Dimensionality reduction of local structure in glassy binary mixtures.

Area of Science:

• Materials Science
• Computational Chemistry
• Statistical Physics

Background:

• Supercooled liquids and glasses exhibit complex disordered microscopic structures.
• Characterizing these structures is crucial for understanding their unique properties.
• Unsupervised learning offers potential for analyzing complex structural data.

Purpose of the Study:

• To apply unsupervised learning for characterizing the disordered microscopic structure of supercooled liquids and glasses.
• To assess the effectiveness of dimensionality reduction on structural descriptors.
• To connect structural features with particle mobility in glassy binary mixtures.

Main Methods:

• Dimensionality reduction of smooth structural descriptors (radial and bond-orientational correlations).
• Analysis of principal component analysis (PCA) and neural network autoencoders.
• Investigation of glassy binary mixtures.

Main Results:

• A few collective variables capture significant structural fluctuations and correlate with particle mobility.
• Fine-grained descriptors offer better insights into mobility-relevant fluctuations but are complex.
• PCA and autoencoders yield similar results, with PCA offering better interpretability.
• Some mixtures exhibit locally favored structures, indicated by bimodal distributions.

Conclusions:

• Unsupervised learning, particularly dimensionality reduction, is effective for analyzing glassy structures.
• Structural features identified are linked to particle dynamics and local ordering.
• The choice of descriptors impacts interpretability and the ability to capture relevant fluctuations.