The technique originated in geostatistics, where it is desirable to Uncertainty) from a limited number of observations. Surrogate models, we can use Gaussian Processes, which are simple surrogates that excel at making predictions (with Regions of interest, and can be approximated with simpler surrogate models with limited losses of fidelity. We can accelerate sampling with surrogate modelingįortunately, the physical properties we are interested in are relatively smooth functions of the LJ parameters in our For us, this means we need a muchįaster way of taking LJ parameter inputs (sigma/rmin_half, epsilon) and getting physical property outputs (densities,Įnthalpies) very quickly. Parameter probability distribution, we need a way to speed things up significantly. So if we want to do Bayesian exploration of a Is even more computationally expensive then most optimization schemes. The catch, however, is that evaluating a Bayesian posterior distribution generally requires Monte Carlo simulation, and Knowledge, and b) because understanding the whole parameter distribution can find parameter sets that escape local Bayesian posterior sampling is a promising method of exploring parameter spaceīayesian inference is a paradigm for evaluating parameter sets that naturally incorporates prior information and a likelihood derived from experimental data into a posterior distribution, which can be used as a metric for fitness.īayesian inference is of interest to us because of a) the way that it parsimoniously incorporates previous model This makes it difficult to efficiently optimize LJ parameters or otherwise explore parameter space due The disadvantage of using physical property measurements in parameterization is that they are generally slow toĮvaluate, because computing a physical property with a force field requires you to run an equilibrium simulation with This is an important step of parameterization because it fits theĬritical LJ parameters and introduces macroscopic constraints that complement microscopic QM constraints. Typically liquid densities and heats of vaporization). Since the earliest days of force field science, LJ parameters have been fitted against physical property datasets ( It down! Lennard-Jones parameterization relies on physical properties This is a lot to parse in one sentence, so let’s break Working on the parameterization of Lennard-Jones models, one strategy I am exploring is using Bayesian inference toĮxplore parameter landscapes with respect to physical properties. In the Open Force Field Initiative, we are always looking for novel ways to advance force field parameterization.