Extracting the properties of constituent phases from the overall response of composites: A deep neural network method

Due to the limitations of available experimental techniques, it is often difficult to directly measure the mechanical properties of a specific constituent phase in a heterogeneous material, e.g., cancerous cells embedded in the extracellular matrix of a tumor. In this paper, we propose a machine learning-based approach to extract the elastic properties of constituent phases from the macroscopically effective responses of a composite material. Through a number of numerical experiments, it is demonstrated that the trained deep neural network (DNN) model can efficiently solve this inverse analysis problem and provide an accurate mapping between the mechanical properties of constituent phases and the microstructures and properties of the composite. We optimize the structures of the deep neural networks and the number of data sets to ensure the high accuracy and good performance of the proposed DNN method. This work holds promise for applications in diverse fields, such as materials science and medical engineering.