Highly accurate prediction of viscosity of epoxy resin and diluent at various temperatures utilizing machine learning

Abstract

Obtaining quantitative structure-property relationships (QSPR) is crucial for the development of new materials, which also helps to reduce the number of trial and improve the efficiency for both research and development. The viscosity of epoxy resin is vital for processing and application, for example, low viscosity can be used as coatings while high viscosity as adhesives. However, due to the wide variety of epoxy resin and its additives, the resin with target viscosities cannot be easily designed and the viscosity cannot be precisely predicted directly from massive formulation of epoxy resin. In the present work, we propose a simple strategy to accurately predict the viscosity of epoxy resin for a wide range of epoxy resins leveraging machine learning (ML) and deep learning (DL). The coarse-grained (CG) methodology is applied to the dataset first and then the dataset is categorized via K-Means clustering algorithm. A high-precision prediction is thus achieved with R2 up to 1.00 among 10 of the classes on train sets. To build a more generalized model without clustering, we compare 5 ML and DL models to select the optimal model under multidimensional evaluations. A prediction model with R2 of 0.96 on the test set is obtained using TensorFlow framework. We further employ our model to predict the viscosity of a commonly used diglycidyl ether of bisphenol-A (DGEBA) epoxy with different diluent proportions at different temperatures, and then we verify the predicted data by using several empirical viscosity equations. As a consequence, the activation energy of DGEBA can be estimated from the relation between viscosity and temperature, and the calculated value (56.40 kJ-mol−1) agrees well with the experimental data (58.16 kJ-mol−1). Our work reveals the great potential of machine learning methods in the prediction of QSPR in materials science.