Machine learning strategies for the structure-property relationship of copolymers

Area of Science:

• Polymer Science
• Materials Science
• Computational Chemistry

Background:

• Establishing structure-property relationships is crucial for copolymer molecular design.
• Existing machine learning (ML) models lack a unified approach to incorporate both chemical composition and sequence distribution for various copolymer types.

Purpose of the Study:

• To develop and validate ML models capable of processing diverse copolymer types (alternating, random, block, gradient) using a unified approach.
• To investigate the performance and generalizability of different ML architectures for copolymer property prediction.

Main Methods:

• Formulated four ML models: feedforward neural network (FFNN), convolutional neural network (CNN), recurrent neural network (RNN), and a combined FFNN/RNN (Fusion) model.
• Systematically validated model performance and generalizability using various copolymer types.

Main Results:

• The recurrent neural network (RNN) architecture demonstrated superior performance and generalizability.
• RNN models effectively process monomer sequence information bidirectionally, leading to better predictions for diverse copolymer structures.

Conclusions:

• Recurrent neural networks (RNNs) are highly suitable ML models for predicting copolymer properties, offering enhanced generalizability.
• The proposed ML approach provides an efficient tool for copolymer evaluation, complementing polymer informatics efforts.