Machine learning workflow for analysis of high-dimensional order parameter space: A case study of polymer...

Area of Science:

• Materials Science
• Computational Chemistry
• Polymer Science

Background:

• Traditional methods for identifying polymer crystallization rely on single order parameters (OPs) with preset cutoffs, leading to sensitivity issues and systematic biases.
• Accurate quantification of crystallinity is crucial for understanding polymer behavior and processing.

Purpose of the Study:

• To develop an integrated machine learning workflow for accurate crystallinity quantification in polymers using atomistic molecular dynamics simulation data.
• To identify a minimal set of order parameters that can reliably capture crystallization labels.

Main Methods:

• Representing each atom with a high-dimensional feature vector combining geometric, thermodynamic-like, and symmetry-based descriptors.
• Employing low-dimensional embeddings and unsupervised clustering to identify crystalline and amorphous atoms.
• Utilizing supervised learning to determine the minimal set of order parameters (q6, S̄i, p2) for accurate classification.

Main Results:

• A machine learning workflow accurately quantifies polymer crystallinity, achieving >98% classification performance using only three order parameters (q6, S̄i, p2).
• The crystallinity index (C-index), derived from logistic regression, provides a robust, bimodal measure of crystallinity.
• A trained model allows for efficient on-the-fly crystallinity computation.

Conclusions:

• The developed workflow offers a data-driven strategy for selecting order parameters and a generalizable metric for monitoring structural transformations in polymer simulations.
• The findings support the hypothesis that entropy dominates early nucleation, with q6 becoming more relevant in later crystallization stages.