Predicting Time-Dependent Durability of FRP–Timber Bonds in Harsh Environments: A Comparative Study of Machine Learning Models
This study presents a comparative evaluation of three machine learning models, XGBoost, AdaBoost, and LightGBM, for predicting the time-dependent bond strength between fiber-reinforced polymer (FRP) and timber in both normal and harsh environments. A dataset was compiled (79 for normal conditions and 265 for harsh environments) incorporating material properties, geometric parameters, exposure time, and solution pH as input features. Hyperparameter optimization was performed for each model, and performance was evaluated using R2, RMSE, MAE, and MSE metrics. SHAP analysis and Partial Dependence Plots were employed to interpret feature importance and model behavior. Under normal conditions, XGBoost achieved the highest predictive accuracy (testing R2 = 0.944, RMSE = 2.170, and MAE = 1.775), outperforming AdaBoost and LightGBM. However, in harsh environments, LightGBM demonstrated superior generalization, with the highest testing R2 of 0.797 and the lowest RMSE of 0.549 and MAE of 0.431, outperforming XGBoost and AdaBoost. AdaBoost exhibited severe overfitting under harsh conditions, with a training-to-testing R2 drop of 0.341. Feature importance analysis by SHAP analysis identified fiber tensile strength and exposure time as the most influential parameters governing bond performance. SHAP force plots demonstrated that fiber properties predominantly enhance bond strength, while pH consistently acts as a decreasing factor under harsh conditions. This research provides a robust predictive framework for FRP–timber bond durability, offering valuable insights for structural design and service life prediction in harsh environments.