Machine-Learning-Enabled Framework in Engineering Plastics Discovery: A Case Study of Designing Polyimides with...

Area of Science:

• Materials Science
• Polymer Chemistry
• Computational Chemistry

Background:

• High-performance engineering plastics are crucial for replacing traditional materials.
• Machine learning (ML) offers potential for discovering novel polymers but faces challenges with data limitations and descriptor accuracy.
• Current ML approaches in polymer development often lack accuracy and efficiency.

Purpose of the Study:

• To develop a rapid and accurate ML approach for designing polyimides (PI) with specific glass-transition temperatures (Tg).
• To identify and interpret key structural descriptors for predicting PI properties.
• To establish a scalable framework for engineering plastics innovation.

Main Methods:

• Collected a dataset of 878 polyimides with experimentally measured glass-transition temperatures.
• Converted polymer structures into "mechanically identifiable" SMILES strings.
• Utilized multiple analysis methods to obtain eight critical descriptors and employed an artificial neural network (ANN) model.

Main Results:

• An ANN-based model achieved high accuracy with a root-mean-square error of approximately 11 K.
• Identified and analyzed the physicochemical meaning of eight critical descriptors.
• Successfully designed and experimentally verified three polyimide candidates (DPIs) with desired Tg values, showing an average deviation of 3.66%.

Conclusions:

• The developed ML approach significantly reduces time and computational resources compared to traditional molecular simulation.
• The study provides a scalable and adaptable framework for future engineering plastics innovation.
• The identified descriptors offer chemical insights, translating "machine language" into practical chemical knowledge.