Machine Learning Prediction of Ultimate Load and Failure Mode of CFRP-Strengthened Steel I-Beams: Source-Aware Validation and Limits of Transferability
Designs · 07.11.2026 · 10 (6) · 125
DOI: 10.3390/designs10060125 ↗Kurz gefasst
Fachartikel in der Zeitschrift Designs, online erschienen am 07.11.2026. Verfasst von Bishal Kumar Jha und 2 weiteren Autor:innen. Die Arbeit nennt 37 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Carbon fibre-reinforced polymer (CFRP) strengthening of steel I-beams can increase flexural capacity, but assessment commonly requires analytical models or finite element analysis (FEA). This paper examines how far machine learning (ML) predictions of ultimate load and failure mode for simply supported steel I-beams with CFRP bonded to the tension flange transfer beyond the studies used for training. A database of 292 records from nine studies (36 experimental, 256 numerical) was compiled with specimen-level provenance for source study, calibration family, failure-label origin and adhesive-property definition. Random-split gradient boosting (GB) reaches R2 = 0.998, but 83% of the records come from two studies. Under leave-one-study-out (LOSO) validation, pooled R2 falls to 0.478 (mean absolute percentage error, MAPE = 36.2%) and is positive for only two of nine held-out studies; holding out calibration families gives R2 = 0.452 (MAPE = 47.7%). Training on numerical records from other calibration families and testing on experiments gives R2 = 0.517 (MAPE = 39.3%). Multiple linear regression matches GB on random splits but fails to generalize to out-of-study data (LOSO R2 = 0.166). When the unstrengthened capacity of the same beam is supplied, a hybrid formulation predicts total load with LOSO MAPE = 4.5%, but its ML increment model is no better than a mean-increment baseline across studies (increment R2 = 0.026 against −0.036). Failure-mode classifiers trained and tested on source-reported labels under LOSO reach 59.4% accuracy, below the 72.2% of a majority-class baseline. Conformalized prediction intervals restore near-nominal 90% coverage under LOSO (87.0%) only with mean widths of about 1.7 times the load. The models are therefore limited to interpolation and screening within the sampled studies, and the results identify the independent experimental evidence needed before design use.
Zitieren
APA
Jha, B. K., El-Sisi, A. E., & Sah, R. K. (2026). Machine Learning Prediction of Ultimate Load and Failure Mode of CFRP-Strengthened Steel I-Beams: Source-Aware Validation and Limits of Transferability. Designs, 10(6), 125. https://doi.org/10.3390/designs10060125
Vancouver
Jha BK, El-Sisi AE, Sah RK. Machine Learning Prediction of Ultimate Load and Failure Mode of CFRP-Strengthened Steel I-Beams: Source-Aware Validation and Limits of Transferability. Designs. 2026;10(6):125. doi:10.3390/designs10060125
BibTeX
@article{Jha2026,
author = {Jha, Bishal Kumar and El-Sisi, Alaa El-Din and Sah, Rohit Kumar},
title = {Machine Learning Prediction of Ultimate Load and Failure Mode of CFRP-Strengthened Steel I-Beams: Source-Aware Validation and Limits of Transferability},
journal = {Designs},
year = {2026},
volume = {10},
number = {6},
pages = {125},
doi = {10.3390/designs10060125}
}
EUVIMED