Evaluating the efficiency of artificial neural networks and tree-based techniques for forecasting the flexural strength of concrete using waste foundry sand

Abstract

The numerous investigations of soft computing algorithms have been done to forecast the flexural strength (FS) of concrete using waste foundry sand (WFS). This research aims to study the application of soft-computing techniques, including random forest (RF), Reduced error pruning tree (REP tree), artificial neural network (ANN), Random tree (RT) and M5P-based model, in forecasting the FS of concrete. For this aim, a dataset of 158 experimental results with a wide range of FS values, ranging from FS 2.21 MPa to 6.98 MPa, was collected from existing literature. The input parameters for the soft computing models included the curing period (CP), slump (SL), fine aggregates (FA), coarse aggregates (CA), water/cement ratio (W/C), cement (C), waste foundry sand (WFS) content, sand (S), WFS blended with other substances (WFS/other), and water (W), with the FS of concrete as the output parameter. Various performance indices were employed to assess the reliability and accuracy of each model, including mean absolute error (MAE), relative root mean square error (RRMSE), coefficient of correlation (R), Nash–Sutcliffe model efficiency coefficient (NSE), root-mean-squared error (RMSE), Wilmott index (WI), and relative absolute error (RAE). Results from the RF model showed high accuracy with R values of 0.9936 and 0.9789, MAE values of 0.1186 and 0.2348, WI values of 0.996 and 0.987, RMSE values of 0.1538 and 0.2931, RAE values of 11.33% and 20.54%, NSE values of 0.986 and 0.953, and RRMSE values of 12.04% and 21.59% during the training as well as testing stages, respectively. The REP Tree model also displayed competitive predictive capability compared to ANN, RT, and M5P models. A sensitivity analysis revealed that the curing period (CP) was the most influential parameter in forecasting FS using the RF-based model. The research emphasises the efficiency of soft computing methods, specifically the random forest, in perfectly assessing the FS of concrete through the utilisation of waste foundry sand. Moreover, it provides researchers with a faster and more economical method to assess the impact of waste foundry sand and additional variables on FS estimation, hence avoiding the necessity for laborious and expensive experimental investigations.