Artificial Neural Network Models Development for Diesel Oxidation Catalyst Characterization

The adoption of numerical models, such as Artificial Neural Networks, to reproduce the behavior of a physical system basing only on data-driven approach, offers a powerful technique in simulation environment. This master thesis work exploits this approach to develop Artificial Neural Network models aimed at simulating the functioning of a Diesel Oxidation Catalyst, one of the first after-treatment system's component present into exhaust line of diesel engine applications. The study begins with the generation of several Design of Experiments to simulate most common operating conditions by means of pre-existing physics-based 1D model to collect data, which serve as the foundation for the construction of several single-output models. Each of them describes a specific variable of the component, ranging from wall temperature to the chemical species exiting from the reactor. Subsequently, to assess the robustness of these models, they are subjected to testing on several emissions driving cycles, commonly used to analyze the impact of engine emissions in real word scenario. Then, a final part focuses on the combination of these models to assess the results of a global model, capable of accurately emulating the physics of the entire diesel oxidation catalyst component. Furthermore, the strength of the discretization approach has been proved during this work, showing improvements in prediction accuracy of multiple-sub volume model's up to three times higher compared to single-volume one. Anyway, the significance of this work lies in its contribution to advance emission control technologies, providing a sophisticated tool which marries the strength of data-driven neural networks with the complexity of thermal and chemical phenomena. The presented methodology has revealed promising solution in the field of engine emission control environment, as the developed neural network models result in predictions of thermal and chemical species which are very close to the measured data, even under most challenging operating conditions. Overall, this work aims to highlight the robustness of these data-driven models, providing a valuable alternative to classical physics-based simulation model.