arxiv: 2605.11454 · v1 · submitted 2026-05-12 · ⚛️ physics.flu-dyn

Recognition: 2 theorem links

· Lean Theorem

Neural Refractive Index Primitives for Flame Field Reconstruction Using Background-Oriented Schlieren

Jingxuan Li, Wei Hu, Xinyi Lu, Yue Zhang, Zheng Wang, Zizhou Liao

Pith reviewed 2026-05-13 02:14 UTC · model grok-4.3

classification ⚛️ physics.flu-dyn

keywords neural refractive index primitivesbackground-oriented schlierenflame field reconstructiontomographymultilayer perceptronhash encodingcombustion diagnostics

0 comments

The pith

Compact neural networks reconstruct continuous three-dimensional refractive index fields in flames from schlieren images.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that the refractive index field itself can act as the sole output of a compact multilayer perceptron for background-oriented schlieren tomography. Adding multiresolution hash encoding captures fine details, automatic discrete gradient losses enforce physical consistency, and a three-dimensional mask focuses the solution on the flame region. Tests on combustion phantoms and real flame recordings show recovery of both broad structures and small-scale turbulence with good noise tolerance and faster convergence than voxel grids or frequency-based alternatives. A reader would care because this supplies a practical route to detailed three-dimensional combustion diagnostics without explicit discretization.

Core claim

By treating the refractive index field as the sole neural primitive inside a compact multilayer perceptron and integrating multiresolution hash encoding, automatic-discrete gradient losses, and a three-dimensional mask, the method produces fast-converging, high-resolution, spatially coherent reconstructions of flame fields from background-oriented schlieren measurements.

What carries the argument

Refractive index field modeled as the sole neural primitive in a compact multilayer perceptron equipped with multiresolution hash encoding

If this is right

Accurate recovery of both large-scale flame structures and fine-scale turbulence on numerical phantoms and real data
Strong robustness to noise in the captured schlieren images
Clear accuracy and coherence gains over frequency-encoding-based and voxel-based methods
Fast convergence to high-resolution solutions during optimization

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Continuous neural field representations could reduce staircasing artifacts when the same tomography setup is applied to other transparent fluid flows.
If paired with high-speed cameras, the approach might support time-resolved reconstruction of unsteady flames without additional regularization.

Load-bearing premise

The refractive index distribution inside a flame can be faithfully captured by the output of a compact multilayer perceptron without introducing non-physical artifacts or losing accuracy in the inverse tomography problem.

What would settle it

Apply the method to a new numerical phantom whose exact refractive index volume is known, then measure whether the recovered field matches the ground truth at fine turbulence scales within the error bounds reported for the original tests.

Figures

Figures reproduced from arXiv: 2605.11454 by Jingxuan Li, Wei Hu, Xinyi Lu, Yue Zhang, Zheng Wang, Zizhou Liao.

**Figure 2.** Figure 2: The overview of the multi-resolution hash encoding scheme functioning in 3D coordinate [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Sampling strategy of multiscale occupancy grids. This is achieved using multiscale occupancy grids that distinguish empty from non-empty regions, as shown in Fig.3. Each voxel stores a single-bit occupancy value, and samples below a threshold are skipped. These grids are independent of the neural encoding and are updated during training based on the absolute refractive-index magnitude. The refractive in… view at source ↗

**Figure 4.** Figure 4: 2D masks and 3D mask. However, a 2D mask only constrains projection-space data and cannot localize occlusions in the 3D volume, allowing spurious refractive-index values to persist. In contrast, a 3D mask directly constrains the volume, enabling rays to skip invalid voxels, reducing computation and suppressing noise. Nicolas et al. applied a volumetric support constraint in direct BOS 3D density inversio… view at source ↗

**Figure 5.** Figure 5: The isosurfaces of the true phantoms’ temperature. In the numerical validation, five flame phantoms with different temperature distributions were used, including four high-fidelity cases from the BLASTnet dataset [18]- [19] and one mathematical phantom [6]. Since the original BLASTnet data contain billions of voxels, a single snapshot from each case was selected and downsampled for efficient reconstructi… view at source ↗

**Figure 6.** Figure 6: Comparison of reconstruction quality obtained with different encoding and gradient [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗

**Figure 7.** Figure 7: Approximating a flame temperature of resolution 280 [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗

**Figure 8.** Figure 8: The effect of the MLP size (a) and the point sampling size (b) on the reconstruction quality. To evaluate the influence of network architecture and sampling on reconstruction fidelity, we performed a parameter sensitivity study using RMSE of the reconstructed temperature field [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗

**Figure 10.** Figure 10: Temperature reconstruction results for Phantom 2 using three methods. Left: isosurfaces at values indicated by the colorbar. Right: temperature slices at different ylocations for hash encoding, frequency encoding, voxel-based CGLS, and the ground truth [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗

