Machine learning discovery of high-temperature polymers.

Area of Science:

• Materials Science
• Computational Chemistry
• Polymer Science

Background:

• Predicting polymer properties like glass transition temperature (Tg) is crucial for material design.
• Existing methods for determining Tg can be time-consuming and resource-intensive.
• A need exists for efficient methods to discover polymers with specific thermal properties.

Purpose of the Study:

• To develop a robust machine learning (ML) model for predicting polymer glass transition temperatures (Tg).
• To establish a structure-property correlation for Tg using polymer chemical structures.
• To enable high-throughput screening for novel high-temperature polymers.

Main Methods:

• Collected a dataset of nearly 13,000 homopolymers from the PoLyInfo database.
• Trained a deep neural network (DNN) model using 6,923 experimental Tg values and Morgan fingerprint representations.
• Validated the model's predictive accuracy against molecular dynamics simulations and experimental data.

Main Results:

• The trained DNN model accurately predicted unknown Tg values for diverse polymer structures.
• The model demonstrated strong transferability and generalization capabilities.
• High-throughput screening of one million hypothetical polymers identified over 65,000 candidates with Tg > 200°C.

Conclusions:

• The developed ML model is effective for predicting polymer Tg and discovering new materials.
• This approach significantly expands the pool of known high-temperature polymers.
• The findings offer a powerful tool for accelerating the design and development of advanced polymers.