Benchmarking Machine Learning Models for Polymer Informatics: An Example of Glass Transition Temperature

Area of Science:

• Polymer Informatics
• Computational Materials Science
• Machine Learning Applications

Background:

• Machine learning (ML) offers an efficient alternative to experimental methods for predicting polymer properties like glass transition temperature (Tg).
• Existing ML models for Tg prediction vary in datasets, structure representations, and feature engineering, hindering direct comparison and optimal model selection.
• A systematic benchmark is needed to evaluate different ML techniques and identify critical factors for accurate and generalizable Tg prediction.

Purpose of the Study:

• To conduct a comprehensive benchmark of 79 diverse ML models for polymer Tg prediction.
• To identify the most influential components in ML model development: structure representation, feature representation, and ML algorithms.
• To assess the generalization ability of ML models on unlabeled data and their sensitivity to polymer topology and molecular weight.

Main Methods:

• Compiled and trained 79 distinct ML models using a large, diverse polymer dataset.
• Investigated various polymer structure representations (monomer, repeat unit, oligomer) and feature representations (e.g., Morgan fingerprinting, RDKit descriptors, molecular embeddings, graph representations).
• Employed diverse ML algorithms including deep neural networks, convolutional neural networks, random forest, support vector machines, LASSO regression, and Gaussian process regression.

Main Results:

• Evaluated model performance on both holdout test sets and an unlabeled dataset from molecular dynamics simulations.
• Demonstrated that the choice of structure representation, feature engineering, and ML algorithm significantly impacts Tg prediction accuracy.
• Highlighted the importance of generalization ability for ML models applied to polymer informatics tasks.

Conclusions:

• This benchmark provides a clear guideline for selecting appropriate ML models for polymer Tg prediction.
• The findings offer valuable insights for optimizing ML model development in polymer informatics beyond Tg prediction.
• Understanding the impact of different components is crucial for advancing data-driven polymer science.