Prediction of three-dimensional flow field inside realistic fibrous filter obtained from x-ray computed tomography images using deep convolutional neural networks

Abstract

Deep-learning models garnered considerable attention in the field of fluid mechanics for physics discovery and approximation-model generation. This study aims to develop an approximation model to predict the flow field inside realistic fibrous filters based on an image-to-image approach to replace three-dimensional (3D) computational fluid dynamics (CFD) simulations, which are computationally expensive and difficult to apply to realistic fibrous filters. A data-driven framework is proposed using deep convolutional neural networks (CNNs) to provide a per-pixel prediction of the flow field. The model inputs are two-dimensional x-ray computed tomography images, whereas the outputs are the 3D distributions of the velocity vectors and pressure. High-resolution 3D CFD simulations are performed to create a database to train and test the CNN model. The model is applied to surgical and N95 face masks. The relative error of the CNN model over the test dataset is approximately 10% in regions with high velocity and pressure, and the model can provide a detailed high-resolution prediction of the flow field with a speedup of about three orders of magnitudes. A strict generalization test is conducted for completely unseen 3D segments with complex microstructures. The model generalizability still needs more improvements; however, the model can provide a low-resolution 3D flow field for those segments that can be used as the initial condition for CFD simulation to reduce the CFD computational time. This framework can be utilized for other types of filters and provides a basis for the design and optimization of fibrous filters.