**Figure 11.** Figure 11: Temperature results for phantom 3. The proposed hash-encoding method shows consistently strong performance across Phantoms 2, 3, and 5, achieving the best over13 [PITH_FULL_IMAGE:figures/full_fig_p013_11.png] view at source ↗

**Figure 13.** Figure 13: Temperature results for phantom 5. Finally, we compare reconstruction quality with and without the mask. The mask is shown in [PITH_FULL_IMAGE:figures/full_fig_p014_13.png] view at source ↗

**Figure 12.** Figure 12: Temperature results for phantom 4 [PITH_FULL_IMAGE:figures/full_fig_p014_12.png] view at source ↗

**Figure 14.** Figure 14: 2D (left) and 3D (right) masks of for phantom 1. 4.2 Experiment test [PITH_FULL_IMAGE:figures/full_fig_p015_14.png] view at source ↗

**Figure 16.** Figure 16: 3D and 2D slices of the reconstructed double flame temperature field using different [PITH_FULL_IMAGE:figures/full_fig_p016_16.png] view at source ↗

**Figure 17.** Figure 17: 3D and 2D slices of the reconstructed diffusion flame temperature field using different [PITH_FULL_IMAGE:figures/full_fig_p016_17.png] view at source ↗

read the original abstract

An improved neural refractive-index-primitive method for background-oriented schlieren tomography is presented, enabling continuous three-dimensional reconstruction of refractive-index fields using a compact multilayer perceptron. The method adopts the refractive-index field as the sole neural primitive and integrates multiresolution hash encoding, automatic-discrete gradient losses, and a three-dimensional mask to enable fast convergence and high-resolution, spatially coherent reconstructions. Tests on numerical combustion phantoms and real flame data demonstrate accurate recovery of both large-scale structures and fine-scale turbulence, strong robustness to noise, and clear advantages over frequency-encoding-based and voxel-based reconstruction methods.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This paper combines refractive-index neural primitives with hash encoding and gradient losses for BOS flame tomography, showing decent results on phantoms and real data but thin quantitative backing.

read the letter

The core idea here is to represent the refractive index field directly with a compact MLP plus multiresolution hash encoding, then add automatic discrete gradient losses and a 3D mask to stabilize the inverse problem from background-oriented schlieren images. That specific bundle is the main incremental step over earlier neural-field or voxel work in this niche. The tests on combustion phantoms recover both large structures and finer turbulence, and the real-flame examples look plausible with claimed noise robustness and visible edges over frequency-encoding and pure voxel baselines. Those are the parts that actually move the needle for someone doing high-resolution combustion diagnostics. The method appears straightforward to implement once the primitives are set, and the choice to stay with refractive index rather than derived quantities avoids some common circularity traps in tomography. On the downside, the abstract gives no error tables, no RMSE or PSNR numbers against ground truth, and no ablation that isolates how much the hash encoding or the gradient term actually buys versus just more careful training. Real-data evaluation stays visual, which is common but limits how strongly we can claim superiority. The 3D mask helps coherence, yet without details on how it is constructed or its sensitivity, it is hard to judge generalization. This work is aimed at fluid dynamicists and combustion experimentalists who already use BOS and want a continuous 3D field without heavy voxel grids. A reader already familiar with neural radiance fields or hash-grid papers will see the tweaks quickly and can judge whether the gains justify the added complexity. It is coherent on its own terms and engages the relevant prior literature, so it clears the bar for serious peer review even if the quantitative case needs tightening in revision.

Referee Report

2 major / 2 minor

Summary. The paper presents an improved neural refractive-index-primitive method for background-oriented schlieren (BOS) tomography. It uses a compact multilayer perceptron as the sole neural primitive for the refractive index field, combined with multiresolution hash encoding, automatic-discrete gradient losses, and a 3D mask. This enables continuous 3D reconstruction of flame refractive-index fields. Validation is performed on numerical combustion phantoms (with ground truth) and real flame data, claiming accurate recovery of large-scale structures and fine-scale turbulence, noise robustness, and advantages over frequency-encoding-based and voxel-based methods.

