Predicting polymer properties based on wavelet transform and Transformer

With the rapid development of polymer material design, traditional experimental methods and single-scale molecular characterization face significant limitations in predicting polymer properties such as glass transition temperature (T_g). These limitations include high experimental costs and insufficient capture of nonlinear structural features. To address these challenges, this work proposes a multi-scale fusion framework that integrates wavelet transform with a Transformer deep learning model to improve the accuracy and robustness of polymer T_g prediction. The wavelet transform enables multi-level decomposition of Morgan fingerprints, extracting both low-frequency and high-frequency features to enhance information density and noise resistance in molecular representations. By incorporating the self-attention mechanism of the Transformer model, the framework achieves effective fusion of multi-scale features and captures long-range dependencies within molecular structures. Furthermore, a Bayesian optimization algorithm is introduced to adaptively adjust both wavelet decomposition levels and Transformer hyperparameters, significantly enhancing the model's generalization performance. Experimental results demonstrate that the proposed framework substantially outperforms traditional single-descriptor models in polyimide T_g prediction tasks. This study establishes a new paradigm for multi-scale feature fusion in polymer property prediction and provides a methodological foundation for high-throughput screening and cross-scale modeling of complex polymer materials.