2021 | OriginalPaper | Buchkapitel
AI-Driven Gasoline Direct Injection Development: A Knowledge-Discovery Framework for Comprehensible Evaluations of Complex Physical Phenomena
verfasst von : Massimiliano Botticelli, Robin Hellmann, Paul Jochmann, Karl Georg Stapf, Erik Schünemann
Erschienen in: Internationaler Motorenkongress 2021
Verlag: Springer Fachmedien Wiesbaden
The complementary analysis through simulations and measurements led to a huge success in the development and the improvement of Gasoline Direct Injection (GDI) systems. The complexity of the physical interactions involved increases dramatically starting from the single component, e.g. the highpressure injector, going to the system level, including spray patterns, engine combustion and emissions. This implies strong non-linear and high multidimensional domains, requiring a significant effort in the generation and the evaluation of the necessary data. Furthermore, new emission regulations and the demand of high power output, as well as high efficiency, require deeper analysis and understanding in the field of GDI engines.In this paper a modular interdisciplinary AI-Based Knowledge-Discovery framework is presented and applied in the analysis of in-cylinder Computational Fluid Dynamics simulations, based on the variation of spray targeting coordinates and injection strategies. In particular, with a limited number of evaluations, the AI is able to explore and exploit the investigated domains, discovering connections and correlations among non-linear and complex phenomena. The framework is based on a novel explainable AI algorithm, able not only to provide robust models but also to allow the understanding of its decisions from a human point-of-view, stepping beyond the concept of black-box AI.After the introduction of the investigated dataset, the main structure and characteristics of the Knowledge-Discovery framework are presented. The nonlinear dependencies among spray targeting and injection strategies are then modeled through the AI algorithm. Beside the high prediction capabilities on engine results achieved by the models, the potential of their interpretability is used to gain a deeper understanding of complex physical phenomena, opening new frontiers for improvement and optimization.