Generating High-Temperature Polymer Dielectrics with a Graph Convolution and Reinforcement Learning Framework

Generating molecular structures with certain desired properties is a fundamental problem for material science research. Deep learning models have demonstrated big potential to identify novel materials by exploring large chemical spaces efficiently. Here, we demonstrate the generation of hypothetical polymers with a high glass-transition temperature (Tg) and a wide band gap (Eg) using a general graph convolutional network-based model through reinforcement learning. Two forward property prediction models on Tg and Eg are employed to provide domain-specific rewards. We further incorporate prior knowledge in the form of adversarial rewards given by a generative adversarial network framework via a data set of exemplar polymer structures. The domain-specific rewards and adversarial rewards are optimized in a reinforcement learning framework through policy gradient methods. Rules of chemistry are incorporated in the molecule generation environment as hard constraints. The generation process is guided to generate promising high-temperature polymer dielectrics for energy storage applications by property targeting and constrained property optimization. This work demonstrates potential for exploring large chemical spaces for the discovery of polymers with desirable properties.