Hybrid support vector regression – Particle swarm optimization for prediction of compressive strength and RCPT of concretes containing metakaolin

This paper develops a hybrid support vector regression (SVR) – particle swarm optimization (PSO) model to predict the compressive strength (CS) and rapid chloride penetration test (RCPT) results of concretes containing metakaolin. The predictive accuracy of SVR models is highly dependent on their learning parameters. Therefore, PSO is exploited to seek the optimal hyper-parameters for SVR in order to improve its generalization capability. Moreover, a SVR-based sequential forward feature selection algorithm is proposed to disclose the most dominant input variables for the prediction of CS and RCPT results. The performance of the hybrid model is compared with the well-known system modeling method of adaptive neural-fuzzy inference system (ANFIS) by using 100 data samples with 25 different mix proportions established by experiments. The results show that the hybrid model has strong potential to predict material properties with high degree of accuracy and robustness.

► A hybrid SVR–PSO model is proposed to predict the properties of concretes. ► Data samples of concrete containing metakaolin were gathered by experiments. ► Dominant input variables are disclosed by a SVR-based feature selection algorithm. ► The SVR–PSO model is utilized to predict the CS and RCPT test results. ► The hybrid model outperforms the ANFIS model in generalization capability.