PSI - Issue 75

Kris Hectors et al. / Procedia Structural Integrity 75 (2025) 102–110 Hectors et al. / Structural Integrity Procedia (2025)

107

6

Fig. 7: Results of the convergence study showing the maximum notch stress as a function of the number of elements in the notch (i.e. number of seeds).

Fig. 8: Scatter plot of the elastic stress concentration factors obtained from the finite element analyses of the models based on the as-manufactured specimens.

Fig. 9: Illustration of the relationship between the opening angle, notch radius, notch height and the stress concentration factor. The unfilled markers correspond to the specimens from which no valid FEA results could be obtained.

Fig. 10: Notch root profile of N35, showing an irregularity at the notch root not captured by the global measurement of notch opening angle.

4. Machine learning Various machine learning models are evaluated for estimating stress concentration factors based on geometry data extracted from profile scans. This work focusses exclusively on conventional machine learning approaches rather than deep learning techniques such as artificial neural networks (ANN). The machine learning models used in this work were chosen from established Python libraries, such as PyTorch and scikit-learn, due to their robust implementation and open-source availability. Linear regression, ridge regression, and linear support vector machines were chosen to evaluate the effectiveness of regularization in a feature space where multicollinearity is likely. PLS was chosen for its ability in handling such multicollinearity and suitability in dealing with small datasets. K-nearest neighbours, random forest and gradient boosting regression were chosen for their ability to capture potentially complex, non-linear relationships between the as-machined notch geometry and the resulting stress concentration. Finally, a voting regression model was used to determine if combining multiple models could produce a more accurate prediction by offsetting the weaknesses of the individual models. From the work described in section 3, 151 valid datapoints could be obtained for model training and testing. The training data incorporates -values derived from finite element analyses described previously, enabling the application of supervised learning algorithms for regression tasks. The models assessed in this work include linear