Prediction and validation of mechanical properties of self-compacting geopolymer concrete using combined machine learning methods a comparative and suitability assessment of the best analysis.

Recent sustainable engineering trends show the re-use of wastes in the production of concrete materials. This was important in two ways. First, there is a great environmental necessity to eliminate these industrial wastes and their usage in a solid waste upcycling system to ensure structural sustainability creates an avenue for this process. Second, it has become important to reduce laboratory and equipment costs by establishing intelligent models through the application of these supplementary cements and optimized for optimal performance of concrete materials. For these reasons, the present research work has applied the intelligent learning abilities of eight (8) ensemble-based and one (1) symbolic regression machine learning methods to predict the strengths (compressive-Fc, flexural-Ff and splitting tensile-Ft) of SCGPC with the "Orange Data Mining" software version 3.36. In this research paper, the influence of the industrial wastes like ground granulated blast furnace slag (GGBS) and fly ash (FA) and alkali activators such as (NaOH and NaSiO) on the performance self-compacting geopolymer concrete (SCGPC) in terms of strength has been studied. This has been executed using 132 mix entries at different curing regimes and partitioned into 75% and 25% for training and validation, respectively. At the end of the process, performance indices were employed to test accuracy and comparatively the best. Also, the Taylor chart-based comparison of the performance of the ensemble-based machine learning models was conducted. The results show that for the compressive strength of the SCGPC (Fc) model, the K-NN outclassed all the ensemble techniques with an average R of 0.99, accuracy of 0.96, and an average error of 0.04%. This is followed in order of superiority by the SVM closing its model with an average R of 0.99, accuracy of 0.955 and average error of 0.045%. Both models ended with equal SSE, MAE, MSE, and RMSE. For the flexural strength of the SCGPC (Ff) model, the K-NN and the SVM performed equally in all the studied indices especially with average R of 0.99 and outclassed all the other ensemble techniques. Finally, for the splitting tensile strength (Ft) of the SCGPC, the K-NN and the SVM again performed equally with an average R of 0.985 and the other performance indices. The RSM showed a strong competition with R (standard = 1) of 0.987, 0.973, and 0.986 and adequate precision (standard = 7) of 129.7, 85.3, and 123.5, for the Fc, Ff, and Ft, respectively. In addition, the RSM as a symbolic regression learning system proposed closed-form equations with which the model can be applied manually to design the production of SCGPC performing at optimal strength with optimized application of the industrial waste materials especially the most influential components, which are the GGBS, FA and NaOH. The NB came least in performance. Overall, the studied ensemble-based ML techniques applied in this present research paper outperformed the techniques used in previous literatures, except the poorly performed NB.