The demand for steel fiber-reinforced concrete (SFRC) in construction has surged, particularly due to its enhanced fire resistance, leading to extensive research on its residual properties after elevated temperature exposure. However, conducting experimental tests can be a time-consuming process, demanding significant resources in terms of time, cost, and human resources. In contrast, machine learning (ML) models can rapidly simulate outcomes, enabling researchers to explore a wide range of scenarios efficiently. The previous literature studies used either ensemble or individual models; however, this study made a unique approach to utilize hybrid models, which are more precise compared to the other types of ML models. Accordingly, a data-rich framework containing 304 data samples was utilized in this study to develop an efficient and robust predictive model for the compressive strength (CS) of SFRC at higher temperatures. Support vector regression (SVR) in combination with three distinct optimization algorithms, specifically, particle swarm optimization (PSO), the firefly algorithm (FFA), and grey wolf optimization (GWO) were utilized to develop the hybrid models. In addition, typical ML techniques such as random forest (RF) and decision tree (DT) were used for comparative analysis. Among the five models, the SVR-FFA hybrid model showcased superior predictive accuracy, with the SVR-GWO model following closely. The correlation coefficients (R) for both models exceeded 0.99. The SVR-FFA model demonstrated relative root mean square error (RMSE) and mean absolute error (MAE) values of 0.0375 and 0.0248, respectively, while the SVR-GWO model exhibited corresponding values of 6.0560 and 5.1326. The SHapley Additive exPlanation (SHAP) technique revealed that the influence of temperature is more significant compared to the heating rate. Moreover, a considerable enhancement was observed in CS as the steel fiber volume fraction (Vf) reached 1.5%; on the other hand, no further improvement was noticed when Vf exceeded the 1.5% threshold. The proposed hybrid models are robust and accurate methods to estimate the CS of SFRC in field applications.
Read full abstract