PSI - Issue 75

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

109

8

(which includes the hyperparameter tuning and LOOCV process) and the remaining 15% for a final, independent test set. The randomized search CV, as described, was performed on the 85% training set, both with and without the application of Recursive Feature Elimination (RFE), while keeping the hyperparameter search space consistent.

Fig. 12: Results of the recursive feature elimination.

Fig. 11: Notch profile feature extraction

4.3. Model evaluation The coefficient of determination 2 and the mean absolute error (MAE) are selected as performance metrics for training and tuning the models. The results of the performance of the models trained on the datasets with and without RFE are shown in Fig. 13 and Fig. 14. The voting regressor is an ensemble machine learning model that combines the predictions of the four best-performing base models. By averaging their outputs it enhances accuracy and robustness compared to individual models. When applied to the dataset without RFE, the voting regressor outperforms the other models. However, when RFE is used, the gradient boosting regressor achieves slightly better overall performance. In general, all models demonstrate high accuracy, with the notable exception of the K-nearest neighbours algorithm, which performs relatively poorly. The case study on specimen N35, described in Section 3.5 is revisited to determine whether the ML model is capable of predicting the exceptionally high value of . Fig. 15 shows a scatter plot of the predicted values on a test set containing the N35 specimen, with the N35 specimen highlighted in red. Notably, the ML models are capable of identifying and predicting the higher value of though underestimating it. The accuracy is slightly worse than for the other predictions given that the number of outliers in the dataset is small.

Fig. 13: Performance metrics from training and testing on the dataset with RFE.

Fig. 14: Performance metrics from training and testing on the dataset without RFE.