A semi-supervised learning technique for the region-wise simplification of chemical kinetics: Application to Hydrogen–Ammonia flames

High-fidelity simulations of complex reacting flows require accurate yet computationally affordable kinetic models. Heuristic simplification methods such as those based on Computational Singular Perturbation (CSP) are used to obtain skeletal kinetic mechanisms by removing unnecessary species/reactions. However, these mechanisms are expected to be accurate only in the proximity of the operating conditions at which the training dataset was generated. This paper proposes a method that uses CSP analysis and clustering techniques together to identify chemical dynamics common to large regions in the operating space and to find a skeletal mechanism representative of that region. The developed workflow is tested to derive skeletal kinetic models for 1D strained premixed H2 ∕NH3 counterflow flames. Starting from a comprehensive database spanning a wide range of equivalence and blending ratios, the algorithm autonomously partitions the operational space into chemically coherent regions. For each region, a representative skeletal mechanism is then generated. These region- wise models overcome the inherent inefficiency of large comprehensive mechanisms and the extrapolation limitations of local models delivering maximum target errors below 2.5% while retaining strictly between 18 and 24 species across all regions. Furthermore, we employ the CSP analysis to confirm a-posteriori that the data-driven clustering is consistent with the fundamental behaviors of the underlying chemical manifold, such as the complex nitrogen sub-mechanisms that limit the degree of reduction in ammonia-heavy regimes. Finally, the scalability of the approach is demonstrated on a larger hydrocarbon mechanism.