Performance evaluation of slag-based concrete at elevated temperatures by a novel machine learning approach

Ground granulated blast furnace slag is a sustainable material and supplementary for cement in the concrete industry. Different behavioral aspects must be assessed to achieve reliable sustainable materials, including post -fire mechanical properties. One robust tool is the machine learning approach to train prediction models. This study proposes a novel machine learning algorithm, hybrid support vector regression and dolphin echolocation algorithm (SVR-DE), to predict the post-fire compressive strength ratio of slag-based concrete. In this regard, SVR hyper-parameters were tuned by the DE optimization algorithm. Four kernel functions were implemented in SVR formulation: linear, sigmoid, polynomial, and RBF. Accordingly, four different models were proposed based on the defined kernels with comparing their accuracy. The training and testing process was applied to 80% and 20% of 124 collected data from relevant studies in the literature, respectively. Although the models with sigmoid and RBF kernel functions yielded a satisfactory fit, both with 0.86 R2 value, the model with polynomial kernel function expressed high accurate results with R2 and RMSE of 0.92 and 0.59. A sensitivity analysis was per-formed eventually to investigate the importance of input parameters. Six parameters displayed a high importance class in the post-fire strength ratio of slag-based concrete, including fine/coarse aggregate, water/cement, slag/ water, slag/superplasticizer, slag/cement, and temperature.

Automated summary

This study proposes a novel machine learning algorithm, SVR-DE, to predict the post-fire compressive strength ratio of slag-based concrete. Among the four different models based on different kernel functions, the model with polynomial kernel function showed the highest accuracy. The sensitivity analysis revealed the importance of six parameters in the post-fire strength ratio of slag-based concrete.