Natural Hazard Probabilistic Risk Assessment Through Surrogate Modeling

Abstract

Assessment of risk under natural hazards is associated with a significant computational burden when comprehensive numerical models and simulation-based methodologies are involved. Despite recent advances in computer and computational science that have contributed in reducing this burden and have undoubtedly increased the popularity of simulation-based frameworks for quantifying/estimating risk in such settings, in many instances, such as for real-time risk estimation, this burden is still considered as prohibitive. This chapter discusses the use of kriging surrogate modeling for addressing this challenge. Kriging establishes a computationally inexpensive input/output relationship based on a database of observations obtained through the initial (expensive) simulation model. The upfront cost for obtaining this database is of course high, but once the surrogate model is established, all future evaluations require small computational effort. For illustration, two different applications are considered, involving two different hazards: seismic risk assessment utilizing stochastic ground motion modeling and real-time hurricane risk estimation. Various implementation issues are discussed, such as (a) advantages of kriging over other surrogate models, (b) approaches for obtaining high efficiency when the output under consideration is high dimensional through integration of principal component analysis, and (c) the incorporation of the prediction error associated with the metamodel into the risk assessment.

Reduced‐order models for the seismic assessment of plan‐irregular low‐rise frame buildings

“…On the other end, surrogate models (e.g. Tsompanakis et al, 25 Taflanidis et al 26 ) are built exclusively on the premise of calibration, essentially fitting the results of the full model by means of meta-models that simulate as closely as possible the building response while being computationally cheaper. Such non-mechanical models establish a functional relation between the model parameters (input) and the model performance (output), for example by employing neural networks (Lagaros and Fragiadakis 27 ), response surface models (Gavin and Yau, 28 Taflanidis and Cheung 29 ) or kriging models (Gidaris at al., 30 Gidaris and Taflanidis 31 ).…”