On the development of a Neural Radiance Field technique for tomographic reconstruction applied to flow diagnostics

Abstract

In this work, a novel neural implicit representation tomography algorithm based on Neural Radiance Fields is developed and demonstrated for 3D flow diagnostics. Neural Radiance Fields (NeRF) originate from the computer vision community that uses a machine learning approach to approximate a scene of interest as a continuous function using a neural network. The NeRF machine learning concept provides some advantages over traditional tomography methods, including i) a continuous approximation of the volume that removes the inherent limitation on volume resolution that is present in discretized representations, ii) a reduction in memory requirements for the volume prediction, and iii) an adaptable tomography framework that can include additional inputs, outputs, imaging models, and constraints. The method developed in this work, FluidNeRF, predicts the intensity per unit volume as a continuous function of 3D spatial (static) or 4D spatial-temporal (time-resolved) coordinates. FluidNeRF trains similarly to other algebraic reconstruction techniques, where the volume approximation is updated by comparing predicted and captured images of the volume. The image rendering technique of FluidNeRF employs an emission-based imaging model. Static and time-resolved FluidNeRF was evaluated using both i) a DNS-generated turbulent mixing jet and ii) an experimental dataset of a low-speed, smoke-entrained jet flow. The synthetic datasets systematically investigated the hyperparameters, camera configuration, and image noise regarding reconstruction quality. Static FluidNeRF was also compared to a traditional ART-based tomography model. The results show that i) FluidNeRF is a viable technique for tomography of flow diagnostics, ii) FluidNeRF produces comparable or superior reconstruction accuracy and is more robust to noise than traditional tomography methods, and iii) the method can scale to larger problems. Additionally, the results proved the FluidNeRF can be expanded to time-resolved reconstructions, which further compresses the volume representation and implicitly constrains the problem in time.