Developing Hybrid Machine Learning Frameworks for Polymer Property Prediction Based on Composition and Sequence...

Area of Science:

• Polymer Science and Engineering
• Materials Science
• Computational Chemistry

Background:

• The glass transition temperature (Tg) is crucial for polymer physical properties.
• Understanding the relationship between polymer composition/sequence and Tg is vital for material design.
• Existing methods for predicting Tg face challenges due to complex structure-property relationships.

Purpose of the Study:

• To investigate the influence of polymer composition and sequence structure on Tg using machine learning (ML).
• To develop and validate advanced ML models for accurate Tg prediction.
• To introduce novel data augmentation and sequence representation techniques for polymer science.

Main Methods:

• Utilized k-nearest neighbor mega-trend diffusion (kNNMTD) for polymer composition data augmentation.
• Employed Random Forest models for Tg prediction based on composition, achieving R²=0.85.
• Applied natural language processing (NLP) techniques and Wasserstein generative adversarial network (WGAN-GP) for polymer sequence representation and augmentation.
• Developed a convolutional neural network-long short-term memory (CNN-LSTM) model for sequence-based Tg prediction, achieving R²=0.95.

Main Results:

• The Random Forest model showed strong performance in predicting Tg from composition data.
• The integrated framework combining NLP and CNN-LSTM achieved high predictive accuracy (R²=0.95, RMSE=0.23) for sequence-based Tg.
• The models demonstrated excellent generalization capabilities across diverse polymer datasets.

Conclusions:

• This study presents an innovative ML framework for predicting polymer Tg by integrating advanced data augmentation and sequence representation techniques.
• The developed models offer a powerful tool for accelerating polymer material design and optimization.
• The findings highlight the potential of AI and NLP in advancing polymer science and engineering.