Machine learning applied to retrieval of temperature and concentration distributions from infrared emission measurements

Tao Ren, Michael F. Modest, Alexander Fateev, Gavin Sutton, Weijie Zhao, Florin RusuApplied Energy 2019 · DOI

Abstract

Researchers have developed a machine learning framework using Multi-Layer Perceptron (MLP) neural networks to instantly retrieve temperature and gas concentration profiles from infrared emission spectra. By training on synthetic data generated from high-fidelity physics databases, the model bypasses the slow, error-prone iterative solvers traditionally used in combustion diagnostics, achieving millisecond-level retrieval with high experimental accuracy.

Introduction

In high-temperature combustion environments, such as jet engines or industrial furnaces, monitoring the internal distribution of temperature and gas concentrations (CO₂, H₂O, CO) is critical for optimizing efficiency and controlling emissions. However, intrusive probes are often impractical as they can melt or disrupt the flow. Consequently, researchers rely on inverse radiation techniques: measuring the infrared light emitted by the flame and working backward to determine the conditions that produced it.

Schematic of the line-of-sight infrared emission measurement setup.

The traditional approach uses gradient-based optimization to iteratively match simulated spectra to measured data. This process is computationally expensive, sensitive to initial guesses, and often fails to find a unique solution because the problem is "ill-posed"—multiple combinations of temperature and gas can produce similar spectral patterns.

Methodology

The authors propose a "global" inverse model based on a Multi-Layer Perceptron (MLP). Instead of solving physics equations for every new measurement, they use physics-based simulations to train a neural network to perform the retrieval in a single step.

  1. Forward Modeling: The HITEMP 2010 spectroscopic database was used to simulate infrared light transport through various gas mixtures.
  2. Training Data: 10,000 synthetic spectra were generated for different profile types (flat, single-peak, and double-peak) across path lengths ranging from 10 cm to 10 m.
  3. Neural Network Architecture: The MLP takes spectral intensities as input and outputs temperature and concentration values for 21 discrete points along the line-of-sight.
  4. Spectral Selection: The study focused on two infrared windows (1800–2500 cm⁻¹ and 3000–4200 cm⁻¹) where combustion gases exhibit strong, overlapping emission bands.
Comparison of simulated infrared emission spectra at various resolutions.

Key Insights and Results

  • Real-Time Performance: Once trained, the MLP retrieves a full profile in approximately 10 milliseconds, enabling potential real-time control of combustion systems.
  • Optimal Spectral Windows: The 1800–2500 cm⁻¹ interval, which includes the CO₂ 4.3 µm band, proved significantly more accurate than higher-wavenumber windows due to a better signal-to-noise ratio.
  • Optical Thickness Advantage: Counter-intuitively, the model performed better on longer gas paths (1 m vs. 10 cm). In "optically thick" paths, absorption creates unique spectral signatures for different locations, which the MLP exploits to improve spatial resolution.
  • Experimental Validation: The model was tested against real data from the NPL standard flame. Predicted temperatures were in excellent agreement with independent Rayleigh scattering measurements, with mean differences as low as 0.19% to 2.08%.

Conclusion

This research demonstrates that neural networks trained entirely on simulated data can accurately interpret physical measurements from FTIR spectrometers, provided the training includes realistic noise. While currently limited to 1D line-of-sight measurements, future work aims to adapt this framework for 3D tomographic reconstructions and include additional species such as soot or NOₓ.

Summary updated

BibTeX
@inproceedings{ren2019machine,
  author = {Tao Ren and Michael F. Modest and Alexander Fateev and Gavin Sutton and Weijie Zhao and Florin Rusu},
  title = {{Machine learning applied to retrieval of temperature and concentration distributions from infrared emission measurements}},
  booktitle = {Applied Energy},
  year = {2019},
  doi = {10.1016/j.apenergy.2019.113448},
}
Figure