Machine learning applied to the retrieval of three-dimensional scalar fields of laminar flames from hyperspectral measurements

Abstract

Accurate measurement of three-dimensional temperature and species mole fraction fields for combustion systems provides comprehensively detailed information for optimizing combustion process and improving combustion efficiency. The state-of-art three-dimensional combustion diagnostic techniques for temperature and species mole fraction reconstructions, either laser-based or radiation imaging-based, require solving problems of huge matrices with iterative processes based on the multiple projection measurements of flame emission or absorption. These techniques are typically computationally intensive, with limited spatial resolution and can be hardly applied to retrieve three-dimensional temperature and multiple species mole fractions simultaneously. In the present study, we extended the machine learning methodology we previously proposed (Ren et al. 2021) for the reconstruction of two-dimensional temperature and mixture species mole fraction fields to three-dimensional for a group of non-axisymmetric flames with different equivalence ratios. The developed method demonstrates its excellent capability to retrieve three-dimensional temperature with CO2, H2O, and CO mole fractions simultaneously for these targeted flames. The accuracy of the machine learning reconstructions was found to be excellent, while computational effort was reduced by at least five orders of magnitude, as opposed to conventional gradient-based optimization methods.