Application Of Neural Network In Optimal Selection Of Aggregate For Preparation Of High Durable Concrete

Coarse aggregates occupy more than 50% of concrete’s volume and have a very wide range of the source, so that it’s hard to control the quality. It is necessary to choose the coarse aggregates with high quality, because its quality directly affects the durability of concrete. In the past, the influence of coarse aggregates on the durability of concrete was mostly studied by experiments. But it wastes time and manpower. Therefore, it is necessary to find a method, which can predict the durability of concrete with different coarse aggregates more convenient and quickly. As a result of the relationship between coarse aggregates and the durability of concrete is nonlinear, it is difficult to establish mathematical expressions. In this paper, the artificial neural network theory is introduced to study the concrete’s durability.To study the impermeability, frost resistance, carbonation resistance of the concrete, different mineralogical coarse aggregates(average particle size, water absorption, density and strength) was selected. When the water-cement ratio is 0.32, the concrete with coarse aggregates of 5~30 mm particle size shows the best impermeability. The concrete with coarse aggregates of 5~20 mm particle size with water-cement ratio of 0.40 and 0.49 shows the best impermeability. The rule between the particle size and the carbonation resistance is similar to the influence of the impermeability. But the influence of the particle size on the frost resistance is not obvious. The concrete with basalt shows the best impermeability at three water-cement ratios. When the water-cement ratio is 0.32, the concret with limestone shows the worst impermeability. At the other water-cement ratio, the concrete with granite shows the worst impermeability. It’s similar to the influence of the type of coarse aggregates on the frost resistance and carbonation resistance.The data was collected to develop the neural network model. Compared with resilient back propagation, additional momentum back propagation and variable learning rate back propagation, the Levenberg-Marquardt back propagation shows better training time, prediction error and convergence. So it was choosed for training.Based on analyzing the factors of coarse aggregates influence the durability of concrete, six factors(water-cement ratio, curing age, particle size, water absorption, apparent density, strength of coarse aggregates) was took as the input layer neurons to predicte the impermeability and carbonation resistance of concrete. At the same time, the chloride diffusion coefficient and the carbonation depth of concrete was took as the output layer neuron. Based on the 108 data sets of the impermeability and the 135 data sets of the carbonation resistance, the model was developed. They were 6-17-1 and 6-15-1 for the impermeability and carbonation resistance of concrete respectively. To predict the frost resistance of concrete, took the six factors as the input layer neurons: water-cement ratio, freezing and thawing cycles, particle size, water absorption, apparent density and strength of coarse aggregates. The relative dynamic elastic modulus of concrete was took as the output layer neuron. Based on the 103 data sets of frost resistance, the model(6-21-1) was developed. The average relative errors of prediction of test results were found to be 4.44%, 4.15%, 5.16% respectively, which met the requirement of engineering(within 6%). It found that the predicted value was closed to the measured value.In order to verify the applicability of the model, some data sets of impermeability and frost resistance of concrete in engineering were collected. The average relative errors of prediction of the date sets were found to be 11.00% and 9.85% respectively. The prediction error was bigger than this paper’s, but it could be met the requirement of engineering.