Convolutional neural network for fast prediction of the effective properties of domains with random inclusions

Abstract

We consider a heterogeneous domain with random inclusions in two-dimensional and three-dimensional formulations. For generation of the train and test datasets, we numerically calculate the effective properties for a given geometry of the heterogeneous domain. We construct a machine learning method to learn a map between local heterogeneous geometries and effective properties. We present numerical results for prediction of the effective properties for 2D and 3D model problems.