Residual mechanical properties of steel-fibre-reinforced concrete with volcanic scoria sand after freeze–thaw cycles using machine learning

In cold regions, using steel fibres (SF) and manufactured volcanic scoria sand (MVSS) for fibre-reinforced concrete formulation can mitigate the problem of deterioration of mechanical properties due to freeze–thaw (F–T) cycles and effectively reduces the consumption of natural sand. To better predict the residual mechanical properties of concrete after freeze–thaw (F–T) cycles, this study develops four machine learning models: back-propagation neural network, convolutional neural network, decision tree and CatBoost. The input variables were water/cement ratio, manufactured volcanic scoria sand replacement rate, steel fibre volume content and F–T cycles. The output values were compressive strength (CS) and splitting tensile strength (STS). Key indicators showed all models exhibited acceptable accuracy. CatBoost outperformed the other methods with root mean squared error of 0.391 and 0.037, mean absolute error of 0.273 and 0.026, mean absolute percentage error of 0.009 and 0.011, scatter index of 0.011 and 0.014, and index of agreement of 0.999 and 0.999 for CS and STS, respectively. The coefficients of determination (R2) are all as high as 0.99. CatBoost shows the highest prediction accuracy. Sensitivity testing of the strength of steel-fibre-reinforced manufactured volcanic scoria sand concrete using CatBoost and it showed F–T cycles were an essential parameter. Finally, scanning electron microscopy shows SF and MVSS can improve the frost resistance of concrete.