Scalable Bayesian Optimization for High-Dimensional Coarse-Grained Model Parameterization

Abstract

Coarse-grained (CG) force field models are extensively utilised in material simulations due to their scalability. Traditionally, these models are parameterized using hybrid strategies that integrate top-down and bottom-up approaches; however, this combination restricts the capacity to jointly optimize all parameters. While Bayesian Optimization (BO) has been explored as an alternative search strategy for identifying optimal parameters, its application has traditionally been limited to low-dimensional problems. This has contributed to the perception that BO is unsuitable for more realistic CG models, which often involve a large number of parameters. In this study, we challenge this assumption by successfully extending BO to optimize a high-dimensional CG model. Specifically, we show that a 41-parameter CG model of Pebax-1657, a copolymer composed of alternating polyamide and polyether segments, can be effectively parameterized using BO, resulting in a model that accurately reproduces key physical properties of its atomistic counterpart. Our optimization framework simultaneously targets density, radius of gyration, and glass transition temperature. It achieves convergence in fewer than 600 iterations, resulting in a CG model that shows consistent improvements across all three properties.