Deep Neural Network Modeling of a Turbulent Jet Ignition System

As a means to address the imminent and severe threat of global climate change, the emissions of the transportation sector must be addressed in the near term. Harnessing limited computational resources in an efficient manner is paramount to improving our existing internal combustion (IC) systems; however, low fidelity models have been shown to have difficulty accurately representing a highly turbulent combustor, such as in Turbulent Jet Ignition (TJI). TJI replaces the spark ignition source of a conventional IC system with a small combustion chamber that injects hot ignition products into the main combustion chamber. This process reduces the probability of engine misfires, particularly at the ultra-lean condition, improving the thermal efficiency of existing IC systems. This study develops a framework for the training and validation of a Deep Neural Network (DNN) model for the prediction of chemical reaction rates as a means to address the poor performance of existing low-fidelity combustion models for TJI. A high-fidelity combustion simulation was performed under the large eddy simulation (LES) equations with a transported probability density function (PDF) combustion model, and validated against experimental results gathered from a single-cycle TJI rig performed by a research group at Purdue University. This simulation data was used to train a DNN model to predict the turbulent reaction rates for a 19-species, 84-reaction reduced order reaction mechanism for the combustion of methane. A supervised learning approach involving stochastic gradient descent with backpropagation was used to optimize the DNN model; additionally, a self-organizing map (SOM) was used to cluster the model input into burnt, unburnt and reacting regimes. The framework was then examined in three cases: initial a-priori validation against the LES-PDF training set, a-posteriori validation of a single-cycle test apparatus developed at Purdue University (Purdue TJI rig), and a-posteriori validation of a single-cylinder engine developed at Argonne National Laboratory (Argonne TJI engine). The study found that the a-priori performance of the model was dependent on the state of the SOM cluster, as the framework was only able to accurately predict species that were active in a given cluster. The framework was found to capture the ignition in the Purdue TJI rig under a low-fidelity Reynolds averaged Navier Stokes (RANS) simulation, while a traditional laminar finite rate chemistry (LFRC) model was unable to capture this event. When applied to the Argonne TJI engine, the framework outperformed the common SAGE model that was unable to capture the pressure rise seen in experimental results. However, the performance of the model was limited by the availability of simulation data, and there exists room for improvement in the model.