A New Database for the Durability of Fiber-Reinforced Polymer Rebar

Abstract Fiber-reinforced polymer (FRP) rebar has gained increasing attention due to its superior corrosion resistance compared with steel rebar. However, a major challenge restricting its wider adoption is the limited understanding of its long-term durability. Given the complex, multifactorial nature of durability, machine learning (ML) is considered as the only feasible approach for accurate prediction. However, the lack of a suitable database has hindered ML-based modeling efforts. This paper presents the development of a comprehensive FRP rebar durability database, comprising 5,210 instances and 80 parameters—the largest of its kind so far. The construction process and database composition are detailed, along with an analysis of key challenges affecting tabulated experimental databases, namely, the missing data issue, missing parameter issue, and small data issue. The superiority of the new database in addressing these challenges is demonstrated. An ablation test was conducted as a preparatory step for generating the complete-case dataset required for subsequent analyses, with results supporting the use of listwise deletion over mean/mode imputation. Permutation importance analysis then reveals that several parameters frequently were overlooked in existing databases. Finally, a learning curve analysis shows that increasing data size improves prediction accuracy, with the most significant reduction observed in multilayer perceptron (MLP): both root-mean-squared error (RMSE) and mean absolute error (MAE) drop by approximately half when the training data increase from 10% to 100%. All analyses demonstrate that the new database establishes a robust foundation for advancing ML-driven predictive modeling of FRP rebar durability, supporting future research and practical applications as a broadly applicable reference across diverse environmental conditions and FRP types.