(353b) Convolutional Neural Network for Predicting Morphology, Flow and Transport Properties of Complex Materials

Physical properties of porous media and materials, such as their permeability, electrical conductivity, and elastic moduli are controlled by their morphology, i.e., their pore-size distribution, pore shapes, and connectivity. Experimental studies provide much insight, but they often are costly. Computational studies that involve model of such materials and media involve assumptions, approximation, and simplifications. On the other hand, advancements in imaging techniques have made it possible to use two- and three-dimensional (3D) images of materials in direct computations without any assumptions or simplifications. With advances in the development of convolutional neural networks (CNNs), it is now possible to combine the CNNs with images of porous materials and media in order to make accurate predictions for their physical properties. But one must first address the issue of lack of a large dataset, as well as the required resolution in the available images.

Using a recently developed stochastic reconstruction technique, we have generated big data for use in a CNN in order to enhance the resolution of porous materials’ images, with the enhancement evaluated quantitively. We have also developed a hybrid and physics-guided deep learning approach for estimating the permeability of porous materials, which enables us to establish a direct link between the morphology and to the flow properties. Since several stochastic reconstruction algorithms have been developed recently, we have also quantified their using a deep-learning method.

References

Kamrava, S., Tahmasebi, P., & Sahimi, M. (2019). Enhancing images of shale formations by a hybrid stochastic and deep learning algorithm. Neural Networks, 118, 310-320.

Kamrava, S., Tahmasebi, P., & Sahimi, M. (2020). Linking morphology of porous media to their macroscopic permeability by deep learning. Transport in Porous Media, 131(2), 427-448.

Kamrava, S., Sahimi, M., & Tahmasebi, P. (2020). Quantifying accuracy of stochastic methods of reconstructing complex materials by deep learning. Physical Review E, 101(4), 043301.