Machine Learning Accelerated Prediction of 3D Granular Flows in Hoppers

Granular materials are crucial components in a broad variety of industrial and natural processes. However, despite their widespread importance, predicting their complex flow behaviour under different conditions remains extremely challenging. The Discrete Element Method (DEM) is the primary computational technique used to simulate granular flows, and while it can produce highly accurate predictions, it is also inherently computationally expensive. We look to overcome this computational bottleneck through the use of a neural network surrogate model for 3D granular flow simulation, and consider the industrially important use case of flow in grain hoppers. We investigate our model performance across a range of different time scales, quantifying its accuracy and generalizability to different hopper geometries. The use of deep learning techniques for the prediction of granular flow dynamics offers an excellent opportunity for providing step change increases in computational efficiency for industrial decision-making and potential application in real-time decision making in diverse manufacturing settings.