be between significantly different force field optimization replicates to understand how much of a difference the changes in parameters make on reproducing physical properties.