FluidNeRF: Learning 3D Flow Fields from 2D Projections

A new method for 3D flow visualization is presented. FluidNeRF is based on the concept of Neural Radiance Fields (NeRF), whereby a volume is represented using a deep neural network and trained via 2D image projections. The machine learning based framework of FluidNeRF allows for the easy integration of multiple constraints, including both data and physics constraints, thus enabling data assimilation, fusion, and compression. FluidNeRF is demonstrated using both synethic image data of a turbulent jet and experimental image data of plumes and sprays obtained with a custom-built array of 18 cameras. FluidNeRF achieves better performance than traditional ART based reconstruction methods with orders of magnitude reduction in computer memory requirements. Details of the FluidNeRF framework are presented, including the influence of hyperparameters, camera array size, various loss terms (data, physics, boundary), and flow complexity.