Utilizing Artificial Neural Network and Multiple Linear Regression to Model the Compressive Strength of Recycled Geopolymer Concrete

Abstract

Based on the heterogeneity nature of concrete constituents and variation in its compressive strength over several orders of magnitude for various types of concrete, predictive methods to evaluate its compressive strength has now been given an appropriate consideration. Therefore, this study investigates, the performance of Artificial Neural Network, ANN in forecasting the compressive strength of hybrid alkali-activated recycled concrete (HAARC) and compared with the traditional Multiple Linear Regression, MLR. The developed models utilized the experimental results where varying material quantities were used. The ANN and MLR models were developed using eight input variables namely; Ordinary Portland cement (OPC), Rice Husk Ash (RHA), Coal Fly Ash (CFA), Crushed granite (CG), Cupola Furnace Slag (CFS), Alkaline Solution (AS), Water-Binder Ratio (WB) and the Concrete Age (CA) while compressive strength was the only response (predicted) variable. The input data were trained, tested and validated using feedforward back-proportion and backward elimination approach for ANN and MLR, respectively. The most probable model architecture containing eight-input layer, thirteen-hidden layer, and one-output layer neurons was selected based on satisfactory performance in terms of means square error MSE, after several trials. MLR results revealed that only three input variables; CFA, CG and CA proves to be statistically significant with p-values below 0.05. Performance evaluations of the developed models using coefficient of determination, R², Mean Square Error, MSE, Root Mean-Square Error, RMSE and Mean Absolute Percentage Error, MAPE, showed that ANN prediction accuracy is better than that of MLR with R² = 0.9972, MSE = 0.4177, RMSE = 1.8201, MAPE = 2.2935 for ANN and R² = 0.7410, MSE = 66.6308, RMSE = 290.4370, MAPE = 385.5221 for MLR.