An iterative sequential Monte Carlo filter for Bayesian calibration of DEM models

Abstract

The nonlinear history-dependent macroscopic behavior of granular materials is rooted in the micromechanics at contacts and irreversible rearrangements of the microstructure. This paper presents an iterative sequential Monte Carlo filter to infer micromechanical parameters for DEM modeling of granular materials from macroscopic measurements. To demonstrate the performance of the new Bayesian filter, the stress–strain behavior of fine glass beads under oedometric compression is considered. The parameter sets are initially sampled uniformly in parameter space and then resampled around highly probable subspaces, which shrink towards optimal solutions iteratively. The proposed calibration approach is fast, efficient and automated, because it uses the posterior distribution after a completed iteration as the proposal distribution for the succeeding iteration, and thereby allocating computational power to more probable simulation runs. The Bayesian filter can also serve as a powerful tool for uncertainty quantification and propagation across various scales in multiscale simulation of granular materials.