Modeling the compressive strength of concrete at different curing regimes using machine learning

In this research work, a representative data of 334 collected from literature of concrete admixed with highly pozzolanic industrial waste materials, which are nano-silica (nSi), nano-alumina (nAl) and micro-silica (mSi) has been deployed to the learning abilities of the artificial neural network (ANN), genetic programming (GP), and the evolutionary polynomial regression (EPR) under the influence of the variable structure neuron interface in the hidden layer and the levels complexity of the techniques. The effect of the varying neuron and levels of complexity has not been studied previously. The datasets were utilized in the ratio of 75% for training and 25% for validation. The output of the nSi-, mSi- and nAl-precursor concrete is the compressive strength tested at different curing regimes. The relative importance values for each input parameter were studied.The results indicated that (Age) has the most influence than the regular concrete components; C, W, FAg, CAg in the mixes then the additives; FA,mSi, nAl, nSi, and PL. The Taylor diagram which compared the accuracies of the developed models and the variance distribution of the models, respectively were carried out. In the relation between the predicted and measured values of the Fc, the GP produced a fit line parametric expression of y = 0.967 × with MAE of 6.97 MPa, RMSE of 8.6 MPa and R2 of 0.693, the EPR produced a fit line parametric expression of y = 0.987x, with MAE of 3.97 MPa, RMSE of 5.3 MPa, and R2of 0.900, and the ANN produced a fit line parametric expression of y = 0.997x, MAE of 2.32 MPa, RMSE of 3.16 MPa, and R2 of 0.966. Generally, it can be deduced that the ANN maintains the decisive model with the lowest error in computation and the least outliers from the ± 25% accuracy envelop. This outcome agrees with previous application of the ANN, with the increased neurons adding to the superiority of the present application of the ANN. Finally, the varying levels of complexity and thestructure neurons of the models’ interface of computation have contributed to the performance of these models as they reduced the error and increased the other performance metrics.