Plastics and sustainability in the same breath: Machine learning-assisted optimization of coarse-grained models for polyvinyl chloride as a common polymer in the built environment

Abstract

The fast-growing construction industry has a vast potential to rise to the plastics challenge by using them in both primary and recycled forms as a sustainable solution to some challenges in the built environment. Improving existing plastics and developing innovative polymers and polymer nanocomposites requires knowledge of interatomic interactions and their influence on macroscopic properties. Coarse-grained (CG) models offer a more computationally efficient alternative to their all-atom counterparts for simulating larger, more representative models of these materials. However, the parameterization and calibration process of CG force fields (CG-FFs) commonly entails solving a nonconvex optimization problem involving numerous local minima, rendering traditional optimization techniques impractical and iterations based on educated guesses inefficient. Here, we develop an approach to efficiently parameterize a CG-FF by coupling a metaheuristic algorithm as the calibrator (optimizer) with support vector regression-based surrogate models trained using molecular dynamics data. The merit of the approach is demonstrated by parameterizing a CG-FF potential for polyvinyl chloride (PVC) as a representative general-purpose plastic with many applications in the construction industry. The generalizability of the CG-FF to large PVC models in both pristine and carbon nanotube-filled composite forms is demonstrated. The CG-FF also accurately reproduces glass transition temperature and thermal conductivity as unseen properties of PVC.