Study of effective elastic properties of heterogeneous materials with an artificial neural network model

This paper is devoted to the estimation of macroscopic elastic properties of composites containing inclusions and pores. Some representative analytical homogenization models are first recalled and their weaknesses are highlighted by comparisons with the reference numerical results obtained from direct simulations by a Fast Fourier Transform (FFT) method. A large data set is then constructed also based on numerical results obtained from FFT simulations for different types of micro-structures and for a large range of elastic properties of constituents. An artificial neural network (ANN)-based model containing two hidden layers is constructed and trained by using this data set. Different types of validation tests of the ANN model are presented. It is found that the ANN-based model can estimate the macroscopic elastic properties of strongly heterogeneous materials with a very good accuracy, for different types of micro-structures and a large range of porosity and inclusion volume fraction.