finite element & artificial nueral network analysis of ecap

Equal channel angular pressing (ecap) is one of the most promising technique among the developed severe plastic deformation (spd) techniques for the induction of strain in bulk metals. during this process, repetitive pressings, the shear strain is accumulated in the billet, leading ultimately to a ufg structure. in this study finite element method (fem) and artificial neural network (ann) were used to simulate ecap deformation of aa6063 aluminum alloy. in this research, in order to generate model according to artificial neural network for the ecap process and evaluation the effect of its die channel angles, a lot of data will be needed for training and testing of artificial neural network. despite the fact that the experimental results are real, due to the high cost and the inability to provide some numerical parameters (such as strain distribution, inhomogeneity index), the finite element simulation software, abaqus/explicit, was used. in this modeling, different states of angle between the channels (φ) and the outer corner angle (ψ) and friction coefficient were considered and the modeling outputs of the maximum force for the extrusion, strain inhomogeneity index and effective strain were considered. in this research, two artificial neural networks were used to estimate the function separately. one of artificial neural networks was chosen to predict the maximum force and another neural network was chosen for the prediction of effective strain and strain heterogeneity index. in order to train the network, after collecting the data from the finite element method simulation, they were subjected to a suitable pre-processing and duplicate data were removed. after that, 80% of the data was used for training, 15% for validation and 5% for testing the network. the type of network used in this research is back propagation and the training algorithm of levenberg–marquardt. neural network has two hidden layers, the first layer has 10 neurons and the second layer has 20 neurons, and its transfer function is tansigmoid. after confirming the validity of the model with experimental data, a number of parameters are considered to be effective in usage of neural network for estimating. the die channel angles (angle between the channels φ and the outer corner angle ψ) and friction coefficient which were subsequently used for training the ann. finally, the outputs of the neural network were compared with the results of the finite element method. the results of the comparisons showed that the trained neural network with an average error of 10% predicted the results of the finite element method well.