Digging deep : A data-driven approach to model reduction in a granular bulldozing scenario

Abstract: The current simulation method for granular dynamics used by the physics engine AGX Dynamics is a nonsmooth variant of the popular Discrete Element Method (DEM). While powerful, there is a need for close to real time simulations of a higher spatial resolution than currently possible. In this thesis a data-driven model reduction approach using machine learning was considered. A data-driven simulation pipeline was presented and partially implemented. The method consists of sampling the velocity and density field of the granular particles and teaching a machine learning algorithm to predict the particles' interaction with a bulldozer blade as well as predicting the time evolution of its velocity field. A procedure for producing training scenarios and training data for the machine learning algorithm was implemented as well as several machine learning algorithms; a linear regressor, a multilayer perceptron and a convolutional neural network. The results showed that the method is promising, however further work will need to show whether or not the pipeline is feasible to implement in a simulation.