Assessing the potential of a chemiluminescence and machine learning-based method for the sensing of premixed ammonia–hydrogen–air turbulent flames

The potential of chemiluminescence to develop non-intrusive sensors for the monitoring and control of turbulent ammonia–hydrogen flames is here investigated experimentally. This study looks into the impact of equivalence ratio (0.35 ≤ ϕ ≤ 1.70), NH3 fuel fraction (0.55 ≤ XNHjavax.xml.bind.JAXBElement@4695dbaa ≤ 0.90) and Reynolds number (4000 ≤ Re ≤ 7000) on UV, visible and infrared chemiluminescence signatures and NOx emission of NH3/H2 turbulent flames within an atmospheric tangential swirl burner. Chemiluminescence spectroscopy is employed to provide detailed information about the excited species (e.g., NO∗, OH∗, NH∗, NH2∗, and NO2∗) in both in-flame and post-flame zones. Findings are compared to previous measurements in laminar flames and similar trends are observed. Many chemiluminescence intensity ratios are investigated but none are found to be potential surrogates of equivalence ratio and NH3 fuel fraction across all the conditions considered. Therefore, a more advanced method based on machine learning is used to infer equivalence ratio and NH3 fuel fraction from the chemiluminescence intensities of more than just two excited radicals at once. This method referred to as Gaussian Process Regression (GPR) is found to provide predictions of equivalence ratio and NH3 fuel fraction with an accuracy better than 0.1 and 0.02, respectively, across the whole range of conditions. GPR is also able to predict the measured NO, N2O and NO2 emissions using only measured chemiluminescence intensities, confirming the potential of chemiluminescence sensors coupled with a machine learning-based method for the monitoring and control of practical NH3/H2 flames.