Efficient techniques for bayesian inverse modeling of large-order computational models

Abstract

Bayesian tools for inverse modeling are based on asymptotic approximations and Stochastic Simulation Algorithms (SSA). Such tools require a number of moderate to large number of system re-analyses. For large-order numerical models of engineering systems, the computational requirements in Bayesian tools can be excessive. Using the Transitional MCMC algorithm, this study proposes efficient techniques for reducing the computational demands to manageable levels. Adaptive surrogate models are used to reduce the number of full system runs by an order of magnitude and parallel computing algorithms are employed to efficiently distribute the Transitional MCMC computations in multi-core CPUs. Applications in structural dynamics are emphasized in this work. Recently developed fast and accurate component mode synthesis techniques, consistent with the finite element parameterization, are implemented to achieve drastic reductions in the system order, resulting in additional substantial computational savings.An example of a bridge model with hundred of thousand of degrees of freedom demonstrates the capabilities of the proposed framework and the remarkable computational savings that can be achieved. © 2013 Taylor & Francis Group, London.