PSI - Issue 70

N. Vignesh Kumar et al. / Procedia Structural Integrity 70 (2025) 453–460

457

3.4 Validation Approach and Limitations While robust 10-fold cross-validation within Realfonzo & Napoli (2011) database was employed, we acknowledge the limitation of not validating against fully independent external datasets, primarily due to challenges in sourcing compatible data. However, the utilized database is comprehensive, covering a wide range of unconfined concrete strengths (fco from approx. 15 MPa to 170 MPa) relevant to many practical construction applications, from residential to high strength structures. Furthermore, to partially simulate external validation and assess robustness across different experimental conditions within the database, we segregated data by research groups by training and testing. This analysis showed only a modest performance decrease (average R2 drop: 0.07 for CFRP, 0.09 for GFRP) compared to random splits, suggesting reasonable model robustness. Nonetheless, future validation on truly external datasets remains desirable to fully confirm generalizability across diverse real-world conditions. 4 Results and Discussion The performance metrics for all parametric models (Stepwise and LASSO) and non-parametric models (SVM, RTE and GPR), evaluated on the testing and training datapoints are presented in Fig. 1. The results shown separately for CFRP and GFRP confined concrete specimens. Performance metrics (Fig. 1) show GPR consistently outperformed other models (SVM, RTE, LASSO, STEPWISE) for both CFRP and GFRP, achieving the highest R2 values. Other non-parametric models (SVM, RTE) followed, while parametric models lagged. While GPR showed high training R2 (CFRP: 0.986, GFRP: 0.997), overfitting was addressed via cross-validation, inherent GPR regularization, and comparison with simpler models. The small gap between training and testing R2 (CFRP: 0.917 test, GFRP: 0.916 test) suggests adequate generalization.

Fig. 1. Performance metrics of parametric and non-parametric models using R 2

a