Learning microstructure–property mapping via label-free 3D convolutional neural network

Predicting the physical property of a class of microstructures is crucial in material design, structural simulation, and design. Property prediction may be conducted millions of times in these studies and is better derived instantly for computational efficiency. This issue is addressed in this study via building a mapping from a 3D microstructure to its effective material property, or called structure–property mapping, using a 3D convolutional neural network (CNN). Unlike the direct approach using labeled simulation data, the mapping is based on the physical knowledge of the structure–property relationship determined by its underlying PDE equations. The knowledge is embedded in the loss function of the CNN framework, which is designed and tested under several different formulations to improve its training convergence. Ultimately, the derived structure–property mapping can instantly predict the associated material property for a given microstructure and has far better generalization ability than the data-labeled approach, as demonstrated via numerical examples.