Effect of number of neurons and layers in an artificial neural network for generalized concrete mix design

Selection of the number of neurons in different layers of an artificial neural network (ANN) is a key decision-making step involved in its successful training. Although the number of neurons in the input layer is decided by the number of input parameters, and similarly, in the output layer the number of neurons is fixed by the output parameters, however, the number of neurons in the hidden layer is not fixed, whereby the overall efficiency of the ANN depends upon the correct modeling of the realistic scenario in the form of a neural network. Concrete can be mixed in a huge variety of methods and compositions to achieve a specific or a combination of specific results. These results have been observed to repeat and follow patterns. Furthermore, for several types of concrete mixes, only actual physical trial data are available with no mix design method; this makes achieving different required results very complex and expensive task. Having an ANN to be able to predict a concrete mix to handle such variations will save a lot of time and money and will prove to be a universal mix design tool. A thorough analysis for deciding the number of neurons for mix design of concrete has been carried out in this research. The results of neural networks for a varying number of neurons and layers have been correlated. The data for a number of concrete mixes were sorted and normalized. Properties of concrete constituents like specific gravity of cement, coarse and fine aggregates, dry density of coarse and fine aggregates, cement type, type of mineral admixture, water-to-cement ratio, type and temperature of curing and hardened properties like compressive strength, modulus of elasticity and tensile strength were taken as input parameters. The contents of constituents of concrete such as water content, cement content, coarse aggregates content, fine aggregates content and content of mineral admixtures are regarded as output parameters. For 17 inputs and 5 outputs, it has been discovered that simple neural network with single or double hidden layers performed better that 3 or more such layers.