Glass Transition Temperature Prediction of Polymers via Graph Reinforcement Learning

An expansive array of graph-based models has been utilized for accurate prediction of the structure-property relation of polymers. However, these approaches notably underutilize unsupervised structural information. Concentrating on the domain of heterocyclic polymers, particularly polyimides, this study delves into the glass transition temperature (<i>T</i>_g) prediction, aiming to fully exploit the potential within both the global and local structures of molecules. To achieve this, a graph reinforcement learning framework termed Molecular Structural Regularized Graph Convolutional Network with Reinforcement Learning (MSRGCN-RL) is proposed. Experimental results highlight the crucial role of both global and local structural regularization in precise <i>T</i>_g prediction. Concurrently, optimization of MSRGCN training through RL proves essential. This research leads the way in integrating Graph Neural Networks (GNNs) with reinforcement learning methodologies for the property prediction of polymers.