Global big data laboratory experiment, integrated with kernel-based algorithm with an improved nonlinear ensemble for compressive strength modeling

With the continuous clamor for a reduction in embodied carbon in cement, rapid solution to climate change, and reduction to resource depletion, studies into substitute binders become crucial. These cementitious binders can potentially lessen our reliance on cement as the only concrete binder while also improving concrete functional properties. Finer particles used in cement microstructure densify the pore structure of concrete and enhance its performance properties. The compressive strength of concrete made from a mixture of ground granulated blast furnace slag (GGBFS), fly ash (FA), and ordinary Portland cement was estimated using kernel regression techniques in this work. The kernel-based method offered was support vector regression (SVR), while robust linear regression (RLR), and multi-linear regression (MLR) were used as regression methods, subsequently, nonlinear average approaches were used to improve the accuracy of the prediction. Eight variables (cement, FA, GGBFS, water, superplasticizer dose [SP], coarse aggregate [CA], fine aggregate [F_ag], age) were employed as input features in 3323 data samples, and their relative value was assessed using linear correlation analysis. Following analysis, three combinations were employed to train the kernel-based models: I (inputs: cement, water, and age|output: CS), II (inputs: cement, water, FA, SP, and age|output: CS), and III (inputs: cement, water, FA, SP, CA, GGBFS, and F_ag|output: CS). The third combination gave the best testing performance with all the proposed models where their R² and MSE results after model evaluation for SVR, RLR, and MLR, are [0.984, 0.8776 and 0.8804] and [0.0019, 0.0131 and 0.0128] respectively. The study concludes that SVR with the combination III (SVR-M3) offered the best performance through effectiveness and efficiency in accurately predicting the compressive strength of the blended concrete. The prediction models should be utilized with the input variable ranges used in this work.