Modelling Charpy impact energy of heat-treated-steel using efficient neural-networks & genetic-algorithms

Although neural networks (NN) are known to represent a powerful tool for mapping non-linear relationships between the inputs and outputs, their structures are typically set up by the modeller either by trial-and-error or based on their own experience. The latter is not only a time-consuming operation but also creates a risk that the best model structure is not necessarily selected. In this paper a genetic algorithm (GA) based strategy of determining the optimal NN-based model structure, together with the best training options for NN modelling will be described. Initial results of this GA-NN based modelling paradigm are promising, with the model performance being significantly improved when compared to previously elicited where the model structures were defined in an 'ad-hoc' fashion.