Machine learning-based multi-objective optimization of NH3/H2 combustion under H2O/N2 dilution with consideration of flame instability

Utilizing engine exhaust waste heat to drive ammonia (NH3) cracking is an effective measure to improve thermal efficiency, while exhaust gas recirculation (EGR) effectively improves emissions. However, efficient and clean combustion strategies for NH3/H2 with H2O/N2 dilution remain to be explored. Accordingly, this study developed an integrated framework combining machine learning (ML) models, multi-objective optimization, and instability analysis for NH3/H2/H2O/N2 mixtures. Firstly, ML models were trained to predict seven parameters, including laminar burning velocity (LBV), pollutant emissions (NO, NO2, N2O), and instability parameters, namely flame thickness (δ), thermal expansion ratio (σ), and effective Lewis number (Leeff). The trained models achieved an overall prediction accuracy exceeding 95%. Then, the ML models were coupled with the genetic algorithm to improve thermal efficiency and emissions under local and global optimization. Finally, the Markstein length (Lb), evaluated from δ, σ, and Leeff, was used to identify the most stable case. Global optimization across 298-500 K, 1-20 atm, and ϕ = 0.4-1.0 identified two combustion strategies: stoichiometric and fuel-lean modes. In both modes, the most stable cases show consistently positive Lb, indicating strong resistance to flame stretch. The stoichiometric optimum is obtained at 497 K and 15.3 atm with 42% NH3, 33% H2, and 25% N2, achieving an LBV of 20.7 cm/s with NOx concentration of 3699 ppm. In contrast, the fuel-lean (ϕ = 0.57) optimum occurs at 425 K and 1.4 atm with 43% NH3, 27% H2, 15% H2O and 15% N2, resulting in an LBV of 8.4 cm/s and an NOx concentration of 4275 ppm.