ThermoSight: From Appearance to Internal Thermofluid Fields via Physics-Guided Neural Reconstruction

Yuhao Fang, Chenglin Lu, Hefei Ling, Chun LouarXiv 2026 · arXiv

TL;DR

ThermoSight is a physics-guided neural reconstruction framework that goes beyond just "looking right" to "being right" by inferring internal temperature and soot fields of flames from a few RGB images. By integrating a differentiable spectral model of light emission and self-absorption, it successfully breaks the inherent ambiguity between temperature and soot concentration, achieving high physical accuracy where standard methods like NeRF or 3D Gaussian Splatting only recover surface appearance.

What Problem Does This Paper Address?

In the world of neural scene reconstruction, success is usually measured by how well a model can synthesize a new view (visual consistency). However, for scientific applications like combustion optimization or fire safety, we need to know what’s happening inside the volume—specifically the 3D distribution of temperature and soot.

This is notoriously difficult because of the temperature–soot ambiguity. A bright pixel in a flame image could be caused by a small amount of very hot soot or a large amount of cooler soot. Furthermore, soot doesn't just emit light; it also absorbs light coming from deeper within the flame. Standard appearance-based methods like NeRF or 3D Gaussian Splatting (3DGS) can match the RGB pixels perfectly without ever learning the true underlying physical states.

Comparison of appearance-only reconstruction vs. physical field reconstruction.

Methods

ThermoSight represents temperature ($T$) and soot volume fraction ($f_v$) as continuous neural fields using a multiresolution hash encoder and a shared MLP. The core innovation is the differentiable spectral image formation model that links these physical fields to the recorded RGB values.

Key components include:

  • Physics-Based Rendering: It uses Planck’s law to model thermal emission and accounts for wavelength-dependent self-absorption (soot attenuating its own light).
  • Spectral Integration: Instead of treating RGB as three independent channels, the model integrates the full spectrum against calibrated camera response functions. This forces the internal fields to be consistent across all color channels simultaneously.
  • Multi-Objective Optimization: The loss function combines intensity fitting with chromaticity consistency (matching color proportions), observation-derived spatial constraints (using masks to define the flame boundary), and soot regularization to prevent non-physical "peaks" in concentration.
Figure 2. Overview of ThermoSight. Calibrated multi-view images supervise continuous temperature and soot fields through differen- tiable spectral rendering and four complementary objectives. The ray-update sketch uses a thin-segment approximation; reconstruction uses the analytic solution in Eq. (6).

Datasets

The authors evaluated ThermoSight using two primary sources:

  1. Synthetic Methane-Flame Data: Generated from high-fidelity Computational Fluid Dynamics (CFD) simulations. This provided a "ground truth" for internal temperature and soot fields that is impossible to measure directly in the real world.
  2. Measured Propane-Flame Data: The authors built a custom experimental platform featuring a propane diffusion burner and a ring of ten calibrated cameras capturing synchronized 12-bit RAW images.
Figure 4. Experimental acquisition platform. Multi-camera ac- quisition of a propane diffusion flame and the associated calibra- tion equipment.

Key Insights and Ideas

  • Appearance $\neq$ Physics: The paper demonstrates that a model can achieve near-perfect RGB reconstruction (high PSNR) while having massive errors in internal temperature. Visual fidelity is a necessary but insufficient condition for scientific imaging.
  • Wavelength Matters: By modeling how soot absorbs light differently at different wavelengths, the authors found they could significantly reduce the temperature-soot ambiguity. A three-channel (RGB) approximation of light transport is much less accurate than a full wavelength-resolved model.
  • Neural Fields as Regularizers: Compared to traditional voxel-based tomography, the continuous nature of neural fields provides a natural spatial coupling that helps "fill in the gaps" when camera views are limited, leading to smoother and more physically plausible reconstructions.

Key Results

  • Accuracy Gains: In synthetic tests, ThermoSight reduced the temperature Root Mean Square Error (RMSE) in hot, sooty regions from 234 K (using a simplified 3-channel model) to 112 K (using the full spectral model), despite both models having nearly identical visual accuracy (~65.2 dB PSNR).
  • Baseline Comparison: ThermoSight outperformed traditional tomographic methods like SIRT and ASD-POCS, which produced fragmented fields or artifacts because they neglected self-absorption.
  • View Efficiency: The system remains robust even with limited data, though the authors recommend at least six views for reliable internal reconstruction. At four views, "field leakage" begins to distort the boundaries.
Qualitative comparison on synthetic data showing ThermoSight's superior recovery of internal fields compared to baselines.

Limitations and Future Work

While powerful, ThermoSight relies heavily on accurate radiometric calibration. If the camera's spectral response or the soot's optical properties (the absorption function $E(\lambda)$) are incorrectly characterized, the physical accuracy of the reconstruction degrades significantly.

The authors suggest several promising directions for future research:

  • Multispectral Expansion: Moving into the infrared spectrum to better capture cooler regions of the flow.
  • Gas-Phase Imaging: Extending the model to recover gas concentrations (like $CO_2$ and $H_2O$) using their specific emission bands.
  • 4D Reconstruction: Incorporating temporal and fluid-dynamic constraints (e.g., the Navier-Stokes equations) to reconstruct evolving flames over time.

Summary updated

Abstract

Most neural reconstruction methods focus on reproducing scene appearance, yet inferring physical states from visual observations is essential for machine understanding of the physical world. Reconstructing the internal temperature and soot fields of radiating thermofluids such as flames is challenging yet essential for improving energy efficiency and assessing heat exposure and fire spread. In flames, temperature and soot concentration jointly govern visible emission, and soot self-absorption alters the light that reaches each camera. With limited views, changes in the two fields can offset one another in the recorded RGB images, making temperature and soot difficult to disentangle. ThermoSight, a physics-guided neural framework for scientific imaging, jointly reconstructs continuous three-dimensional temperature and soot volume fraction fields from limited calibrated flame images without field-level ground truth during optimization. A differentiable spectral model links these fields to measured RGB through thermal emission, self-absorption, and calibrated camera responses. Temperature-dependent emission spectra and wavelength-dependent attenuation constrain shared fields across RGB channels, reducing temperature-soot ambiguity. Chromaticity consistency, observation-derived spatial constraints, and soot regularization complement intensity fitting. In a synthetic ablation, wavelength-resolved attenuation reduces temperature RMSE in the hot, sooty region from 234 to 112 K relative to a three-channel approximation, with nearly identical RGB PSNR (65.23 vs. 65.24 dB). Measured data yield coherent fields and competitive RGB reconstructions. These results demonstrate that physics-guided neural fields recover interpretable internal physical quantities beyond visually consistent appearance.
BibTeX
@misc{fang2026thermosight,
  author = {Yuhao Fang and Chenglin Lu and Hefei Ling and Chun Lou},
  title = {{ThermoSight: From Appearance to Internal Thermofluid Fields via Physics-Guided Neural Reconstruction}},
  year = {2026},
  eprint = {2611.20786},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.20786},
}
Figure