Machine learning driven chemiluminescence-based modelling of combustion parameters in premixed swirling NH3/H2 flames

This work develops predictive models for estimating equivalence ratio ( ϕ ), ammonia fraction (xNH 3 ) and noxious emissions (NOx) from the flames of turbulent premixed NH 3 /H 2 fuel blend stabilised using a tangential swirl burner. Bayesian Regularisation Artificial Neural Network (BR-ANN) is utilized to estimate both ϕ and xNH 3 with excited ratios NH∗/OH∗, violet/OH∗ and NH 2 ∗/NO 2 ∗ as inputs. NOx was predicted with NO∗, OH∗, NH∗, NH 2 ∗ and NO 2 ∗ as input variables. The coefficient of determination (R 2 ) was 0.98,0.95, 0.99,0.99 and 0.97 for the ϕ , xNH 3 , NO, NO 2 and N 2 O models, respectively. The models show better performance when compared to the conventional ratio-based method of inferring crucial combustion features. The developed models operate within the ranges of ammonia-hydrogen blend (0.55 ≤ xNH 3 ≤ 0.90), Reynolds numbers (4000 ≤ Re ≤ 7000), equivalence ratios (0.60 ≤ ϕ ≤ 1.40), room temperature and atmospheric pressure. The models have been explicitly presented in mathematical equations enabling easy deployment in a software. These models will serve as a crucial step towards the development of non-invasive sensors that will help designers easily determine location of interest, predict reaction zones formation according to ϕ and xNH 3 and advance approaches that abate NOx emissions. • Excited radicals used to predict ϕ ,xNH 3 and NOx in NH 3 /H 2 turbulent flames. • Bayesian Regularisation algorithm deployed to develop models. • The models performed better than the conventional ratio-based method. • First study which models are represented explicitly with mathematical equations. • Vital tool for developing non-invasive sensors for NOx free NH 3 /H 2 turbulent flames.