Application of an HMC Based Approximate Method for Combined Identification of Hydraulic Conductivity and Piping Region Interface

Abstract

The detection of piping and the extent to which it progresses is important in the serviceability of natural dams and levees, where the soil is usually poorly consolidated and filters or drainage zones are absent. Practically, information of the spatial variation of hydraulic conductivity is also unavailable in the domain of interest. An approximate version of a Hamiltonian Monte Carlo (HMC) based method for combined probabilistic inversion is detailed. The inverse problem consists of Karhunen-Loève (KL) expansion parameters and solid-void interface parameters which are determined simultaneously. The interface parameter updates are carried out using a reversible proposal in a mesh moving framework. To maintain computational efficiency, the parameter update is enforced in an approximate sense, where the covariance matrix for the KL expansion is kept constant throughout the analysis. Synthetic data from a numerical experiment on a domain containing a predefined piping region, is used to validate the approximate method. A total of 30000 samples are generated using the HMC sampler. Results show that the Markov chains converge to the stationary distribution. A good match is also observed between the inferred mean interface, and the true interface and the true spatial distribution of hydraulic conductivity is obtained. Show more