TY - GEN
T1 - Bayesian Uncertainty Quantification for Geomechanical Models at Micro and Macro Scales
AU - Cheng, Hongyang
AU - Magnanimo, Vanessa
AU - Shuku, Takayuki
AU - Luding, Stefan
AU - Weinhart, Thomas
N1 - Publisher Copyright:
© 2021, The Author(s), under exclusive license to Springer Nature Switzerland AG.
PY - 2021
Y1 - 2021
N2 - Uncertainty exists in geomaterials at contact, microstructural, and continuum scales. To develop predictive, robust multi-scale models for geotechnical problems, the new challenge is to allow for the propagation of model/parameter uncertainty (conditioned on laboratory/field measurements) between micro and macro scales. We aim to first quantify these uncertainties using an iterative Bayesian filtering framework. The framework utilizes the recursive Bayes’ rule to quantify the evolution of parameter uncertainties over time, and the nonparametric Gaussian mixture model to iteratively resample parameter space. Using the iteratively trained mixture to guide resampling, model evaluations are allocated asymptotically close to posterior modes, thus greatly reducing the computation cost. In this paper, we first respectively quantify the parameter uncertainty of models that are discrete and continuum in nature, namely a discrete particle and an elasto-plastic model. We then link the two models by conditioning their uncertainties on the same stress-strain response, thereby revealing micro-macro parameter correlations and their uncertainties. The micro-macro correlations obtained can be either general for any granular materials that share similar polydispersity or conditioned on the laboratory data of specific ones.
AB - Uncertainty exists in geomaterials at contact, microstructural, and continuum scales. To develop predictive, robust multi-scale models for geotechnical problems, the new challenge is to allow for the propagation of model/parameter uncertainty (conditioned on laboratory/field measurements) between micro and macro scales. We aim to first quantify these uncertainties using an iterative Bayesian filtering framework. The framework utilizes the recursive Bayes’ rule to quantify the evolution of parameter uncertainties over time, and the nonparametric Gaussian mixture model to iteratively resample parameter space. Using the iteratively trained mixture to guide resampling, model evaluations are allocated asymptotically close to posterior modes, thus greatly reducing the computation cost. In this paper, we first respectively quantify the parameter uncertainty of models that are discrete and continuum in nature, namely a discrete particle and an elasto-plastic model. We then link the two models by conditioning their uncertainties on the same stress-strain response, thereby revealing micro-macro parameter correlations and their uncertainties. The micro-macro correlations obtained can be either general for any granular materials that share similar polydispersity or conditioned on the laboratory data of specific ones.
KW - Bayesian filtering
KW - Constitutive modeling
KW - Discrete element modeling
KW - Micro-macro parameter correlations
KW - Uncertainty quantification
UR - http://www.scopus.com/inward/record.url?scp=85101581078&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85101581078&partnerID=8YFLogxK
U2 - 10.1007/978-3-030-64514-4_90
DO - 10.1007/978-3-030-64514-4_90
M3 - Conference contribution
AN - SCOPUS:85101581078
SN - 9783030645137
T3 - Lecture Notes in Civil Engineering
SP - 837
EP - 845
BT - Challenges and Innovations in Geomechanics - Proceedings of the 16th International Conference of IACMAG - Volume 1
A2 - Barla, Marco
A2 - Di Donna, Alice
A2 - Sterpi, Donatella
PB - Springer Science and Business Media Deutschland GmbH
T2 - 16th International Conference of the International Association for Computer Methods and Advances in Geomechanics, IACMAG 2021
Y2 - 5 May 2021 through 8 May 2021
ER -