Efficient two-dimensional scalar fields reconstruction of laminar flames from infrared hyperspectral measurements with a machine learning approach

Abstract

The latest hyperspectral measurements of combustion flames by Rhoby et al. (2014) provided extensive spatially and spectrally resolved information of flame radiation, which has been explored to retrieve two-dimensional, multi-scalar values of these flames with the conventional gradient-based optimization method. The drawback of that method is that the inverse radiation problem was solved through iterations with computationally intensive radiative heat transfer calculations and high-resolution wide-spectrum modeling, making the retrieving process very time-consuming. In the present study, we propose a machine learning based efficient inverse radiation model to retrieve two-dimensional temperature, CO2, H2O, and CO mole fractions of laminar flames from hyperspectral measurements. The model is trained with synthetic numerical data and is tested against previously made OH-laser absorption measurements and chemical equilibrium calculations for ethylene laminar flames with different equivalence ratios. The training data generation process, machine learning model architecture, model training, and validations are discussed in detail. Results have shown that the proposed machine learning based inverse radiation model is both accurate and efficient.