GrainLearning: an efficient Bayesian uncertainty quantification framework for discrete element simulations of granular materials

Cheng, H. (Speaker), Takayuki Shuku (Contributor), Weinhart, T. (Contributor), Luding, S. (Contributor)

Activity: Talk or presentationOral presentation

Description

The nonlinear, history-dependent macroscopic behavior of a granular material is rooted in the micromechanics between constituent particles and the irreversible change in the microstructure. The discrete element method (DEM) can predict the evolution of the microstructure resulting from interparticle interactions. However, micromechanical parameters at contact and particle levels are generally unknown because of the diversity of granular materials with respect to their surfaces, shapes, disorder and anisotropy.

GrainLearning, a Bayesian filtering framework specially developed for DEM simulations of granular materials can iteratively explore parameter/solution space conditioned on experimental data, and efficiently quantify and propagate posterior uncertainties. It utilizes Bayesian nonparametric density estimation to iteratively refine the proposal density, thus enabling an efficient multi-level sampling for global optima.

As an example, the probability distribution of the micromechanical parameters, conditioned on the experimental measurements of granular flow, is approximated, with rapid convergence within a few iterations. Six micromechanical parameters, i.e., contact-level Young's modulus, restitution coefficients, interparticle friction, rolling stiffness and rolling friction, are chosen as relevant for the macroscopic behavior. The a posteriori expectation of each micromechanical parameter converges within a few iterations, leading to an excellent agreement between the experimental data and the numerical predictions. As new result, the proposed framework provides a deeper understanding of the correlations among micromechanical parameters and between the micro- and macro-parameters/quantities of interest, including their uncertainties. Therefore, the iterative Bayesian filtering framework has a great potential for quantifying parameter uncertainties and their propagation across various scales in granular materials.
Period21 Jul 2019 - 26 Jul 2019
Event title8th International Conference on Discrete Element Methods, DEM 2019: MS-03: Open-source development
Event typeConference
Conference number8
LocationEnschede, Netherlands
Degree of RecognitionInternational