Hybrid physics-machine learning model for multispecies and temperature inference from FTIR spectra: Application to ammonia flames

Abstract

Fourier-transform infrared (FTIR) spectroscopy offers a powerful, non-intrusive diagnostic tool for in-situ measurements of temperature and species concentrations in combustion systems. However, in practical applications, FTIR spectra often suffer from low spectral resolution, strong band overlap, and significant variation in species concentration levels, making quantitative interpretation a challenging inverse problem. In this work, we present a hybrid physics-machine learning framework for inferring temperature, path length, and species mole fractions from FTIR emission spectra of ammonia flames. The model is trained on high-fidelity synthetic spectra generated via line-by-line radiative transfer using HITEMP/HITRAN spectroscopic databases. To address challenges of spectral overlap, minor-species detectability, and measurement noise, the architecture incorporates physics-based regularization and a self-supervised spectrum reconstruction module that enforces consistency with the radiative transfer equation. Our hybrid approach enables robust multi-target inference across species spanning several orders of magnitude in concentration. Compared to standard partial least squares (PLS) regression and ablated models, the proposed framework achieves superior accuracy and noise robustness while remaining compact and interpretable. Additionally, the co-trained reconstruction module exhibits effective denoising capabilities, highlighting the physical relevance of the learned spectral representation. This framework provides a foundation for practical, generalizable FTIR diagnostics and opens pathways toward spatially resolved inference in complex combustion environments.