Abstract: Bayesian calibration/validation and uncertainty propagation for discrete particle models of granular materials
In granular materials, uncertainty exists at contact, microstructural and continuum scales. To create predictive virtual prototypes of complex granular processes, the new challenge for particle simulations is to allow for probabilistic predictions which quantifies all sources of uncertainty and propagate them to macroscopic quantities of industrial interest. We present a material informatics framework, aided by advanced machine learning algorithms, to bridge particle models and experiments with Bayesian inference.
The framework utilizes state-of-the-art Bayesian inference and machine learning methods to quantify and model uncertainties at multiple scales in granular materials. The novel Bayesian calibration/validation tool recursively updates the posterior distribution of model parameters in time and iterates the process with new sets of parameters drawn from a proposal density. Over iterations, the proposal density is progressively localized near the posterior modes. The Dirichlet process Gaussian mixture is trained for estimating the proposal density from sparse and high dimensional data.
We show how to practically use the Bayesian calibration tool to quantify the probability distribution of micromechanical parameters, conditioned on the observations obtained from angle of repose, rotating drum and direct shear experiments. We further propagate the parameter uncertainties to the macroscopic quantities of interest to showcase the micro-macro capability of the framework
Dr. Hongyang Cheng is a postdoctoral researcher in the Multi-Scale Mechanics group in the Faculty of Engineering Technology and the MESA+ Institute for Nanotechnology of the University of Twente. He holds a PhD and a M.Eng. degree in geotechnical engineering of Hiroshima University in Japan. His research interests lie in the predictive computational science for granular materials, for which advanced experiments (e.g., micro-CT imaging), numerical simulations (e.g., DEM/LBM/FEM coupling) and Bayesian inference (e.g., sequential Monte Carlo methods) are the three cornerstones. He has developed and coupled several open-source software packages (e.g., YADE, MercuryDPM, and LB3D) for solving various geomechanics/geophysics problems, including soil-geomembrane interaction (FEM-DEM) and wave propagation in dry/saturated granular media (LBM-DEM). Dr. Cheng’s main research now focuses on the development of an intelligent, unified material informatics framework for Bayesian uncertainty quantification, validation, selection and unsupervised learning of granular mechanics at various scales.