Modelling the effective thermal conductivity of an unidirectional composite by the use of artificial neural networks

This paper explores the application of pattern recognition and artificial intelligence techniques in the characterization of a multiphase realistic disordered composite and in the design of a multiple regression model to estimate effective thermal conductivity. An image database of computer simulated microstructures was generated. Some descriptors based on boundary and area shapes of Voronoi cells were extracted for each fiber distribution. Several approaches have been used to reduce the high original dimensionality. Selected features can be introduced as inputs in a multiple regression model. This procedure provides an alternative to the finite element method for the computation of effective thermal conductivity. Different regression models (classical and neural approaches) have been considered and a randomised resampling procedure has been designed in order to choose the best estimation model from a statistical point of view.