Data-Driven Ammonia Combustion Simulation using Neural Ordinary Differential Equations

Detailed combustion mechanisms feature a complex system involving many species and elementary reactions. The complexity of the system usually makes the computational cost very expensive. Recently, Owoyele et al. (Energy and AI, 2022) developed the Neural Ordinary Differential Equations (NODE) for calculations of chemical reactions to accelerate the calculation of hydrogen combustion mechanisms considering 9 species and 21 elementary reactions. However, the application of NODE to chemical reactions was not capable of reproducing the ammonia combustion behavior well, which has much more complex combustion mechanisms involving 59 species and 356 elementary reactions (Okafor et al., Combustion and Flame, 2018). In the present work, the application of NODE is further extended for ammonia combustion mechanisms, and new normalization and training strategies have been proposed. The results show that the optimized model could reproduce the temporal evolution of the 0D ammonia combustion well for wide ranges of equivalence ratios and temperatures and give accurate ignition delay predictions. In addition, the model performance is further explored to investigate the generality to predict untrained data.