Jointly Embedded Machine Learning Approach for Polymer Combustion Properties

High energy density polymeric binders are a class of polymer materials that can be used in lieu of inert binders in high energy density mixtures. By using higher energy binders, the overall internal energy of the mixture can be designed intentionally and proactively. In this presentation, we will showcase our recent efforts to develop a machine learning approach to learn, predict, and design novel energetic polymers. The scarcity of data available for energetic polymers is a particular challenge that we overcome through transfer learning techniques. Generally-speaking, transfer learning is a class of machine learning algorithm that assists the learning of general trends within one dataset using alternate datasets. In our approach, we use a feature transfer learning approach based on low-level physiochemical data that may be obtained for any molecule. We first train the model to learn the form of repeat unit structures using an open synthetic dataset containing 1 million polymer repeat units. Then, the model is trained on a dataset containing <170 polymer repeat units for which thermochemical properties are known. The model is then developed to perform generation functions and new polymers may be proposed with desirable attributes. The resulting machine-learned molecule property estimates are then compared with theoretical thermochemical models. Through the transfer learning approach, we will also address the importance of synthesizability and discuss how the proposed techniques can be used to increase the likelihood of developing synthesizable candidate polymer molecules.