A data-driven sparse learning approach to reduce chemical reaction mechanisms

Reducing detailed chemical reaction mechanisms is a crucial strategy for mitigating the computational cost of reacting flow simulations. In this work, we propose a novel sparse learning (SL) approach that leverages reaction sparsity to systematically identify influential reactions for mechanism reduction. Specifically, the SL method learns an optimized weight vector to rank reaction importance, enabling the construction of compact reduced mechanisms by retaining species involved in the most influential reactions. The approach is extensively validated against fundamental combustion properties and turbulence-chemistry interactions across various hydrocarbon fuel/air systems. The results demonstrate that the SL-based reduced mechanisms accurately predict ignition delay times, laminar flame speeds, species mole fractions, and turbulence-chemistry interactions over a broad range of operating conditions. Furthermore, comparative analysis with existing reduction methods shows that the SL method yields more compact mechanisms while maintaining similar accuracy levels, particularly for large-scale mechanisms with extensive species and reactions. These findings highlight the potential of SL as an effective tool for developing reduced chemical mechanisms with improved efficiency and scalability. Novelty and Significance Statement The novelty of this work lies in the development of a sparse learning (SL) approach for chemical mechanism reduction, which systematically explores reaction sparsity by identifying influential reactions through statistically learned weight criteria. This method enables the construction of highly compact reduced mechanisms while preserving predictive accuracy. Comparative assessments demonstrate that SL outperforms existing reduction techniques, such as DRGEP and DRGEPSA, by yielding mechanisms with fewer species under the same error constraints. Moreover, SL achieves more extensive reductions than state-of-the-art methods while maintaining comparable maximum relative errors. This work introduces a novel data-driven strategy for efficient mechanism reduction, offering significant potential for advancing computational combustion modeling.