Significance. If the central claims hold with quantitative support, the work could meaningfully advance non-intrusive 3D diagnostics in combustion and fluid dynamics by providing a continuous, high-resolution representation that avoids discretization artifacts common in traditional tomography. The integration of hash encoding and gradient losses for physical inverse problems is a timely contribution, but its impact depends on rigorous error metrics and ablation studies that are currently absent from the presented material.

major comments (2)

[Abstract and Results] Abstract and Results section: The central claims of 'accurate recovery' and 'clear advantages' over baselines are stated without any quantitative error metrics (e.g., RMSE, PSNR, or L2 norms against ground truth on phantoms), ablation tables, or statistical comparisons. This leaves the superiority and robustness assertions unverified and load-bearing for the paper's contribution.
[Method] Method description (around the neural primitive and losses): The automatic-discrete gradient losses and 3D mask are presented as enabling physical consistency and fast convergence, but no derivation, hyperparameter sensitivity analysis, or proof of non-introduction of non-physical artifacts is provided. This directly affects the weakest assumption that the MLP faithfully represents flame RI fields for the inverse problem.

minor comments (2)

Notation for the refractive index field and encoding should be standardized across equations and figures for clarity.
Figure captions for phantom and real-data reconstructions would benefit from explicit scale bars and error maps to aid visual assessment.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. The comments highlight opportunities to strengthen the quantitative support for our claims and to provide more explicit details on the method components. We address each major comment below and will revise the manuscript to incorporate the suggested improvements, including additional metrics, derivations, and analyses.

read point-by-point responses

Referee: [Abstract and Results] Abstract and Results section: The central claims of 'accurate recovery' and 'clear advantages' over baselines are stated without any quantitative error metrics (e.g., RMSE, PSNR, or L2 norms against ground truth on phantoms), ablation tables, or statistical comparisons. This leaves the superiority and robustness assertions unverified and load-bearing for the paper's contribution.

Authors: We agree that the current presentation relies primarily on qualitative visual comparisons between our method, frequency-encoding baselines, and voxel-based approaches. While these figures demonstrate clearer recovery of large-scale structures and fine-scale turbulence on both phantoms and real flames, we acknowledge the absence of explicit numerical error metrics and ablation tables. In the revised manuscript, we will add a quantitative comparison table reporting RMSE, PSNR, and L2 norms on the numerical combustion phantoms (with ground truth available), as well as ablation studies isolating the contributions of multiresolution hash encoding and the automatic-discrete gradient losses. This will provide the statistical verification needed to support the claims of accuracy and advantages. revision: yes
Referee: [Method] Method description (around the neural primitive and losses): The automatic-discrete gradient losses and 3D mask are presented as enabling physical consistency and fast convergence, but no derivation, hyperparameter sensitivity analysis, or proof of non-introduction of non-physical artifacts is provided. This directly affects the weakest assumption that the MLP faithfully represents flame RI fields for the inverse problem.

Authors: The automatic-discrete gradient losses are constructed by applying a discrete finite-difference approximation to the refractive-index field sampled from the hash-encoded MLP, with the resulting gradient term incorporated into the BOS forward model loss to promote physical consistency and smoothness. The 3D mask is a simple spatial bounding volume derived from the experimental setup to restrict optimization to the flame region. We recognize that the manuscript would benefit from a more explicit step-by-step derivation of these terms and supporting analyses. In the revision, we will expand Section 3 with a detailed derivation of the gradient loss, include hyperparameter sensitivity results (e.g., varying the gradient-loss weight and its effect on convergence and reconstruction quality), and add discussion plus phantom evidence showing that the approach avoids non-physical artifacts, as the recovered fields align with expected combustion structures without introducing spurious oscillations or discontinuities. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper proposes a neural representation for refractive-index fields in BOS tomography using a compact MLP with multiresolution hash encoding, automatic-discrete gradient losses, and a 3D mask. This is an empirical modeling choice tested on numerical phantoms (with ground truth) and real flame data, with quantitative comparisons to frequency-encoding and voxel baselines. No load-bearing step reduces by construction to self-definition, fitted inputs renamed as predictions, or self-citation chains; the representational power is validated externally rather than assumed tautologically. The derivation remains self-contained as a proposed architecture whose accuracy is assessed against independent benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that a neural network can accurately invert the BOS tomography problem for continuous refractive index fields; no explicit free parameters or invented physical entities are stated in the abstract.

axioms (1)

domain assumption A compact multilayer perceptron with multiresolution hash encoding can represent the continuous 3D refractive index field of a flame with sufficient fidelity for tomography reconstruction.
This is the core modeling choice that enables the continuous representation and fast convergence described.

pith-pipeline@v0.9.0 · 5405 in / 1121 out tokens · 54422 ms · 2026-05-13T02:14:45.618234+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
The method adopts the refractive-index field as the sole neural primitive and integrates multiresolution hash encoding, automatic-discrete gradient losses, and a three-dimensional mask to enable fast convergence and high-resolution, spatially coherent reconstructions.
IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean reality_from_one_distinction unclear
Tests on numerical combustion phantoms and real flame data demonstrate accurate recovery of both large-scale structures and fine-scale turbulence

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages

[1]

Raffel, Background-oriented schlieren techniques, Experiments in Fluids 56 (2015) 60

M. Raffel, Background-oriented schlieren techniques, Experiments in Fluids 56 (2015) 60

work page 2015
[2]

