Surrogate Modelling for the Estimation of Stress Intensity Factor
Proceedings Publication Date
Presenter
Dr. Smitha Koduru
Presenter
Author
Durlabh Bartaula, Ngandu Balekelayi, Smitha Koduru, Yong Li, Samer Adeeb
Part of the proceedings of
Abstract

Prediction of fatigue crack growth (FCG) in oil and gas pipelines is a challenging topic due to the sensitivity of the crack growth to the material parameters, crack size, and variations in the cyclic stress. Furthermore, uncertainty in load magnitudes and the material variability requires a large volume of data to improve reliability of the fatigue crack growth predictions even when the crack size is reliably measured. As full-scale experimental tests are expensive and time consuming, experimental data is limited. Numerical solution techniques, such as eXtended Finite Element Method (XFEM), are used to study the FCG under variable material properties, crack sizes and loading conditions.

Numerical techniques, while being less expensive than full-scale tests, are yet computationally expensive for rapid application in the field. Therefore, simplified models that can rapidly provide the remaining fatigue with reasonable confidence are needed. For a simplified approach, well-known fatigue growth models such as Paris Law, and Dowing-Begley Law are commonly used. However, these models require an accurate estimate of the crack driving parameter known as “stress intensity factor” (SIF), which is a measure of stress intensity at the crack tip due to remote applied load as well as residual stresses. 

In the present study, SIF is estimated using XFEM results for various combinations of the crack size and pipe geometry and load intensity. Three approaches to develop surrogate models to predict SIF were selected: namely (1) Gaussian Process Regression (GPR), (2) Neural Networks (NN), and (3) Support Vector Machines (SVM). The model parameters for both the approaches are optimized using a training data set and the models are validated against a test data set. The performance of the three approaches is compared based on the requirements of data size, data quality (i.e., presence of noise in the training data), and accuracy of the predictions.

To view the video or download the paper please register here for free

You already have access? Sign in now.