UniPolymer Framework Accelerates Polyimide Design with AI

This paper introduces UniPolymer, a comprehensive and unified AI framework designed to streamline the challenging process of designing polyimide structures with specific glass transition temperatures (Tg). Traditional methods often focus solely on generating structures based on targets, lacking a robust mechanism to assess the consistency between the generated structure and the desired properties. This leads to inefficient development cycles due to low-quality candidates entering expensive experimental stages. UniPolymer addresses this by integrating property prediction, target-conditioned generation, candidate evaluation, and structure recommendation. It first establishes a reliable structure-property relationship through self-supervised chemical semantic learning, structural consistency enhancement, and multi-scale information fusion. This foundational step ensures a strong understanding of how molecular structure relates to Tg. Subsequently, the framework employs a continuous-discrete joint Tg representation to guide the autoregressive generation of SELFIES (Simplified Molecular-Input Line-Entry System) for new polyimide candidates. These generated structures are then rigorously evaluated using a frozen property predictor and polyimide-specific structural constraints. Candidates are ranked based on their deviation from the target Tg, preventing unsuitable structures from proceeding to high-cost experimental validation. Experimental results demonstrate UniPolymer's effectiveness, achieving an R^2 of 0.93 for property prediction and a 73.79% candidate evaluation pass rate, outperforming baselines and significantly reducing the number of invalid experiments.