Universal Prediction of Molecular Weight Distribution for Entangled Linear Polymers from Linear Rheology Using Neural Networks

Abstract We present a machine learning methodology for inferring the molecular weight distribution (MWD) of arbitrary polydisperse entangled linear polymers from their linear rheology. Using a state-of-the-art tube model, we generate large synthetic rheology datasets to train neural networks (NNs) to accurately predict MWDs from frequency-sweep rheology measurements. We take advantage of the universality of the dynamics of flexible polymers, where the shape of the relaxation spectrum is identical for polymers with different backbone chemistries but the same number of entanglements. We therefore use the NN to predict the distribution in the number of entanglements from normalised rheology data, and this distribution is subsequently converted to an MWD using the mean entanglement molecular weight. Our primary interest is in industrially relevant materials, so we focus on broad, highly polydisperse MWDs. However, we also introduce a classification scheme to determine, solely from the rheology, whether a sample is best described as polydisperse, monodisperse, or bidisperse, and deploy dedicated NNs for each of these cases. We present resulting predictions for a range of industrially relevant polymer chemistries, including commercial samples of polyethylene (HDPE and LLDPE), with good agreement with experimental gel permeation chromatography (GPC) data. An open-access GUI software tool is released alongside this work to enable easy adoption of our methodology in both academic and industrial settings.