Dr. Hongyang Cheng
Postdoctoral researcher at the University of Twente, the Netherlands
Postdoctoral researcher at the University of Twente, the Netherlands
During my PhD, I initiated my own independent research on Bayesian calibration/uncertainty quantification for DEM simulations, with the help of Dr Takayuki Shuku at Okayama University and Dr. Klaus Thoeni at the University of Newcastle. We have done the provisional work to demonstrate the feasibility in J1 and developed the Bayesian framework further by utilizing the posterior probability distribution, approximated by Dirichlet process Gaussian mixtures model, to adaptively sample parameter/solution space [J4].
We introduced the concept of Bayesian calibration to geotechnics [5b.4] and powder technology (Cheng et al., 2018d). The Bayesian calibration tool [5b.4] fills the longstanding gap between macroexperiments and microsimulations. Using advanced machine learning, I further boosted the efficiency of the Bayesian framework (Cheng et al., 2018b) and implemented it as a distinguished feature of the opensource DEM code YADE. The method applies to not only micromodels, e.g., microbeads (Cheng et al., 2018b) and sand [5b.4], but also macroscopic constitutive laws. 

The novelty of the iterative Bayesian filter [J4] consists in recursively updating the posterior distribution of model parameters in time and iterating the process with new samples 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. As an example, the posterior distribution of the micromechanical parameters, conditioned on the experimentally measured stressstrain behavior of a granular assembly, is approximated.
The a priori particle configuration is obtained from 3D Xray computed tomography images. The a posteriori expectation of each micromechanical parameter converges within four iterations, leading to excellent agreement between the experimental results and the numerical predictions. With the previously approximated posterior chosen as the new proposal density, the model evaluations are limited to those within highly probable parameter subspaces. As new result, the proposed framework provides a deeper understanding of the correlations between the uncertainty of the micromechanical parameters and the macroscopic quantities of interest, conditioned on the experimental data. Therefore, the iterative Bayesian filtering framework has a great potential for quantifying uncertainties and their propagation across various scales in granular materials. 







Wrapping granular soils in geosynthetic containers, such as soilbags, results in a considerable increase in the bearing capacity due to the effective restraint on the dilatancy of the soil. We numerically investigates the stress states and fabric anisotropies in the wrapped soil using the DEM, providing a novel perspective for new insights into the reinforcement mechanisms and the development of constitutive relations for soilbags.
The two most anticipated loading conditions, namely, unconfined compression and simple shear, are considered, and numerical predictions are compared to experimental results. During unconfined compression, both global and local p–q stress paths evolve linearly, having the same slope until the global failure of the wrapping geosynthetic. Under simple shear, the global stress path approaches the critical state line first and then turns to the compression line of the wrapped soil. Some local loading–unloading stress paths are observed, which may account for the high damping of soilbags during cyclic shear. The reduced fabric anisotropies of the normal and tangential force chains suggest greater confinement from the lateral sides of the geosynthetic container in either loading course. The performance and mechanisms of the soilbag earth reinforcement method, i.e., confinement and interlocking, can be better understood based on these new findings on the stress states and fabric anisotropies. 

The performance of geosyntheticwrapped and layered soil commonly used for constructing retaining structures are numerically investigates using the DEM. The geotextile and the soil that consist of the reinforced soil are calibrated with tensile, shear box, and triaxial test results. The discrete particle models of the two reinforced soils assume geotextile as regularly positioned particles linked with stretching springs.
The discrete models are loaded under triaxial compression in periodic boundary condition, in order to evaluate the influence of reinforcement form on the mechanical behavior of reinforced soil. The simulations show that the wrapped soil is able to sustain much greater stress than the layered one. Although similar global volumetric deformation is found in both cases, the soil wrapped in a container dilates less than that sandwiched between a pair of sheets. 
Elastic wave propagation provides a noninvasive way to probe granular materials. The DEM using particle configuration as input, allows a micromechanical interpretation on the acoustic response of a given granular system. We compare static and dynamic numerical probing methods, from which wave velocities are either deduced from elastic moduli or extracted from the time/frequencydomain signals.
The dependence of wave velocities on key characteristics, i.e., perturbation magnitude and direction for static probing, and maximum travel distance and inserted signals for dynamic probing, is investigated. It is found that processing the frequencydomain signals obtained from dynamic probing leads to reproducible wave velocities at all wavenumbers, irrespective of the perturbation characteristics, whereas the maximum travel distance and input signals for the time domain analysis have to be carefully chosen, so as to obtain the same longwavelength limits as from the frequency domain. Static and dynamic probes are applied to calibrated representative volumes of glass beads, subjected to cyclic oedometric compression. Although the perturbation magnitudes are selected to reveal the elastic moduli, the deduced wave velocities are consistently lower than the longwavelength limits at various stress states, and thus sensitive to sample size. While the static probes investigate the influence of stress history on modulus degradation, dynamic probing offers insights about how dispersion relations evolve during cyclic compression. Interestingly, immediately after each load reversal the incremental behavior is reversibly elastoplastic, until it becomes truly elastic with further unload/reload. With repeating unload/reload, the Pressure or Shearwave dispersion relations become increasingly scalable with respect to their longwavelength limits. 

