FluidNeRF: A machine learning framework for 3D flow field reconstructions

A new machine learning framework for 3D flow field reconstructions, termed here as FluidNeRF, 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 image projections. FluidNeRF is an inherently modular framework that seamlessly allows for data assimilation, fusion, and compression. The methodology is demonstrated here using image projections of a passive scalar field generated from a DNS simulation of a turbulent jet. 3D flow field reconstructions are compared with reconstructions obtained with an adaptive simultaneous algebraic reconstruction technique (ASART) algorithm. The influence of hyperparameters and experimental arrangement (e.g. number of cameras) on the 3D reconstruction quality are presented. FluidNeRF is shown to produce comparable or better reconstruction quality than ASART with orders of magnitude reduction in computer memory requirements. Incorporation of temporal information and physical priors via additional loss functions will be introduced. Preliminary experimental results will be demonstrated using a vortex ring facility