Controlling green sand mould properties using artificial neural networks and genetic algorithms — A comparison

Optimum formulation of the green sand mixture has remained as a critical problem for the foundryman, for several years. In the present work, experiments were conducted with varying grain fineness number, clay percentage, moisture percentage, mulled time and hardness with an objective to formulate the green sand mixture optimally. The corresponding mould properties namely green compression strength, green shear strength, permeability, dry compression strength and dry shear strength have been measured. first investigation, the above data were fed to a back propagation artificial neural network. The inputs to the neural network were mould properties and the outputs were grain fineness number, clay percentage, moisture percentage, mulled time and hardness. (In this paper, outputs of neural networks and genetic algorithms will be termed as ‘mould controlling parameters’ or simply ‘controlling parameters’). After training, the network was asked to predict the set of controlling parameters, which could yield a particular set of desired mould properties. (This particular set of desired mould properties was one of the sets of mould properties obtained from experiments and the same was excluded in the training of neural network). The set of controlling parameters predicted by the neural network was compared with the set of experimental controlling parameters by calculating percentage error. second investigation, the same inputs and outputs (used to train the neural network) were fed to a micro genetic algorithm. The objective function was designed such that it predicts the set of controlling parameters, which would realize the same set of desired mould properties. The set of controlling parameters predicted by the genetic algorithm was compared with the set of experimental controlling parameters by calculating percentage error. The comparison revealed that the percentage errors obtained by genetic algorithm were small in most of the cases. It was also found that genetic algorithms could offer few advantages over artificial neural networks in the present task.