Predicting the crystalline phase generation effectively in monosized granular matter using machine learning

When monosized granular matter is subjected to continuous mechanical disturbance, crystallization can be observed. The granular crystallization process remains elusive and difficult to capture and forecast because of the complex interactions of particles and long periods of evolution. This study aims to establish a machine learning model that can effectively identify the crystalline phase precursors during the granular crystallization process. We simulate the cyclic shear test of a monosized sphere packing using the discrete element method. A machine learning (ML) model for predicting the generation of the crystalline phase is developed from particles' local structural information using the eXtreme Gradient Boosting algorithm. The predictive power of the ML model shows significant prediction horizon dependence. The local volume fraction is identified as one of the most important structural signatures in the crystalline phase generation. Our work presents a general and data-centric framework that could be used for granular crystallization problems.

Automated summary

The study aims to establish a machine learning model that can identify the precursors of crystalline phases during the granular crystallization process. Through simulation experiments and the use of discrete element method, the study finds that the local volume fraction is an important structural signature in the generation of crystalline phases.