Insights on the internal dynamics of bi-disperse granular flows from machine learning

In granular flows, grains exhibit heterogeneous dynamics featuring large distributions of forces and velocities. Conventional statistical methods have previously revealed how these dynamical properties scale with the grain size in monodisperse flows. We explore here whether they differ between small and large grains in bi-disperse flows. In simulated silo flows comprised of dense and collisional zones, we use a machine learning classifier to attempt to distinguish small from large grains based on features such as velocity, acceleration and force. Results show that a classification based on grain velocity is not possible, which suggests that large and small grains feature statistically similar velocities. In the dense zones, classification based on force only fails too, indicating that small and large grains are subjected to similar forces. However, classification based on force and acceleration succeeds. This indicates that the classifier is sensitive to the correlation between forces and acceleration, i.e. Newton's second law, and can thus detect differences in grain size via their mass. These results highlight the potential for machine learning to assist with better understanding the behaviour of granular flows and similar disordered fluids.

Automated summary

This study explores the differences between small and large grains in bi-disperse flows and attempts to distinguish them based on features such as velocity, acceleration, and force using a machine learning classifier. The results show that classification based on grain velocity is not possible, indicating statistically similar velocities for small and large grains. In the dense zones, classification based on force only fails too, but succeeds when both force and acceleration are considered, highlighting the sensitivity of the classifier to the correlation between forces and acceleration.