Machine-learning prediction of residual tensile strength of GFRP rebars in alkaline environments

Assessing and forecasting the long-term behavior of glass fiber-reinforced polymer (GFRP) rebars continue to pose significant challenges. To address this, the present study introduces a predictive framework aimed at estimating the retained tensile strength of GFRP rebars when exposed to alkaline environments. A total of 350 experimental data points were collected through keyword-based literature retrieval and subsequently augmented using a Gaussian Copula-based generative model. Six machine learning algorithms, combined with grid search and five-fold cross-validation, were employed to develop predictive models for residual tensile strength in alkaline environments. SHAP (SHapley Additive exPlanations) was used to conduct parameter importance and sensitivity analyses. Unlike prior ML-based durability studies that often suffer from data scarcity and limited generalization, this work uniquely integrates a statistical generative approach to rigorously expand small-scale experimental datasets while preserving their underlying physical correlations, thereby significantly enhancing model robustness. The results show that the Gaussian Copula model effectively captures the primary distributional characteristics of the experimental data, and that XGB (eXtreme Gradient Boosting) is the most suitable model for predicting residual tensile strength. Among the input parameters, environmental temperature has the greatest influence on residual tensile strength, while pH exhibits the least impact. Fiber content and rebar diameter positively affect residual tensile strength—higher values lead to greater strength. Conversely, higher pH, elevated temperature, and longer exposure duration negatively influence residual tensile strength. These findings on parameter importance and sensitivity provide valuable insights for durability studies of GFRP rebars.