Machine learning can be used to predict the behaviour of glass under different temperature and pressure constraints based solely on the initial positions of the individual particles it is made of, reports a paper published this week in Nature Physics. These findings may provide a deeper understanding of the mechanical properties of a range of systems including granular materials, colloidal suspensions and biological cells.
Glassy systems typically behave as solids, yet appear like liquids on the microscopic level. Their particles are arranged in a disordered manner. Understanding the precise causes for the extremely slow dynamics of glasses has long been a challenge facing physicists and materials scientists.
Victor Bapst and colleagues use a category of machine learning models known as graph neural networks to predict the dynamics of glassy systems. The authors developed an algorithm that learns a glassy system’s features and corresponding physical properties by using the system’s type of particles and position as the only input. Most notably, this method captures the interactions between particles that are both close together and farther apart. Therefore, the algorithm can predict the locations and movement of the rearranging particles over very long time scales, for a range of different temperatures, pressures and densities. As such, it outperforms existing machine learning approaches used to study this problem.
The authors conclude that this algorithm may be robust enough to apply to other systems beyond glasses.
Planetary science: Modelling electrolyte transport in water-rich exoplanetsNature Communications
Robotics: Taking millimetre-scale origami robots for a spinNature Communications