Development of Kinetic Mechanisms for Varied CN Controlled Fuels using Response Surface Surrogate Modeling

Data-driven chemical kinetic mechanisms have been developed for efficient simulations of jet fuel combustion. Due to the time-consuming simulation and optimization process, achieving many solutions that fit the given experimental data is unfeasible. Therefore, the extent to which the optimized solution is unique is unknown. Using machine learning to approximate the chemical kinetic simulation from the optimized reaction rate coefficients to the desired output, which in this case is ignition delay, provides the ability to quickly generate a large number of possible solutions with only a single initial large time investment of generating the training data for the model. The model training set was developed first through a uniform random search and then through an uncertainty-based adaptive search scheme to better cover the search space around the areas where the output of ignition delay was near the desired experimental ignition delay target. From this model, one hundred unique solutions were calculated using stochastic gradient descent. While these results matched the experimental ignition delay target within a defined tolerance, the solutions showed differing behavior when compared under thermodynamic conditions different from the targeted experimental data. When comparing species time histories from the set of solutions, greater variance in the solutions was found at lower temperatures, but as the temperature increased, this variance decreased. At higher pressures than the target data, ignition delay in the intermediate and low temperature regions showed more deviation from the mean solution than at the target pressure, while the high temperature ignition delay showed a similar level of deviation as the target data conditions. This approach can be a possible solution to develop chemical kinetic models of newly introduced sustainable aviation fuels with limited resources.