Surrogate modeling for structural response prediction of a building class

Abstract

Buildings are vital for critical community functions, and it is of great importance to efficiently invest the limited societal resources in the design of the buildings. To achieve this goal, careful assessment of risk from future hazards is required. In practice, the risk to building structures is regulated by structural design codes through target reliability levels, which are reflected in many code factors, including partial safety factors, load combination factors, and modification factors. Optimizing the target reliability levels often requires running a large number of nonlinear dynamic analyses of complex finite element models, which imposes a significant burden on the computational resources. Computation time can be reduced considerably by employing surrogate models that can efficiently approximate the relation between input design and hazard intensity variables with building response output. This paper explores the different surrogate models, namely, support vector machines, kriging, and neural networks, for structural response prediction of a building class so that they can be used in the target reliability index optimization of a building class (a group of buildings with the same load-carrying characteristics). The uniqueness of this study is in developing a single surrogate model for a group of buildings instead of a single building design. The investigation on the performance of the surrogate models in structural response prediction is conducted for mid-rise office building class by comparing the computation time, accuracy, and robustness.