APPLICATION OF NEURAL NETWORK FOR PREDICTING PROPERTIES OF CONCRETE USING BIDA NATURAL GRAVEL AS COARSE AGGREGATE

Abstract

The production of crushed granite, which is one of the conventional coarse aggregates used in concrete is expensive and energy demanding. This production process leads to the emission of dust particles and also generates noise leading to environmental hazards, which are harmful to humans. But rapid increase in population creates drive towards increase in infrastructural development. This has, in the last decade, overstretched crushed granite in an alarming rate. Many approaches have been used to develop models for predicting the properties of concrete containing various aggregate types. This study is solely focused on applying Artificial Neural Network (ANN) in predicting properties of concrete using Bida Natural Gravel (BNG) as coarse aggregate due to its abundance in the environ. Physical and mechanical properties of the fine and coarse aggregates were determined. Three water-cement ratios (w/c) of 0.40, 0.50 and 0.60. and three coarse aggregate-total aggregate (ca/ta) ratios of 0.55, 0.6 and 0.65 as well as three total aggregate-cement ratios (ta/c) ratios of 3.00, 4.50 and 6.00 was used for factor settings. Full factorial experimental design was used to generate twenty-seven (27) experimental data points. Combinations of constituent materials in each experimental data point were used to produce concrete mixtures. Slump of the concrete were determined and the compressive, flexural, splitting tensile strengths and modulus of elasticity were determined at 28 days curing age. The highest slump of 270 mm was recorded using w/c of 0.60, ca/ta of 0.55 and ta/c of 3.00 while zero slump was recorded using w/c of 0.40, ca/ta of 0.65, 0.6.0 and 0.55 and ta/c ratio of 6.00. Highest compressive, flexural and splitting tensile strengths of 44.30, 7.60 and 3.42 N/mm2 as well as modulus of elasticity of 32.74 kN/mm2 was recorded using low w/c ratio of 0.40, medium ratio of 0.55 and low ta/c ratio of 3.00 while the lowest compressive, flexural and splitting tensile strengths of 7.79, 1.60 and 0.57 N/mm2 respectively and elastic modulus of 4.09 kN/mm2 was recorded using low w/c ratio of 0.40, medium ca/ta ratio of 0.60 and high ta/c ratio of 6.00. The results obtained were augmented using a MATLAB script and the augmented data sets were used to develop two case ANN models for slump, compressive, flexural, splitting tensile strength and modulus of elasticity using a MATLAB back propagation, feed-forward ANN algorithm. Mean Square Error (MSE), Root Mean Square Error (RMSE) and Regression (R) were used to examine the performance of the models. A 5-89-1 ANN architecture with a tangent sigmoid activation function was found to be sufficient in predicting slump data for concrete using Bida Natural Gravel (BNG) as aggregate. A 5-69-1 ANN architecture with tangent sigmoid activation function was found to be sufficient in predicting compressive strength data, while a 5-91-1 ANN architecture with logistic sigmoid activation function was found to perform best in predicting flexural strength of concrete using BNG as coarse aggregate. Architecture with 5 input neurons, 91 hidden neurons and 1 output neuron (5-91-1) was adjudged to best predict the splitting tensile strength of concrete containing BNG using tangent sigmoid activation function, while a 5-67-1 ANN architecture with a logistic sigmoid activation function was selected for predicting the elastic modulus of concrete containing BNG. The ANN models developed herein can be used in predicting properties of concrete using BNG as coarse aggregate with 98% accuracy.