Augmenting Polymer Datasets by Iterative Rearrangement.

Area of Science:

• Polymer Science
• Materials Informatics
• Computational Chemistry

Background:

• Accurate polymer property prediction relies on effective molecular representations.
• Existing representations may not fully capture sequence-dependent information.
• Data augmentation techniques have shown promise in other scientific domains.

Purpose of the Study:

• To investigate data augmentation via molecular representation rearrangement for polymer property prediction.
• To assess if augmenting polymer data improves machine learning model performance.
• To compare augmented representations against common molecular representations.

Main Methods:

• Iterative rearrangement of polymer molecular representations while preserving connectivity.
• Training machine learning models on three polymer datasets using original and augmented representations.
• Evaluating model performance and comparing augmented embeddings to non-augmented ones.

Main Results:

• Data augmentation did not yield significant improvements over equivalent non-augmented representations.
• Augmentation provided molecular embeddings with more information in sequence-dominated datasets.
• Performance gains were observed when target properties were primarily influenced by polymer sequence.

Conclusions:

• Polymer data augmentation via representation rearrangement shows limited general improvement in property prediction.
• This technique can enhance molecular embeddings for sequence-dependent properties.
• Further research may explore optimized augmentation strategies for polymer informatics.