High-strength concrete compressive strength prediction using an evolutionary computational intelligence algorithm

Abstract

Abstract The most crucial mechanical property of concrete is compression strength (CS). Insufficient compressive strength can therefore result in severe failure and is very difficult to fix. Therefore, predicting concrete strength accurately and early is a key challenge for researchers and concrete designers. High-Strength Concrete (HSC) is an extremely complicated material, making it challenging to simulate its behaviour. The CS of HSC was predicted in this research using an Adaptive Neuro-fuzzy Inference system (ANFIS), Backpropagation neural networks (BPNN), Gaussian Process Regression (GPR), and NARX neural network (NARX) In the initial case, whereas in the second case, an ensemble model of k-Nearest Neighbor (k-NN) was proposed due to the poor performance of model combination M1 & M2 in ANFIS, BPNN, NARX and M1 in GPR. The output variable is the 28-day CS (MP) and the input variables are cement (Ce) Kg/m3, water (W) Kg/m3, superplasticizer (S) Kg/m3, coarse aggregate (CA) Kg/m3, and Fine aggregate (FA) Kg/m3. The outcomes depict that the suggested approach is predictively consistent for forecasting the CS of HSC, to sum up. The MATLAB 2019a toolkit was employed to generate the MLs learning models (ANFIS, BPNN, GPR, and NARX), whereas E-Views 11.0 was used for pre-and post-processing of the data, respectively. The model for BPNN and NARX modelling was trained and validated using MATLAB code. The outcome depicts that, the Combination M3 partakes the preeminent performance evaluation criterion when associated to the other models, where ANFIS-M3 prediction outperforms all other models with NSE, R2, R = 1, and MAPE = 0.261 & 0.006 in both the calibration and verification phases, correspondingly, in the first case, In contrast, the ensemble of BPNN and GPR surpasses all other models in the second scenario, with NSE, R2, R = 1, and MAPE = 0.000, in both calibration and verification phases Comparisons of total performance showed that the proposed models can be a valuable tool for predicting the CS of HSC.