HHT: A Polymer Property Prediction Framework Leveraging Natural Language

Predicting the macroscopic properties of polymers solely from 1D chemical sequences (e.g., SMILES) is inherently an ill-posed problem, as it neglects critical higher-order factors such as chain topology, condensed-state morphology, and processing history. To bridge this "semantic gap" between low-level structural representations and high-level physical functionalities, we propose HHT (Highly Hybrid Text), a novel framework that integrates Large Language Models (LLMs) into polymer property prediction. HHT employs a dual-stream architecture that synergizes implicit topological features extracted by a pre-trained chemical language model (PolyBERT) with explicit mechanistic reasoning generated by a scientific LLM. A dynamic Adaptive Gating Mechanism is introduced to weigh the contribution of structural versus semantic features based on task-specific dependencies. Extensive evaluations on seven benchmark datasets demonstrate that HHT outperforms many baseline models. Crucially, HHT transforms the traditional "black-box" fitting paradigm into a "whitebox" interpretable system. We first validate the model's focus on mechanistic logic via semantic attention visualization, confirming its ability to prioritize physical descriptors over statistical noise. Furthermore, through counterfactual sensitivity analysis, we demonstrate that the model correctly infers property shifts induced by textual morphological perturbations, demonstrating that natural language can serve as a high-dimensional, physically consistent feature modality for materials science.