Probabilistic Models to Approximate Highly Repetitive Linear and Nonlinear Finite Element Analyses of Nuclear Components

In the development and technical support of nuclear plants, Engineers have to deal with highly repetitive finite element analyses that involve modeling of local variations of the initial design, local flaws due to corrosion-erosion effects, material properties degradation, and modifications of the loading conditions. This paper presents the development of generic models that emulate the behavior of a complex finite element model in a simplified form, with the statistical representation based on a sampling of base-model data for a variety of test cases. An improved Latin Hypercube algorithm is employed to generate the sampling points based on the number and the range of the variables that are considered in the design space. Four filling methods of the approximation models are discussed in this study: response surface, nonlinear, neural networks, and piecewise polynomial model. Furthermore, a bootstrapping procedure is employed to improve the confidence intervals of the original coefficients, and the single-factor or double-factor analysis of variance is applied to determine whether a significant influence exists between the investigated factors. Two numerical examples highlight the accuracy and efficiency of the methods. The first example is the linear elastic analysis of a pipe bend under pressure loading. The objective of the probabilistic assessment is to determine the relation between the loading conditions as well as the geometrical aspects of this elbow (pipe wall thickness, outside diameter, elbow radius, and maximum ovality tolerance) and the maximum stress in the elbow. The second example is an axisymmetric nozzle under primary and secondary cycling loads. Variations of the geometrical dimensions, nonlinear material properties, and cycling loading are taken as the input parameters, whereas the response variable is defined in terms of Melan’s theorem translated into the Nonlinear Superposition Method.