Generative Foundation Model for On-demand Reverse Polymer Design

Forward screening and reverse design of drug molecules, inorganic molecules, and polymers with better properties are crucial engines for shortening the laboratory-to-market cycle. Particularly, due to the lack of large-scale datasets, polymer discovery based on materials informatics is more formidable. Despite this, polymer scientists have developed a series of machine learning models on polymer structure-property relationships using only small polymer datasets, thereby driving the forward screening process of polymers. However, the success of this paradigm ultimately hinges on the capacity of the candidate pool, while exhaustively enumerating all polymer structures through human imagination is challenging. Therefore, achieving on-demand reverse design of polymers is crucial. In this work, we curate a polymer dataset containing nearly one million polymeric structure-property pairs based on expert intuition. Using this dataset, we propose a generative pre-trained model for polymer on-demand generation using a large language model. The model produce polymers with 99.27\% chemical validity in top-1 generation mode (approximately 200k generated polymers), marking the highest reported success rate among polymer generative models. In addition, the average $R^2$ between the properties of the generated molecules and their expected values across 15 predefined properties is 0.96. To further assess the pre-trained model's performance in generating additional user-defined polymer properties for downstream tasks, we conduct fine-tuning experiments on three publicly available small polymer datasets using semi-template and template-free generation paradigm. Through these extensive experiments, we demonstrate that our pre-trained model and fine-tuned models 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.