Developing efficient deep learning model for predicting copolymer properties.

Area of Science:

• Polymer Science
• Materials Science
• Computational Chemistry

Background:

• Deep learning models show promise for predicting polymer properties.
• Model performance depends heavily on network topology and training data volume.
• Lack of standardized protocols for architecture selection and limited homogeneous polymer data hinder development.

Purpose of the Study:

• To address bottlenecks in deep learning model development for polymers.
• To propose strategies for efficient model creation with minimal data and hyperparameters.
• To assess the impact of topology and data volume on model performance.

Main Methods:

• Linear layer-by-layer expansion to identify optimal neural network topology.
• Feature extraction to map discrete polymer sequences to a continuous latent space.
• Implementation for predicting copolymer radius of gyration and copolymer compatibilizer properties.

Main Results:

• Demonstrated a method to systematically identify suitable deep learning architectures.
• Developed a technique to reduce the data requirements for training polymer models.
• Successfully applied the methods to predict specific sequence-defined polymer properties.

Conclusions:

• Efficient deep learning models for polymer property prediction can be built with minimal data and hyperparameters.
• The proposed strategies overcome key limitations in current deep learning applications for polymers.
• This work facilitates the rapid development of predictive models in polymer science.