An optimization methodology to design a front fendor for minimal air drag using ANN and firefly algorithm

 For introduction on meta-modelling/surrogate modelling.

An optimization study of front fendor of motorcycle to reduce the drag is performed in this project. An ANN based meta model is used here to imitate the CFD analysis and replace the CFD computation time. The firefly algorithm is used to perform a shape optimization study of fendor shape. The shape is parametrically defined via Bezier Curves, with the design constraints. The Artificial neural networks are used here to create the meta model to replace the solver, and hence reducing the computation time. The results of parametrized model shape are obtained using ANN and the optimization is carried out using firefly algorithm, which leads to optimized design with respect to aerodynamic drag within the design constraints. The outcome of the study was to develop a methodology with the combination of ANN and firefly algorithm with its application on this problem. Response surface methodology for design optimization is widely popular as it aims at reducing the expensive analysis methods and their associated numerical errors.

    Artificial neural network (ANN) is a powerful modelling technique that offers several advantages over conventional modelling techniques because they can model based on no assumptions concerning the nature of the phenomenological mechanisms and understanding the mathematical background of problem underlying the process and the ability to learn linear and non-linear relationships between variables directly from a set of examples. The ANN model is potentially more accurate by including all the experimental data. The reporting ability of feed-forward architecture of ANN and , also known as multilayered perceptron (MLP) with back-propagation (BP) algorithm, was selected and trained in this study to develop predictive model. The first step of ANN modelling was to optimize a neural network with the aim of obtaining an ANN mode with a minimal dimension and minimal errors in training and testing. The design of experiments and their respective analysis yield was used for training the network. Depending on the nonlinearity of the problem and the number of parameters, an ANN model may require a high computational cost to create. It is noted that ANNs perform better than the other techniques, especially RSM when highly non-linear behaviour is the case. Also, this technique can build an efficient model using a small number of experiments; however, the technique accuracy would be better when a larger number of experiments are used to develop a model.

The firefly algorithm, a meta heuristic algorithm, was used to optimize the response surface generated by ANN, to find out the design within constraints with least aerodynamic drag. The firefly algorithm, an evolutionary algorithm similar to GA, imitates the behaviour of fireflies and their movement based on the brightness of the flash generated by them. The CFD analysis of 1 run usually takes 8-10 hours, however after learning, ANN can predict the results in few seconds, within a certain accuracy.