Learning glass transition temperatures via dimensionality reduction with data from computer simulations: Polymers as...

Area of Science:

• Materials Science
• Computational Chemistry
• Polymer Physics

Background:

• Machine learning offers advanced pattern recognition in complex datasets.
• Understanding the glass transition temperature (Tg) is crucial for polymer performance.
• All-atom molecular dynamics simulations provide detailed insights into polymer behavior.

Purpose of the Study:

• To evaluate the glass transition temperature (Tg) using machine learning on molecular dynamics simulations.
• To compare the efficacy of Principal Component Analysis (PCA) and Diffusion Maps (DM) for Tg determination.
• To explore various molecular descriptors for capturing polymer state transitions.

Main Methods:

• Employed all-atom molecular dynamics simulations for polylactide (PLA) and poly(3-hydroxybutyrate) (PHB).
• Utilized Principal Component Analysis (PCA) and Diffusion Maps (DM) for dimensionality reduction.
• Applied Gaussian Mixture Models (GMMs) to analyze low-dimensional representations and quantify log-likelihoods.
• Calculated Tg by observing the overlap of log-likelihood distributions during simulated cooling.

Main Results:

• Diffusion Maps (DM) with radial distribution functions (RDF) and mean square displacements (MSDs) accurately predicted Tg for PLA and PHB, matching simulation data.
• DM-transformed dihedral angle (DA) and relative square displacement (RSD) data yielded Tg values consistent with experimental findings.
• Principal Component Analysis (PCA) showed less reliable Tg predictions compared to DM across the tested descriptors.
• A clear separation into melt and glass states was identified using GMMs on PCA and DM projections.

Conclusions:

• The combination of atomistic simulations and Diffusion Maps (DM) with Gaussian Mixture Models (GMMs) provides a robust framework for calculating Tg.
• DM demonstrates superior performance over PCA in predicting Tg from molecular dynamics data for the studied polymers.
• This integrated approach offers a unified method for studying the glass transition across diverse molecular descriptors in glass-forming materials.