Temperature field prediction for various porous media considering variable boundary conditions using deep learning method

The prediction of temperature fields in porous media is challenging owing to the variable boundary conditions attributed to the working condition and topological structure. In this study, a supervised convolutional neural network (CNN) is built to predict the temperature field and effective thermal conductivity of sphere-packed and irregular porous media under various boundary conditions. Datasets of temperature fields, obtained using lattice Boltzmann method (LBM) simulations based on three-dimensional sphere-packed porous media, are employed to a CNN for training. The CNN achieves an accurate and fast prediction of the temperature field and effective thermal conductivity with different boundary conditions (the temperature differences between the inlet and outlet are 25, 50, 100, 125, 150, 175 degrees C). The relative errors for the effective thermal conductivity between the CNN and LBM are 0.7-22.8% for the sphere-packed porous media and 3.1-16.0% for the irregular porous media. For a typical case of sphere-packed porous medium with a porosity of 0.6, the computation time using CNN is 1.53 x 10(-2) h, while that of LBM is approximately 720 h. These findings mean that the CNN is promising for the prediction of the heat transport properties of porous media with different morphologies and variable boundary conditions.