The deep material network can be used to learn the proper topological structure of a given RVE. In the following example, treemaps of trained material networks for three different 2D RVEs with network depth after 10000 epochs (1% training error) are shown.
Material networks can also be trained for 3D RVEs. Applications to polymer nanocomposite and CFRP systems are addressed.
​
Geometric RVE information, such as the volume fraction, can be accurately extracted from pure mechanical data. The material network is intrinsically parameterized for material design purpose.
​
Other than the virtual material testing data from RVE analysis, experimental data can also be incorporated in the offline-training process.
Histories of training and validation errors
Evolution of topological structures of 2D and 3D material networks