On-Demand Reverse Design of Polymers with PolyTAO

The forward screening and reverse design of drug molecules, inorganic molecules, and polymers with enhanced properties are vital for accelerating the transition from laboratory research to market application. Specifically, due to the scarcity of large-scale datasets, the discovery of polymers via materials informatics is particularly challenging. Nonetheless, scientists have developed various machine learning models for polymer structure-property relationships using only small polymer datasets, thereby advancing the forward screening process of polymers. However, the success of this approach ultimately depends on the diversity of the candidate pool, and exhaustively enumerating all possible polymer structures through human imagination is impractical. Consequently, achieving on-demand reverse design of polymers is essential. In this work, we curate an immense polymer dataset containing nearly one million polymeric structure-property pairs based on expert knowledge. Leveraging this dataset, we propose a Transformer-Assisted Oriented pretrained model for on-demand polymer generation (PolyTAO). This model produces polymers with 99.27% chemical validity in top-1 generation mode (approximately 200k generated polymers), representing the highest reported success rate among polymer generative models. Additionally, the average R2 between the properties of the generated polymers and their expected values across 15 predefined properties is 0.96. To further evaluate the pretrained model's performance in generating polymers with additional user-defined properties for downstream tasks, we conduct fine-tuning experiments on three publicly available small polymer datasets using both semi-template and template-free generation paradigms. Through these extensive experiments, we demonstrate that our pretrained model and its fine-tuned versions are capable of achieving on-demand reverse design of polymers with specified properties, whether in semi-template generation or the more challenging template-free generation scenarios, showcasing its potential as a unified pretrained foundation model for polymer generation.