Prompt2Poly: Ask, Specify, Create – A Dialogue-Based Large Language Model for Targeted TSMPs Design

Abstract

Designing polymers with targeted properties is a complex task that often relies on expert knowledge and multi-step computational workflows. Existing approaches typically require rigid input formats and post hoc property filtering, limiting accessibility and responsiveness to user intent. We introduce Prompt2Poly, the first conversational framework that leverages large language models (LLMs) to generate thermoset shape memory polymers (TSMPs) based on user-defined property requirements. Prompt2Poly allows users to specify desired properties, such as glass transition temperature (Tg), rubbery modulus (Er), or particular chemical groups — directly through natural language prompts, enabling guided generation of polymer samples. The framework interprets these prompts, extracts the target constraints, and generates chemically valid, two monomer polymer structures aligned with the specified goals. Built on fine-tuned LLMs, Prompt2Poly supports multi-turn dialogue, enabling iterative refinement and deeper alignment with user intent. Using a curated dataset of TSMPs, we demonstrate that the framework generates novel, diverse, and property-consistent polymer candidates across various scenarios. Compared to strong base models (GPT-4 and Llama-3.2, prior to fine-tuning), Prompt2Poly achieves notable gains in chemical validity, novelty, and alignment with chemical groups and target Tg, and Er. By integrating property-aware generation and conversational interaction, Prompt2Poly demonstrates the potential of human-AI collaboration in materials science, making TSMPs design more intuitive, flexible, and accessible.