58476 - Toward Machine Learned Highly Reduced Kinetic Models for Natural Gas Combustion 

Accurate low dimension chemical kinetic models for methane are an essential component in the design of efficient gas turbine combustors. Kinetic models coupled to computational fluid dynamics (CFD) and chemical reactor networks (CRN) provide quick and efficient ways to test the effect of operating conditions, fuel composition and combustor design compared to physical experiments.

However, detailed chemical kinetic mechanisms which describe the pertinent combustion phenomena to excellent detail, are too computationally expensive for use in CFD or CRN. The computational cost of a mechanism scales with the number of species included in the transport and conservation equations. A mechanism must contain of the order of 20 species or less to enable efficient use of CFD and 50 species or less for use in CRN as a rapid development tool. We propose a novel three-step methodology for the production of compact kinetic models from a detailed mechanism that retains the high fidelity of the original mechanism.

In the first step, a compact kinetic model is obtained by removing all non-essential species using path flux analysis (PFA) from the NUIG18_17_C3 detailed mechanism containing 118 species. The reduction of the detailed mechanism is followed by two rounds of optimisation. The model is first optimised to the detailed model’s prediction of selected species (OH,H,CO and CH₄) profiles in perfectly stirred reactor (PSR) simulations and then re-optimised to the detailed model’s prediction of the laminar flame speed. The best performing model from this second optimisation is selected as the optimised compact model.

To perform the optimisation, an in-house Machine Learning for Optimisation of Chemical Kinetics (MLOCK) algorithm is developed. The MLOCK algorithm is a Monte Carlo-based method that perturbs all three Arrhenius parameters of selected reactions and assesses the suitability of the new parameters through an objective error function which quantifies the error between the compact model and the detailed model’s prediction of the optimisation target. A coarse grid-like search of the parameter bound space is performed followed by the identification of a “genetic seed” which constrains the algorithm to a promising region of the bound space.

This strategy is demonstrated through the production of a 19 species and a more compact 15 species methane-air combustion models. Both compact models are validated across a range of 0D and 1D calculations across both lean and rich conditions and shows good agreement to the parent detailed mechanism. The 15 species model is shown to outperform the current state-of-art models in both accuracy and range of conditions the model is valid over.