De Novo Design of Polyimides Leveraging Deep Reinforcement Learning Agent

Abstract Designing molecular structures has long been a central pursuit in organic films with super properties. However, the vast chemical space of candidate molecules poses a challenge in screening optimal materials with exceptional performance. Herein, a multi‐objective performance‐oriented strategy driven by deep reinforcement learning and train an agent, DAPiGen is proposed, for a de novo template‐free polyimide creation. The agent combines the property predictors identified from four machine learning models and a fragment‐based generation architecture with active fragments extracted from polyimides as fundamental building blocks. Its successful use is demonstrated to create several polyimides for flexible display scenarios, i.e., with specific properties such as higher transparency, lower coefficient of linear thermal expansion, superior tensile strength, and elevated glass transition temperature. Experiment validation and structural importance analysis demonstrate the efficacy and reliability of the proposed research approach. The scalable strategy presented in this work stands as a paradigm for the inverse design of a spectrum of polymeric materials, offering guidance for other structural engineering endeavors.