Umi Yamamoto1,Heyi Liang2,Kenji Yoshimoto1,Masahiro Kitabata1,Juan de Pablo2
Toray Industries, Inc.1,The University of Chicago2
Umi Yamamoto1,Heyi Liang2,Kenji Yoshimoto1,Masahiro Kitabata1,Juan de Pablo2
Toray Industries, Inc.1,The University of Chicago2
We report two latest advances in enabling fast and accurate prediction of polymer rheology using a bottom-up multi-scale simulation method and a top-down machine-learning approach.<br/>First, a multi-scale simulation method is presented for the prediction of the viscoelastic properties of entangled homopolymer melts. Starting from an atomistic model of a polymer, and introducing two coarser representations (a bead-spring model and a slip-spring model) that successively operate at longer time and length scales, we demonstrate that linear rheology of syndiotactic and atactic polystyrene melts is predicted in good agreement with experiments without requiring experimental parameters as input. Very recently, the same method is applied to another industrial polymer, Nylon 6, leading to reasonable agreement for the molecular-weight dependence of the melt viscosity.<br/>Second, we present a machine-learning framework to construct a predictive model for the dynamic moduli of linear entangled homopolymer. As an initial attempt, using computationally generated data based on Likhtman-McLeish model as a training dataset, it is shown that a straightforward supervised learning with standard algorithms (support vector machine, kernel ridge regression, etc.) provides not-too-bad prediction when an input parameter, the number of entanglements in the present case, is extrapolated. However, the predictive power is non-trivially improved by introducing a polymer physics idea into the learning procedure. Namely, by constructing individual predictive models that specialize in the prediction of distinct relaxation behavior at short, intermediate and long-time scale, and merging them into a single model using a frequency-dependent ensemble method, we are able to predict the storage and loss modulus in quantitative agreement with the training data. This approach is expected to be applicable to more complicated cases where a wider variety of input parameters, e.g. monomer chemical structures, polymer architectures, molecular weight distributions, etc. enter. <br/>Finally, integration of the above two approaches, i.e. generation of polymer rheology data using the multi-scale simulation and construction of accurate predictive models by machine-learning the simulation data, leads to an "ideal" framework where a fast and accurate predictive model for the polymer dynamic modulii can be established via purely in-silico techniques. We will discuss an outlook and latest progress on this front.