Reconstruction of porous media geometry from sparse velocity measurements using convolution neural networks

The study of transport phenomena, using non-destructive imaging technologies and simulations of fluid flow through complex geometries, has become increasingly vital, particularly in the biomedical domain. Such approaches provide valuable insight into critical processes, including blood flow in tissues, drug delivery, and extracellular transport, contributing to our understanding of physiology and medical interventions. However, significant challenges persist in extracting geometric features from imaging and reconstructing the flow domain for direct simulations. Recent advancements in machine learning techniques offer a promising solution by regenerating domain geometry based on relatively sparse sampling of flow velocity. Motivated by this, we apply deep learning techniques to predict porous media geometry, utilizing convolutional neural networks (CNNs). A CNN is trained on the velocity field or particle trajectories acquired from 2D Lattice Boltzmann simulations. We conduct an analysis in which we vary the complexities of porous media geometry and attain promising estimations of the flow domain using sparse velocity measurements. Extending this methodology to experimental data will notably aid in the development of quantitatively accurate predictive models with biomedical applications.