Evaluating Polymer Representations via Quantifying Structure-Property Relationships.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Machine learning (ML) is increasingly used for structure-property relationship (SPR) quantification in materials.
• Effective material representation is crucial for ML model performance.
• Applications in specialized domains like polymers are hindered by a lack of benchmark databases and suitable representations.

Purpose of the Study:

• To investigate and compare different polymer representations for SPR quantification.
• To evaluate the impact of various representation learning schemes on polymer property prediction.
• To establish a benchmark for polymer representation quality using a subset of the PolyInfo database.

Main Methods:

• Utilized a benchmark database derived from the PolyInfo polymer database.
• Investigated three polymer representations: Morgan fingerprint (MF), molecular embedding (ME), and molecular graph (MG).
• Evaluated supervised learning (SL), semisupervised learning (SSL), and transfer learning (TL) schemes for ME generation.
• Quantified SPRs for polymer density, melting temperature, and glass transition temperature.

Main Results:

• Molecular embedding (ME) significantly outperformed MF and MG across all tested properties, achieving high R² values (e.g., 0.865 for glass transition temperature).
• Molecular graph (MG) demonstrated poor performance, likely due to limited training data.
• Different learning schemes (SL, SSL, TL) for MEs resulted in varying performance, indicating sensitivity to the learning approach.
• Combining multiple MEs did not improve predictive performance, suggesting no additional information gain.

Conclusions:

• Molecular embeddings represent the most effective approach for quantitative structure-property relationship prediction in polymers among the evaluated methods.
• The choice of learning scheme for molecular embeddings critically influences predictive accuracy.
• Further research into optimized representation learning strategies for polymers is warranted.