Predição de propriedades mecânicas de compósitos unidirecionais através de redes neurais artificiais

Abstract

The composite materials are a new highlight in the technological advancement, consequently leading to the development of new researches due to its growing demand in the most diverse areas. Among these researches, arise those that have the objective to facilitate the application of these materials, through a fast estimation of its mechanical properties, without the need for experimental procedures, with this being the main factor in the projects preparation. Thus the micromechanical models appeared, which gained importance due to its practicality, such as the Mix Rule and the Halpin-Tsai equations. Recently, new computational models are combining micromechanical models and perfecting them to obtain maximum accuracy, as for instance in the Artifical Neural Networks application. Therefore, this work aims to create an Artificial Neural Network (ANN) architecture capable of modeling the shear modulus and ultimate longitudinal stress of unidirectional composites. When the ANN´s are trained and tested, they will serve as computational tools, similar to functions, where an input is supplied to obtain a desired output. To achieve this goal, it was necessary a collection of data in literature, which were divided in a training group and a testing group, with the cross validation between them being performed. Seven different types of architectures were developed, three for the G12 and four for the Xt, each of these with two, three and four inputs. Among these models, three of them are considered mixed models, which combines values from the output of the ANN with values obtained from the micromechanical models, such as the Halpin-Tsai. After the ANN training, a comparative analysis was performed between the values from the ANN and the experimental values, with quantitative and qualitative analysis being performed with the Halpin-Tsai model as a base for comparison, presenting higher values for the correlation coefficient and smaller values for the root mean square error.