Developing Universal Machine Learning Model for Predicting Polymer Properties

Machine learning models are gaining popularity and potency in predicting polymers' properties. These models can be built using pre-existing polymer structure-property data. They are very useful for rapid prediction of their properties for unknown structures and, thus, potentially accelerate their design. However, building an efficient, transferable, and universal machine learning model requires addressing several challenges. First, there are no general guidelines for selecting model architecture, hyperparameters, and polymer fingerprinting. Second, they need large volumes of homogeneous structure-property data, which is not readily available for many polymer design problems. Third, these models are interpolative in nature. Their extrapolation capability is poorly understood. To address these problems, here we propose a simple approach to build computationally cheap high-fidelity deep neural network models with optimal hyperparameters, introduce efficient polymer fingerprinting and data selection algorithm for universal model building. We implement these approaches for predicting several polymer properties, including single molecule radius of gyration, polymer compatibilizer, adhesion-free energy, and glass transition temperature, demonstrating the generality of these strategies.