Prediction of Random Fields of Mechanical Properties from Microstructure Images using Convolutional Neural Networks
Please login to view abstract download link
For the accurate simulation of the mechanical behavior of random composite materials, establishing microstructure consistent mechanical properties is of great importance. This can be achieved by computing random fields of material properties from microstructure images using a moving window technique, which typically involves numerical homogenization of a large number of stochastic volume elements (SVEs). In this work, the expensive FE-based homogenization procedure is substituted with a convolutional neural network (CNN), which takes as input an image of the SVE and returns its apparent mechanical properties, thus leading to nearly instant random field prediction. Training is performed on randomly generated SVEs with varying volume fractions and inclusion positions, which are representative of the SVEs obtained during processing of the initial microstructure image. The trained CNN shows good performance in predicting the mechanical properties of SVEs and is applied, along with the moving window technique, for the prediction of random property fields. The proposed CNN-based method is tested on simulated and real microstructures, leading to accurate and nearly instant (in a few seconds) random field predictions, which could otherwise take several hours using the conventional FE-based approach.
