An Investigation Into Neural Radiance Fields for Spray Reconstruction

Two-phase impinging injector sprays were reconstructed, in 3D, using the novel view synthesis method, Neural Radiance Fields (NeRF). This reconstruction technique utilizes a deep neural network to represent a 3D scene and is applied here to qualitatively visualize the sprays of two types of injector geometries imaged with an array of 18 cameras and pulsed LED illumination. Pentad and triplet injectors were used to produce sample data at four pressure drop values, 40, 100, 150, and 200 psi. NeRF shows significant potential as a 3D spray reconstruction method due to its ability to discern between areas dense with liquid and areas with little to no liquid revealing the asymmetric 3D topology of the different injector types. It also demonstrated the ability to reconstruct large individual droplets. The long-term potential of NeRF to be adapted for quantitative spray measurements is partially motivated by the use of an optical density parameter and Beer-Lambert law in the training of the network as a means to locate objects in a scene. Although the use of the optical density parameter in this work was contrived (i.e. used for convenience and not rigorously grounded in the physics of optics and sprays), volume integrated quantities of optical density revealed expected trends with the overall liquid volume increasing monotonically with pressure.