Predicting the compressive strength of a quaternary blend concrete using Bayesian regularized neural network

Concrete produced with ordinary Portland cement (OPC) along with insertion of supplementary materials increases the level of nonlinearity. Due to this increased non-linearity and difficulty in modeling numerically, the focus has increased on the exploration of computational intelligent models like artificial neural network (ANN) to estimate different concrete properties. In this study, a quaternary blend concrete was developed with OPC, fly ash (FA), metakaolin (MK) and rice husk ash (RHA). The experimental data were further used in training the proposed ANN models to approximate its compressive strength. The proposed neural network models were trained and optimized using three different regularization algorithms; the scaled conjugate gradient “trainsc” (SCG), Levenberg–Marquardt “trainlm” (LM) and Bayesian regularized “trainbr” (BR) algorithms. The percent proportion of OPC, FA, MK and RHA making up the quaternary blends and curing days are the five features used as input variables, while the compressive strength of each of the individual concrete mixture is the output variable (target). It was found out that ANN optimized with Bayesian regularization function performed best with the highest correlation coefficient, and lowest MAE, MSE and RMSE. The results obtained from the ANN approach show significant improvement with the experimental observations.