Predicting effective thermal conductivity of fibrous and particulate composite materials using convolutional neural network

Abstract

Thermal insulation composite materials, such as aerogels and cellular foams, typically exhibit great performance such as low thermal conductivity and complex microstructures at the nanometer or micrometer scale. However, traditional micromechanics-based methods like homogenization and finite element analysis may not accurately predict their effective thermal conductivity due to a limited understanding of microscopic heat transfer mechanisms and the vast amounts of multi-scale microstructural data. In this paper, the effective thermal conductivity of composite materials is predicted using a convolutional neural network (CNN). To model the microstructure of the composites, digital images are generated using the Quartet Structure Generation Set (QSGS) method and the Random Generation-Growth Method (RGGM). The Lattice Boltzmann Method (LBM) is then used to predict the effective thermal conductivities. The CNN model is trained and validated using the microstructural images as input data and the conductivities predicted by LBM as output data. Subsequently, Parametric studies are conducted to investigate the material characteristics, including volume fraction and microstructural anisotropy. Additionally, the CNN predictions are compared with the results of experiments, analytical models, and computational methods. Finally, the proposed CNN model is utilized to predict the thermal conductivity of materials with novel microstructures that are not included in the training set. By capturing the scattering characteristics of heterogeneous materials through artificial intelligence, the CNN model predicts thermal conductivity more efficiently than traditional methods. This highlights the potential of machine learning to advance materials science and accelerate development of materials with desired properties.