B. E. Schmidt, B. F. Bathel, S. J. Grauer, M. J. Hargather, J. T. Heineck, M. Raffel, Twenty-five years of background-oriented schlieren: Advances and novel applica- tions, in: AIAA SciTech Forum, 2025, p. 1627

work page 2025
[3]

S. J. Grauer, A. Unterberger, A. Rit- tler, K. J. Daun, A. M. Kempf, K. Mohri, Instantaneous 3d flame imaging by background-oriented schlieren tomog- raphy, Combustion and Flame 196 (2018) 394–405

work page 2018
[4]

F. J. W. A. Martins, C. T. Foo, A. Un- terberger, et al., Analyzing 3d fields of re- fractive index, emission and temperature in spray-flame nanoparticle synthesis via tomographic imaging, Applications in En- ergy and Combustion Science 16 (2023) 100213

work page 2023
[5]

Zheng, J

Y. Zheng, J. Chen, Y. He, M. Li, X. Chen, S. Xu, X. Wen, Y. Liu, W. Cai, Investi- gation on heat exchange of iron particle combustion based on simultaneous multi- physics field measurements, Combustion and Flame 279 (2025) 114309

work page 2025
[6]

H. Cai, Y. Song, Y. Ji, Z. Li, A. He, Di- rect background-oriented schlieren tomog- raphy using radial basis functions, Optics Express 30 (2022) 19100–19120

work page 2022
[7]

H. C. Liu, J. Q. Huang, L. Li, W. W. Cai, Volumetric imaging of flame refractive in- dex, density, and temperature using bos tomography, Science China Technological Sciences 64 (2021) 1–12

work page 2021
[8]

Gordon, R

R. Gordon, R. Bender, G. T. Her- man, Algebraic reconstruction techniques for three-dimensional imaging, Journal of 17 Theoretical Biology 29 (3) (1970) 471– 481

work page 1970
[9]

Atcheson, I

B. Atcheson, I. Ihrke, W. Heidrich, et al., Time-resolved 3d capture of non- stationary gas flows, ACM Transactions on Graphics 27 (2008)

work page 2008
[10]

Nicolas, V

F. Nicolas, V. Todoroff, A. Plyer, et al., A direct approach for instantaneous 3d den- sity field reconstruction from bos, Exper- iments in Fluids 57 (2016)

work page 2016
[11]

J. P. Molnar, S. J. Grauer, Algorithm for time-resolved bos tomography, in: AIAA SciTech Forum, 2025

work page 2025
[12]

S. Cai, Z. Wang, F. Fuest, Y. J. Jeon, C. Gray, G. E. Karniadakis, Inferring 3d velocity and pressure fields from bos via pinns, Journal of Fluid Mechanics 915 (2021) A102

work page 2021
[13]

J. P. Molnar, S. J. Grauer, O. Leon, et al., Physics-informed bos of turbulent jets, in: AIAA SciTech Forum, 2023

work page 2023
[14]

Y. He, Y. Zheng, S. Xu, C. Liu, D. Peng, Y. Liu, W. Cai, Neural refractive index field for bos tomography, Physics of Flu- ids 37 (2025) 017143

work page 2025
[15]

Mildenhall, P

B. Mildenhall, P. P. Srinivasan, M. Tan- cik, J. T. Barron, R. Ramamoorthi, R. Ng, Nerf: Representing scenes as neural radi- ance fields, Communications of the ACM 65 (2022) 99–106

work page 2022
[16]

Muller, A

T. Muller, A. Evans, C. Schied, A. Keller, Instant neural graphics primitives with multiresolution hash encoding, ACM Transactions on Graphics 41 (2022)

work page 2022
[17]

J. Li, X. Meng, Y. Xiong, et al., Neural de- flection field for sparse-view bos, Physics of Fluids 36 (2024) 121701

work page 2024
[18]

Brouzet, M

D. Brouzet, M. Talei, M. J. Brear, B. Cuenot, Impact of chemical modelling on flame acoustics, Journal of Fluid Me- chanics 915 (2021) A3

work page 2021
[19]

Sharma, W

P. Sharma, W. T. Chung, M. Ihme, Physics-informed ml for reacting flows, Combustion and Flame 277 (2025) 114190

work page 2025
[20]

X. Qin, X. Xiao, I. K. Puri, S. K. Ag- garwal, Effect of composition on tempera- ture reconstruction from refractive index, Combustion and Flame 128 (2002) 1–12

work page 2002
[21]

Biguri, M

A. Biguri, M. Dosanjh, S. Hancock, et al., Tigre: a matlab-gpu toolbox for cbct re- construction, Biomedical Physics & Engi- neering Express 2 (2016) 055010

work page 2016
[22]

J. Tang, X. Chen, J. Wang, et al., Compressible-composable nerf, NeurIPS 35 (2022) 14798–14809. 18

work page 2022