Surrogate Modeling for Stochastic Assessment of Engineering Structures

In many engineering problems, the response function such as the strain or stress field of the structure, its load-bearing capacity, deflection, etc., comes from a finite element method discretization and is therefore very expensive to evaluate. For this reason, methods that replace the original computationally expensive (high-fidelity) model with a simpler (low-fidelity) model that is fast to evaluate are desirable. This paper is focused on the comparison of two surrogate modeling techniques and their potential for stochastic analysis of engineering structures; polynomial chaos expansion and artificial neural network are compared in two typical engineering applications. The first example represents a typical engineering problem with a known analytical solution, themaximum deflection of a fixed beam loaded with a single force. The second example represents a real-world implicitly defined and computationally demanding engineering problem, an existing bridge made of post-tensioned concrete girders.