Interpretable machine learning for GFRP bar durability prediction from sparse experimental data

Ensuring the long-term durability of glass-fiber-reinforced polymer (GFRP) bars in aggressive environments remains a critical challenge in structural engineering, particularly when experimental data are scarce. Traditional durability prediction methods often rely on simplified empirical equations or machine learning (ML) frameworks that require large datasets, reducing their applicability to small-scale studies. To address this gap, the present study aims to establish a data-efficient and interpretable ML framework for predicting the tensile strength retention (TSR) of GFRP bars under various environmental conditions using minimal experimental input. An experimental dataset of 78 samples was developed, with TSR modeled as a function of immersion time, temperature, pH, and bar diameter. Two predictive approaches were evaluated: an XGBoost model trained on CTGAN-augmented data and a transformer-based TabPFN model applied directly to the original dataset. The model performance was assessed using R², MAE, RMSE, and A20 metrics, while interpretability was achieved through SHAP-based feature attribution. The TabPFN model achieved superior accuracy (R² = 0.954, MAE = 2.71) compared with XGBoost (R² = 0.78, MAE = 5.48), effectively capturing the nonlinear decline in TSR with increasing temperature and immersion time. These findings demonstrate that an accurate and physics-consistent durability prediction of GFRP bars can be achieved using a small experimental dataset without retraining or data augmentation. The proposed framework provides a scalable and interpretable solution for performance-based durability design of GFRP-reinforced structures, which is particularly valuable when long-term experimental testing is impractical.