CONICET | Buscador de Institutos y Recursos Humanos

This article describes the development of an intelligent system for the prediction of mechanical properties of a material based on metallographic images. It consists of two parts: an algorithm that quantifies features of the microstructure from metallographic images and a system for the prediction of the mechanical properties of the investigated material. The first trials ofthe system were tested on spheroidal graphite cast irons. At first algorithm determines the proportions of graphite, ferrite and ausferrite from metallographic images. These readings are used as inputs into an oriented artificial neural network that predicts tensile strength, yield strength and fracture toughness according to the dataset, which was used for the training of the neural network. The developed neural network was of a feed-forward form and consisted of one input layer of neurons, five hidden layers of neurons and one output layer of neurons. The Tan-Sigmoid transfer function was used in the hidden layers. The training of the neural network was performed on relatively small database of mechanical properties of dual phase austempered ductile iron. In spite of the small number of data incorporated so far, the made predictions had acceptable small errors. The maximum training error of the developed artificial neural network was 5.38 %. In order to test the ability of the system to predict the properties of unknown specimens, samples were excluded from the database and the oriented neural network was used to predict their mechanical properties from their metallography. The maximum error in the prediction of mechanical properties was 7.3 %. The system can be adapted to different materials, including specific phases and different features of the microstructure. Larger databases should lead to an increased accuracy of the system.

enviar mensaje