Reduction of number of simulations for parameters identification of DEM models using neural network and design of experiments

ABSTRACT

Discrete Element Method (DEM) has been a tool used to simulate the flow of granular material, covering a wide range of industries. For DEM modeling results to be representative, it is necessary to identify the values of the contact law parameters of the material. In this procedure, a major difficulty is the computational cost when handling small particles of irregular shapes. Thus, in DEM simulations, it is common practice to use upscaled spheres particles with the inclusion of contact rolling friction parameter to increase the shear strength. For typical calibration of DEM parameters using upscaled particles, laboratory experiments are performed, and one or more material’s bulk properties (macroscopic properties) are measured. Then, the contact law parameters are adjusted until the prediction behavior of the bulk be met (generally the angle of repose). In general, this parameter identification process can take a long time, as many numerical simulations will be required due to the multi-dimensionality of the parameter space. In order to reduce the number of DEM simulations to determine an adequate set of input parameters, this work presents a method using Design of Experiments (DOE) for the DEM simulations using only 1/16 of a Full Factorial. This set of DEM simulations allows the creation of an Artificial Neural Network (ANN) that simulates the DEM simulations, making the remained simulations of the full factorial to be performed by the neural network created. Thus, with the neural network regression model, an appropriate set of input parameters for the DEM model is determined, which provides the desired macroscopic behavior of the particulate material. The proposed method was applied in a case study from literature. As a result, the number of DEM simulations was reduced by 66.7% to identify a set of contact parameters for predicting the angle of repose of a cohesive iron ore.

Keywords
DEM model; parameter identification; neural networks; design of experiments