Modelling and Validation of Behaviour of Mushy State Rolled Al-4.5Cu-5TiB2 Composite using Neural Network Techniques

Abstract

Aluminium alloy matrix composites reinforced with in situ formed TiB2 particles are found to possess excellent mechanical properties as well as high stability at elevated temperatures. Forming of these composites by conventional methods is difficult due to their tendency of cracking. The problem is overcome by subjecting these composites to mushy state forming. Studies on mushy state rolling of Al-4.5Cu-5TiB2 composite have witnessed formation of bimodal equiaxed grains having spheroidal morphology from one that is essentially dendritic in as cast condition. Resulting mechanical and wear properties of mushy state rolled Al-4.5Cu-5TiB2 composite are also observed to be superior to that of as cast composite. The data on the grain sizes, hardness, wear and tensile properties of mushy state rolled composite has been expanded by using neural network techniques. This is done to have better understanding of the relationship between mushy state rolling process parameters and the resulting mechanical and wear properties. Artificial Neural Networks with feed forward architecture, and trained using backpropagation algorithm have been used to predict bimodal grain sizes, hardness, tensile and wear properties of Al-4.5Cu-5TiB2 composite rolled from mushy state in as cast and in pre hot rolled condition. The models have been validated by conducting mushy state rolling experiments. The composite samples in as cast and in pre hot rolled condition are mushy state rolled at pre set points within and outside the bounds of data used for training the ANN models. The validity of the models is established by way of comparison of the validation experiment results with the values predicted by models. The ANN models formulated for grain size, hardness, wear and tensile properties prediction are found to predict the corresponding outputs quite accurately, within the acceptable limits of prediction errors. Artificial Neural Networks though known for non linear mapping of complex systems, are static mapping tools in the sense that the knowledge update is based on static data provided for training the network. Simple recurrent neural networks (SRN) such as the one proposed by Elman have the capacity of dynamic learning. Thecomputational power of Elman networks has been thought to be comparable to that of finite state machines. However, such extended simple recurrent neural networks when adopted, are found to possess severely hampered learning capabilities due to convergence problems. A novel way of overcoming the problem of convergence is proposed through this work by using a Hybrid Recurrent Neural Network (HRNN) modelled from an ANN. The HRNN is modelled by borrowing weights into an Elman Simple Recurrent Neural Network having similar architecture (devoid of context layers). Such an HRNN formulated is found to converge excellently in a significantly less time as compared to an ANN for the same value of preset MSE. The prediction errors of HRNN and prediction errors resulting from ANN predictions when subjected to statistical testing are found to be equivalent. The predictions resulting from HRNNs modelled for prediction of duplex grain sizes, hardness, tensile and wear properties are seen to be in close agreement with the predictions made by the ANN models. However, it is seen that the overall time required for training HRNNs which includes the time required for training of partially trained ANNs, is significantly reduced. Thus it is observed that an HRNN modelled from a partially trained ANN has equivalent prediction capability and is superior to ANN in terms of computational time. Graphical user interface (GUI) has been designed using available API libraries which include two main modules, namely, ANN and RNN. Each model has the sub components for prediction of grain size, hardness, wear rate and tensile properties. There is provision to obtain outputs by manually feeding the input values as well for plotting line graphs by varying one parameter at a time, keeping other parameters constant. The GUI is also designed to generate bar plots by varying each mushy state processing input parameter at a time. Use of GUI is made in optimising the mushy state processing parameters for obtaining the best possible hardness values and minimum wear rate for mushy state rolled Al-4.5Cu-5TiB2 composite.