Comparing machine learning techniques for predicting glassy dynamics.

Area of Science:

• Condensed matter physics
• Computational materials science
• Statistical mechanics

Background:

• Understanding the relationship between structure and dynamics in glasses is crucial.
• Machine learning (ML) methods are increasingly used to predict dynamics in supercooled liquids.
• Comparing different ML techniques is challenging due to simultaneous variations in algorithms and structural descriptors.

Purpose of the Study:

• To quantitatively compare the performance of different ML algorithms for predicting dynamics in glasses.
• To evaluate the effectiveness of advanced structural descriptors in ML models for glass dynamics.
• To identify the most efficient ML approach for predicting dynamic propensity.

Main Methods:

• Utilized three ML algorithms: linear regression, neural networks, and graph neural networks.
• Employed a recursive set of order parameters as structural input, as introduced by Boattini et al.
• Applied these methods to predict the dynamic propensity of a glassy binary hard-sphere mixture.

Main Results:

• All three ML methods achieved nearly equal accuracy in predicting dynamics when using advanced structural descriptors.
• Linear regression demonstrated significantly faster training times compared to neural networks and graph neural networks.
• The choice of advanced structural descriptors proved more impactful than the complexity of the ML algorithm.

Conclusions:

• Advanced structural descriptors enhance the predictive power of various ML algorithms for glass dynamics.
• Linear regression offers a computationally efficient and accurate alternative for predicting dynamics in supercooled liquids.
• The findings suggest linear regression is the preferred method when balancing accuracy and computational cost in glass dynamics prediction.