Forecasting the strength characteristics of concrete incorporating waste foundry sand using advance machine algorithms including deep learning

The incorporation of waste foundry sand (WFS) into concrete has been recognized as a sustainable approach to improve the strength properties of waste foundry sand concrete (WFSC). However, machine learning (ML) techniques are still necessary to forecast the characteristics of WFSC and evaluate the dominant input features for the suitable mix design. For this purpose, the present work selected five ML-based techniques based on gene expression programming (GEP), deep neural network (DNN), and optimizable Gaussian process regressor (OGPR) to predict the mechanical characteristics of WFSC. To build up the predictive models, a database containing 397 values of compressive strength (CS) and 169 values of flexural strength (FS) is collected from published literature. The models' performance was evaluated via various statistical metrics and additionally, external validation criteria were employed to validate the developed models. Furthermore, the Shapley additive explanation (SHAP) was carried out to interpret the model's prediction. The DNN2 model exhibited superior performance, with R-values of 0.996 (training), 0.999 (testing), and 0.997 (validation) for the compressive strength estimation. In contrast, the GEP2 model showed poor accuracy in estimating the CS compared to other developed models, with R-values of 0.851, 0.901, and 0.844 for the training, testing, and validation sets, respectively. Similarly, for the flexural strength estimation, the DNN2 model provided R-values of 0.999 for training, 0.996 for testing, and 0.999 for validation sets, indicating its robust performance. The SHAP analysis revealed that the age, water-cement ratio, and coarse aggregate-to-cement ratio have the prime influence in determining flexural and compressive strength, respectively. The comparison of the models provided that the DNN2 model accurately estimated the output with high accuracy and lower error values and might be utilized in practical fields to reduce labor and cost by optimizing the mix combinations. Finally, for future studies, it is recommended to utilize ensemble and hybrid algorithms, as well as post-hoc explanatory techniques, to forecast the characteristics of WFSC accurately.