NeRF-Inspired Neural Fields for 3D Super-resolution of Reacting-Flows
Abstract
The multiscale nature of turbulent reacting flows and the importance of fine-scale structures pose challenges for super-resolution. We investigate a 3D super-resolution framework inspired by Neural Radiance Fields (NeRF), called Neural Field Super-Resolution (NFSR), to reconstruct fine-resolution fields from coarsened reacting-flow fields. The proposed formulation adapts the continuous coordinate-conditioned representation of NeRF to physical flow-field reconstruction by conditioning spatial coordinates with local coarse-grid thermochemical information, while eliminating radiance prediction and volume rendering. Principal component analysis is used to provide a compact and low-dimensional representation of the thermochemical state, and a hierarchical extension, H-NFSR, progressively reconstructs the solution through intermediate resolution levels. The framework is evaluated using three-dimensional direct numerical simulation data for a turbulent premixed hydrogen-air flame, with an overall spatial coarsening factor of eight in each coordinate direction. The results demonstrate consistently better prediction measures and a higher ability to reconstruct fine-scale information than tricubic interpolation on the unseen test realization. The hierarchical formulation reduces visible coarse-cell imprinting. In contrast, biased sampling mainly benefits the most localized structures. Three-dimensional isosurface comparisons further show that H-NFSR recovers much of the small and intermediate features of the flow, including surface wrinkling and local flame curvature effects. These results highlight the potential of coordinate-conditioned neural fields as a flexible framework for three-dimensional coarse-to-fine reconstruction of reacting-flow simulations.
Related
- A Neural Implicit Flow Super-Resolution Approach for Reacting Flows
- Acoustic coherence-aware oriented 3D photoacoustic microscopy super-resolution with noise-suppression
- Towards isotropic super-resolution with dual reflection-transmission tomographic diffractive microscopy
- Universal Computational Super-Resolution Framework for Imaging of Quantum Dots.
- Synthetic-Aperture Super-Resolution Imaging via Spatial-Frequency Shift in Near-Field Diffraction
BibTeX
@misc{amarathunga2026nerfinspired,
title = {{NeRF-Inspired Neural Fields for 3D Super-resolution of Reacting-Flows}},
author = {Dakshaka Vishwajith Amarathunga and Shang-Ru Li and Tarek Echekki},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.27284},
eprint = {2610.27284},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.27284},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - UNPB
TI - NeRF-Inspired Neural Fields for 3D Super-resolution of Reacting-Flows
AU - Dakshaka Vishwajith Amarathunga
AU - Shang-Ru Li
AU - Tarek Echekki
PY - 2026
DA - 2026/10/19
PB - arXiv
DO - 10.48550/arxiv.2610.27284
UR - https://arxiv.org/abs/2610.27284
AB - The multiscale nature of turbulent reacting flows and the importance of fine-scale structures pose challenges for super-resolution. We investigate a 3D super-resolution framework inspired by Neural Radiance Fields (NeRF), called Neural Field Super-Resolution (NFSR), to reconstruct fine-resolution fields from coarsened reacting-flow fields. The proposed formulation adapts the continuous coordinate-conditioned representation of NeRF to physical flow-field reconstruction by conditioning spatial coordinates with local coarse-grid thermochemical information, while eliminating radiance prediction and volume rendering. Principal component analysis is used to provide a compact and low-dimensional representation of the thermochemical state, and a hierarchical extension, H-NFSR, progressively reconstructs the solution through intermediate resolution levels. The framework is evaluated using three-dimensional direct numerical simulation data for a turbulent premixed hydrogen-air flame, with an overall spatial coarsening factor of eight in each coordinate direction. The results demonstrate consistently better prediction measures and a higher ability to reconstruct fine-scale information than tricubic interpolation on the unseen test realization. The hierarchical formulation reduces visible coarse-cell imprinting. In contrast, biased sampling mainly benefits the most localized structures. Three-dimensional isosurface comparisons further show that H-NFSR recovers much of the small and intermediate features of the flow, including surface wrinkling and local flame curvature effects. These results highlight the potential of coordinate-conditioned neural fields as a flexible framework for three-dimensional coarse-to-fine reconstruction of reacting-flow simulations.
ER -