Session: 03-01-01: General Topics of Aerospace Materials

Paper Number: 105542

105542 - Predicting Effective Thermal Conductivity of Heterogeneous Microstructure Materials by Convolutional Neural Network 

Thermal insulation materials are usually porous or fibrous with a low thermal conductivity. Their microstructures are very complex at the nanometer or micrometer scales, such as aerogels and cellular foams. Traditional micromechanics-based methods, such as, homogenization and finite element method, etc., can hardly predict the effective thermal conductivity accurately, due to the limited understanding of microscopic mechanism of heat transfer and the large amount microstructural data with multi-scale characteristics.

In this paper, the effective thermal conductivity of composite materials is predicted by the convolutional neural network (CNN). Firstly, the digital image models of composite microstructure are generated by Quartet Structure Generation Set (QSGS) method and Random Generation-Growth Method (RGGM). Then, the effective thermal conductivities are predicted by the Lattice Boltzmann Method (LBM). The accuracy of LBM is verified by the experiment results of aerogel materials. With these microstructural images as the input data and predicted conductivities as the output data, the CNN model can be trained and validated. Then, the characteristics including material porosity, anisotropy of microstructure, and size effect are investigated through the parametric studies. In addition, CNN predictions are compared with both analytical and computational micromechanics methods. Finally, the proposed CNN model is also used to predict the thermal conductivity of materials with novel microstructures, which is beyond the training sets.

This work demonstrates that with sufficient and accurate training data, CNN model predicts the effective thermal conductivity of heterogeneous materials more efficiently than traditional methods. It can implicitly capture the scattering characteristics of microstructures and cross-link them to macroscopic behaviors of the materials by the artificial intelligence.

Presenting Author: Chengcheng Shen University of Chinese Academy of Sciences

Presenting Author Biography: Chengcheng Shen is a Phd candidate majoring in computer application technology. Her research interests are in the areas of scientifc computations with neural network methods to understand the multiscale phenomena in materials and solids on the basis of their microstructures.

Authors:

Chengcheng Shen University of Chinese Academy of Sciences
Haifeng Zhao University of Chinese Academy of Sciences
Qiang Sheng University of Chinese Academy of Sciences