Machine learned compact kinetic model for liquid fuel combustion

A novel data-intensive methodology to produce a high fidelity, extremely-reduced “compact” kinetic model for a high boiling point complex liquid fuel is proposed and demonstrated. A five-component surrogate definition for the liquid fuel is developed that displays a high accuracy to the experimentally-derived combustion property targets. The calculations of the Lawrence Livermore National Lab diesel surrogate model containing 6476 species are used to serve as gas turbine industry-defined performance targets for this surrogate.

Acknowledging that the retention of a multi-component surrogate definition is a limitation on the size of the model, the surrogate fuel is consolidated into a single virtual molecule. Subsequently, the reaction mechanism is simplified by replacing high carbon number chemistry with a virtual scheme. This scheme links the virtual fuel molecule to low carbon number chemistry using four virtual species and forty-four virtual reactions, resulting in a reduction to 429 species in the model.

The Machine Learned Optimisation of Chemical Kinetics (MLOCK) algorithm is adapted to “compact” this model. Compaction is the over-reduction and optimisation of a kinetic model. Path flux analysis generates an overly-reduced model with 31 species that has a poor replication of the detailed model calculations. To address this, virtual reaction rate constants of important virtual reactions are numerically optimized to detailed model high temperature calculations. MLOCK systematically perturbs all three virtual Arrhenius reaction rate constant parameters to generate and evaluate numerous model candidates, refining the search space based on prior results, finding better models. A low temperature virtual reaction network, comprising one new virtual species and three new virtual reactions, is appended to the high temperature compact model. MLOCK is employed to reoptimize the model to calculations at low and intermediate temperatures.

The application of this methodology results in a 32-species compact model in ChemKin/Cantera format, which retains fidelities in the range of 76 to 92 % across a comprehensive range of gas-turbine relevant performance calculations for low, intermediate and high temperatures.