Prediction of fatigue life for butt-welded joints using multi-fidelity surrogate modelling

Abstract

Welding is considered as one of the most efficient and reliable joining technologies in fabrications of metallic components for aerospace, maritime, and civil engineering applications. However, fatigue associated failures are inevitable in welded joints due to several aspects, such as local high-stress concentrations, high-residual stresses as well as material and geometry imperfections. Fatigue strength assessments are often performed experimentally or numerically. However, parameter uncertainties regarding geometry, experimental conditions and inadequate consideration of imperfections can lead to inaccurate evaluations. Recently, machine learning (ML) models have been developed for fatigue assessments in terms of computational efficiency for various engineering tasks. Despite the benefits brought by the ML based fatigue assessments, it is still challenging in prediction models as it requires large databases of experimental data for training and validation of models for accurate predictions. In this study, a multi-fidelity (MF) surrogate model which can predict the fatigue life of butt-welded joints is developed and validated. The MF model takes advantage of high-fidelity models which were developed from the 3D scan data of specimens, and simplified low-fidelity models for which less computational resources are needed for data generation. Additive-scaling function concept is employed for MF modelling, and surrogate and discrepancy models are built using Kernal Polynomial Least Square Kriging and eXtreme Gradient Boosting algorithms. The proposed MF model can provide predictions while keeping the balance between accuracy and computational efficiency with a small amount of sample points.