Development of a Neural Radiance Field Technique for Volumetric Molecular Tagging

Abstract

Tomographic reconstruction of three-dimensional flow fields from limited angular views is a fundamental challenge in experimental flow diagnostics. Conventional grid-based methods such as MART, SART, and RLD amplify noise and produce artifacts under the narrow angular sampling imposed by compact plenoptic imaging systems. In their discretized formulation, spatial resolution depends on the voxel grid spacing, the computational cost is high, and each time instant requires a separate reconstruction. Neural Radiance Fields provide a continuous, gridless alternative to these limitations, and this representation has been applied to flow diagnostics through FluidNeRF. FluidNeRF has not, however, been demonstrated for Molecular Tagging Velocimetry. Existing implementations rely on perspective camera models with wide angular coverage and on Fourier positional encoding, which requires a large network and long training times. Existing FluidNeRF implementations rely on perspective camera models with wide angular coverage and on Fourier positional encoding, which requires a large network and long training times. In this work, the FluidNeRF framework was adapted to the orthographic ray geometry of the Fourier Integral Microscope (FIMic) for the reconstruction of time-resolved scalar fields in MTV. Multiresolution hash encoding additionally replaced Fourier positional encoding, which reduced the required network size and thereby the computational cost of training.The framework was tested on two synthetic datasets generated using the calibrated FIMic imaging geometry: Poiseuille flow and turbulent channel flow. A systematic investigation of the encoding parameters and network size was performed to optimize the network and to enable time-resolved reconstruction of two time instants simultaneously. The optimized framework was then applied to experimental volumetric MTV measurements of a axisymmetric stagnation jet. The hash-encoded framework slightly exceeded the reconstruction fidelity of the Fourier-encoded implementation while reducing training time by more than an order of magnitude. Beamlet morphology and axial resolution were preserved throughout the volume. Applied to experimental data, a single network captured the temporal evolution of the tagged beamlets, and the reconstructions showed good agreement with Richardson–Lucy deconvolution. The results establish time-resolved neural implicit representations as a computationally efficient alternative to conventional grid-based tomography for plenoptic MTV.