Prediction of Thermal Conductivity and Fluid Permeability of Porous Materials by Image-Based Simulations Using Deep Generative Model-Created Three-Dimensional Images

Open-cell porous materials are used as filters, catalysts, and heat exchangers because of their high specific surface area and excellent fluid permeability. Various properties of open-pore porous materials depend on the three-dimensional pore structure. Therefore, it is essential to obtain three-dimensional images to understand the structures dominating the properties. X-ray computed tomography (CT) and serial sectioning are common methods to acquire three-dimensional images. X-ray CT can obtain three-dimensional images in less than 1 hour, but the resolution is limited in general industrial CT apparatuses. In the serial sectioning, the sample is ground and observed repeatedly with an optical microscope or scanning electron microscope. While relatively high resolution can be achieved, the time required to obtain a three-dimensional image is long.<br/>A generative adversarial network (GAN) is one of the deep generative models that create images similar to real images. Recently, it has been shown that three-dimensional images can be efficiently constructed by using a deep generative model called SliceGAN. This model enables the rapid creation of three-dimensional images by inputting one cross-sectional image of isotropic material or three cross-sectional images of anisotropic materials as training data. This model also contributes to creating high-resolution three-dimensional images if the cross-sectional images are acquired in a high resolution. However, it remains unclear whether the SliceGAN can reproduce the morphology (pore and solid parts) and related properties (thermal conductivity, fluid permeability, elastic modulus, and so on) of porous materials.<br/>In this study, we fabricated open-cell porous aluminum (Al) with porosities of 65–80% and an average pore size of approximately 400 μm by the space holder method using sodium chloride (NaCl) particles. The three-dimensional image of porous Al was taken by X-ray CT. In addition, the three-dimensional image was also constructed by SliceGAN using one cross-sectional X-ray CT image of porous Al. The image-based simulations to calculate the thermal conductivity and fluid permeability of porous Al were performed using the obtained three-dimensional images. Consistency of the porosity, pore size, and tortuosity between X-ray CT and SliceGAN images was also compared. The usefulness of SliceGAN for generating three-dimensional images of porous materials and predicting their properties was clarified by comparing the simulation results using X-ray CT images with those using SliceGAN images.<br/>When SliceGAN was trained, the input image was divided into small segments with 64 x 64 pixels. When the input image had a large pixel number, the representative structure of the porous Al was not included in the segments. In contrast, when the input image had a small pixel number, the input image did not reflect the structure of porous Al because of low resolution. Therefore, the pixel number of the input image was adjusted to a reasonable number. In this study, the number of pixels in a cross-sectional image was adjusted to include one or more cells in the segment. Under these conditions, SliceGAN could construct three-dimensional images similar to X-ray CT images.<br/>The thermal conductivity calculated using SliceGAN images was almost consistent with the thermal conductivity calculated using X-ray CT images and followed empirical power law in which the thermal conductivity is proportional to a power of the relative density. The fluid permeability calculated using SliceGAN images was also in good agreement with the permeability calculated using X-ray CT images. These consistencies were attributed to the consistency of related pore structures (porosity, pore size, and tortuosity) between SliceGAN and X-ray CT images. Thus, SliceGAN could generate three-dimensional images reflecting the characteristics of real open-cell materials by adequately adjusting the input images.