Efficient multiscale modeling of heterogeneous materials using deep neural networks

Material modeling using modern numerical methods accelerates the design process and reduces the costs of developing new products. However, for multiscale modeling of heterogeneous materials, the well-established homogenization techniques remain computationally expensive for high accuracy levels. In this contribution, a machine learning approach, convolutional neural networks (CNNs), is proposed as a computationally efficient solution method that is capable of providing a high level of accuracy. In this work, the data-set used for the training process, as well as the numerical tests, consists of artificial/real microstructural images (input). Whereas, the output is the homogenized stress of a given representative volume element RVE. The model performance is demonstrated by means of examples and compared with traditional homogenization methods. As the examples illustrate, high accuracy in predicting the homogenized stresses, along with a significant reduction in the computation time, were achieved using the developed CNN model.

Automated summary

Material modeling using modern numerical methods accelerates the design process and reduces costs. The well-established homogenization techniques for multiscale modeling of heterogeneous materials are computationally expensive. This paper proposes the use of convolutional neural networks (CNNs) as a computationally efficient solution with high accuracy. The CNN model, trained on artificial/real microstructural images, accurately predicts the homogenized stresses of representative volume elements (RVEs) with reduced computation time.