Dual Embedding: A Fine-Tuned Language Model Approach for Accurate Polymer Glass Transition Temperature Prediction

Recent years have witnessed major advances in polymer informatics, yet accurately predicting polymer properties, such as the glass transition temperature (T_g), remains a challenge. Language models like BERT have been leveraged to derive embeddings from polymer representations (e.g., SMILES). However, similarity between embedding vectors in these latent spaces primarily reflects chemical structural similarity, with limited alignment to physicochemical properties. Here, we introduce a dual-embedding framework that enhances T_g prediction by combining a conventional BERT-based embedding with a fine-tuned counterpart explicitly trained so that vector similarity reflects proximity in T_g values. We evaluate our approach across four benchmarks: a heterogeneous data set compared against 25 machine learning baselines, along with three additional data sets focused on homopolymers and polyimides. The dual embedding outperforms standard BERT-based embeddings, achieving up to a 20% reduction in RMSE and surpassing alternative models such as graph-based and descriptor-based approaches. These results demonstrate that embedding molecular properties directly into representations can advance polymer informatics beyond structure-centric paradigms.