Introducing PolySea: An LLM-Based Polymer Smart Evolution Agent

This is the authors' abstract. We don't add key points for this paper.

The integration of artificial intelligence with materials science has opened new frontiers in accelerated materials discovery. However, general-purpose large language models (LLMs) often struggle with domain-specific challenges, necessitating the development of specialized models. Here, we introduce PolySea, a domain-specific LLM tailored for polymer informatics, designed to address key limitations in polymer property prediction, inverse design, and knowledge extraction. PolySea is trained on a meticulously curated dataset, integrating high-fidelity polymer property data from PolyInfo with structured polymer knowledge distilled from expert-curated sources. By leveraging LoRA-based fine-tuning, we mitigate catastrophic forgetting while enhancing computational efficiency, ensuring optimal retention of both general linguistic capabilities and polymer-specific knowledge. PolySea demonstrates state-of-the-art performance across diverse polymer-related tasks. On regression benchmarks, it achieves an R² score of 0.97, while delivering 79% classification accuracy in thermal stability prediction. Comparative assessments against leading general-purpose LLMs—including ChatGPT-o1 and DeepSeek-R1—highlight PolySea’s superior precision, particularly in on-demand polymer design, where it generates novel polymer structures unseen in training yet aligned with target property constraints. The generated polymers are rigorously validated using a graph neural network surrogate model, Polymer Genome and density functional theory experiments, confirming their feasibility for real-world synthesis. Our findings underscore the transformative potential of domain-adapted LLMs in accelerating polymer informatics. By bridging the gap between AI and materials science, PolySea not only establishes a new paradigm for polymer design but also paves the way for the development of specialized AI models across broader scientific disciplines.