Data-Driven Modeling of Granular Column Collapse

Data-driven methods have recently shown their potential in engineering problems. In this paper, a data-driven model is proposed and tested for 2D granular column collapse, a typical problem involving flow-like behavior of granular material. Although many theoretical and numerical models have been developed to investigate granular flows, data-driven methods have yet to be explored. This model can predict granular flows based on the sequential relations over time with data collected from numerical solutions. Two frameworks, a Lagrangian framework and an Eulerian framework, are proposed to give rules for collecting data, and radial basis function networks are used to infer the sequential relations. It is shown that the data-driven model gives reasonable predictions compared to results obtained through direct solution. When the size of data in both frameworks are close, the performance in the Eulerian framework is better, especially for predicting states of columns with different geometric parameters. A sensitivity analysis to explore the influence of time increment and grid spacing is also conducted.