Comparative study on deep learning prediction of directional thermal conductivity of anisotropic porous media

Abstract

Accurate prediction of the directional thermal conductivity has significant guiding significance for the application of the anisotropic porous media. The prediction of directional thermal conductivity can be achieved with high efficiency by building a machine learning model. However, the traditional prediction methods of porous media thermal conductivity are usually based on image recognition machine learning model, which requires a lot of computing resources and training data, bringing great challenges to machine learning prediction. Therefore, this study proposes to use a parameter based multilayer perceptron model to establish a prediction method from the control parameters of porous media generation to the directional thermal conductivity, to improve the prediction efficiency and prediction accuracy with a small number of data sets. To compare the advantages of the proposed methods, we build three machine learning models for comparison: multilayer perceptron model, lightweight convolutional neural network, and VGG19 convolutional neural network. The results show that for a small number of training data sets, the multilayer perceptron model based on control parameters is superior to the convolutional neural network model based on image prediction. The MRE of the MLP is improved by 2.24 % compared to the lightweight CNN. In addition, for a limited dataset, the prediction accuracy can be effectively improved by lightweight CNN model, and the MRE is improved by 7.95 %.