Fast and Accurate Prediction of Polymer Viscoelasticity via Physics-Based Ensemble Learning

We present a machine learning framework to construct a predictive model for dynamic moduli of linear entangled homopolymers. Using well accepted computational data as a training data set, it is shown that a straightforward supervised learning with standard algorithms (support vector machine, kernel ridge regression, etc.) provides reasonable prediction accuracy when an input parameter, the number of entanglements in the present case, is extrapolated. However, the predictive power is further improved to a non-trivial extent by integrating a polymer physic s idea with the learning procedure. Namely, by constructing individual predictive models that specialize in the representation of the distinct relaxation behavior at short, intermediate and long time scale, and merging them into a single model using a frequency dependent ensemble method, we can predict the storage and loss modulus in quantitative agreement with the training data.