Explainable hybrid machine learning models for predicting crack width and stress increase in prestressed concrete beams under challenging conditions with GUI integration

Abstract

Prestressed concrete beams are widely used in modern infrastructure; however, accurate prediction of maximum crack width and the increase in prestressing steel stress (Δfps) remains challenging because of the nonlinear interactions among geometric, material, and prestressing parameters. This study develops five hybrid machine-learning models: MOGB-RC, CatBoost-GA, LightGBM-HHO, ANN-DE, and ELM-WOA, to predict these two responses. An initial database of 803 data records, compiled from published experimental studies, was subjected to ensemble-based anomaly detection, yielding 624 samples for model development and evaluation. The proposed models achieved high predictive accuracy and robustness, with MOGB-RC showing the best overall performance. Sensitivity analysis indicated that the prestressing steel area and effective depth were the most influential parameters, followed by the effective prestressing force and the section height. The low individual correlation coefficients further indicate that crack-width development is governed predominantly by nonlinear multivariable interactions rather than isolated parameter effects. A graphical user interface was additionally developed to facilitate rapid implementation of the trained models without requiring advanced programming expertise. The proposed framework provides an accurate, interpretable, and practical data-driven approach for evaluating cracking and prestressing response in prestressed concrete beams.