New training strategies for neural networks with application to quaternary Al–Mg–Sc–Cr alloy design problems

• The training of a neural network in multiple stages.
• A dual stage multi-resource data training scheme using multi-objective genetic algorithm.
• Development of efficient neural network model focusing on missing, but most informative domains of the dataset.
• The scheme is used for Al–Mg–Cr–Sc alloy system.

This study concerns the training of a neural network in multiple stages considering minimization of errors from multiple data/pattern resources. The paper proposed a dual stage multi-resource data training scheme using multi-objective genetic algorithm. The training scheme has been used for the design and development of efficient neural network model focusing on missing, but most informative domains of the data set by means of introducing only a few patterns from missing domain treated separately during the later stage of training. The trained model has been used to design a quaternary Al–Mg–Cr–Sc alloy system, from the information subsets of binary Al–Cr and the ternary Al–Mg–Sc alloys. The validity of the proposed algorithm has been discussed in light of the evolution of the ageing characteristics of the new aluminium alloy system.

Figure optionsDownload as PowerPoint slide