Machine learning models for estimating the compressive strength of rubberized concrete subjected to elevated temperature: Optimization and hyper-tuning

Abstract

The incorporation of rubber fibers (RFs) brings about significant divergence in the characteristics of rubberized concrete when contrasted with traditional varieties. Thus, raising concerns about performance under elevated temperature and prolonged exposure. This study effectively addresses the challenges of incorporating rubber fibers in concrete by using artificial neural network (ANN), gene expression programming (GEP), and bagging to examine the impact of input factors such as water-to-cement ratio (W/C), rubber fiber content (RF), elevated temperature (T), and exposure duration (t) on air-cooled compressive strength (CSA). The comprehensive literature review and advanced modeling techniques reveal that ANN excels in capturing complex relationships. In addition, GEP provides clear and accurate models through its unique approach, and Bagging enhances model stability and accuracy. These methods together offer a robust framework for estimating the CSA of rubberized concrete. Thus, contributing valuable insights for optimizing its use in construction. All three models exhibited strong performance, with the ANN emerging as the most effective choice among the evaluated models. Notably, ANN displayed the highest coefficient of determination (R2) value of 0.984, indicating its superior predictive accuracy compared to both GEP (0.982), and bagging (0.970). Moreover, ANN demonstrated the lowest mean absolute error (MAE) score of 0.621 and root mean square error (RMSE) of 0.867, underscoring its precision in forecasting the CSA of rubberized concrete with minimal deviation from experimental values. In addition, the SHapley Additive exPlaination (SHAP) method is employed to comprehend the model estimations. The ICE and PDP plots demonstrate an initial increase in CSA up to 150°C, followed by a significant decrease as temperature rises. Furthermore, CSA decreases with higher RF contents, and linearly declines with increasing W/C ratio. The SHAP analysis provides clear evidence of the strong negative correlation between T and CSA, along with a negative association with RF. A graphical user interface has been developed to estimate the CSA of rubberized concrete, enabling efficient and user-friendly model interaction without the need for physical experimentation.