Abstract
Reinforced Concrete (RC) structures are often compromised by cracks and corrosion, necessitating effective retrofitting strategies to restore their structural integrity. Ultra-High-Performance Concrete (UHPC) has emerged as a promising material for strengthening damaged RC beams, significantly enhancing their load-bearing capacity and durability. However, accurately predicting the flexural performance of UHPC-strengthened RC beams remains a challenge due to complex material interactions and limited datasets. This addresses this gap by developing a data-driven framework that combines generative data augmentation and explainable machine learning to predict the ultimate moment resistance (Mu) of strengthened beams. A curated dataset of 160 experimental cases was expanded using a Tabular Generative Adversarial Network (TGAN), and six ensemble learning models were trained and evaluated. Among them, Categorical Boosting (CatBoost) demonstrated superior performance with an R² of 0.90 on the test set. SHapley Additive exPlanations (SHAP) were employed to explain model predictions, revealing that the UHPC reinforcement ratio, longitudinal reinforcement ratio of beam, beam width, and concrete compressive strength are the most influential factors. The proposed approach not only improves prediction accuracy and model robustness but also provides interpretable insights to support rational design decisions in structural retrofitting.
•First use of GANs to predict moment improvement in UHPC-strengthened damaged RC beams.•CatBoost achieved high accuracy on both the training set (R2=0.99) and test set (R2=0.90).•The Shapely-based model explanations reveal features importance and sensitivity for engineering design.