(681k) Machine Generated Coarse-Grained Force-Fields for Efficiently and Effectively Addressing Transferability

A key objective of the Materials Genome Initiative (MGI) is to create a coherent framework to perform coarse-grained (CG) simulations in a streamlined and efficient manner to reduce the time and cost of novel material design and discovery. In order to gain more widespread application in that endeavor, it is imperative to make progress on one of the main limitations of coarse-grained potentials and efficiently and effectively address the issue of transferability. In this study, we develop a coarse-grained potential for polystyrene at a resolution of one CG site per monomer using Iterative Boltzmann Inversion (IBI) in conjunction with Uncertainty Quantification (UQ) techniques. We also introduce a new Spectral Monte Carlo (SMC) method to calculate the necessary distribution functions for the IBI framework which is orders of magnitude more efficient than typical histogram methods. Using a minimal set of IBI-parameterized CG force-fields derived across a range of temperatures and pressures (which constitutes an initial guess), Bayesian methods are subsequently applied to efficiently derive machine generated CG force-fields which are accurate at arbitrary state points. Multiple properties including density, cohesive energy density, and compressibility are used to calibrate the Bayesian derived potential. We show the algorithm is able to accurately reproduce the calibration properties across the entire range of state points, and examine the effect of the machine generated force-field on other more problematic coarse-grained properties such as diffusion and modulus. We conclude with an outlook at how these analyses can applied to other properties in more complicated material systems to better understand the boundaries of applicability of these CG potentials.