A Novel Machine Learning Based Lumping Approach for the Reduction of Large Kinetic Mechanisms for Plasma-Assisted Combustion Applications: Article No. 113252

The development of skeletal mechanisms has become essential for multi-dimensional simulations of plasma-assisted combustion (PAC). However, reduction tools developed for traditional combustion applications are not always applicable to PAC, due to the complex interplay between non-equilibrium plasma and combustion kinetics. Plasma direct relation graph with error propagation (P-DRGEP) is a recent plasma-specific reduction method developed in order to incorporate plasma energy branching in the reduction. In the first part of this work, the applicability of P-DRGEP to large kinetic mechanisms is investigated. A detailed isooctane/air plasma mechanism containing 2805 species and 18457 reactions is reduced to 415 species and 4716 reactions, keeping errors on ignition time within 3% for a wide range of initial conditions: from 750 K to 1200 K, 10 atm and equivalence ratios from 0.75 to 1.50. The second part focuses on isomer lumping, which is another reduction technique widely used in combustion. When applied to PAC, it is shown that the resulting lumped mechanism produces poor results. A novel plasma-specific isomer lumping strategy using machine learning is proposed instead. With the supervised algorithm of gradient boosting, predictive regression models are generated, which describe rate coefficients of lumped reactions adequately. These models are trained with simulation data. Leveraging this newly proposed lumping approach on the reduced mechanism, allows for an additional 28% reduction in the number of species and 19% reduction in the number of reactions. Two different versions are presented: in the first one the models are trained using one input feature (1D), while in the second one, two input features are selected (2D). The resulting lumped mechanism is shown to produce accurate predictions of PAC over the entire parameter space of interest, while significantly decreasing the computational time. Indicatively, with the 1D version the maximum error on ignition time in this range of conditions is 6%. The 2D approach produces even lower errors, which do not exceed 3%.