FluidNeRF: A Scalar-Field Reconstruction Technique for Flow Diagnostics using Neural Radiance Fields

View Video Presentation: https://doi.org/10.2514/6.2023-0412.vid In this work, a Neural Radiance Field approach (FluidNeRF) is investigated as a novel tomographic reconstruction technique for 3D flow diagnostics. Neural Radiance Fields (NeRF) is a machine learning concept that represents a 3D scene using a continuous function of 3D location, where the continuous function is approximated by a neural network. The neural network outputs the intensity of light per unit volume at a point in the volume. Image projections are rendered using emission-based line-of-sight integration by querying the neural network at sampling points along camera rays. The NeRF is updated using the loss between captured and rendered projections similar to other multi-camera tomography techniques. FluidNeRF is compared to the adaptive simultaneous algebraic reconstruction technique (ASART) for a CFD-generated volume of turbulent mixing jet. Image and volume based metrics are presented comparing both ASART and FluidNeRF to ground truth data. Preliminary results show that FluidNeRF outperforms ASART in most scenarios albeit with increased computational time. Potential benefits of FluidNeRF over ASART are discussed including improvements in fidelity, adaptability and scalability.