A New Class of AI-based Engine Models

An interesting field of application for neural networks is the mapping of system components in 100 to 1000-fold real-time while maintaining high quality. However, due to the limited extrapolation capability of neural networks, data for all possible operating states must be available during the training phase. Since the necessary amount of data cannot be generated by test bench experiments, fast, predictable and easily parallelizable simulation models must be used for this purpose. For example, the work process calculation in combination with phenomenological models can sufficiently represent the fundamental processes in the combustion chamber.This paper describes how a data set with almost 9.3 million different operating points of a gasoline engine is in the beginning generated using software tools available at the FKFS. Special characteristics of certain motor parameters regarding the network training are shown, as well as the reduction of the data set before the training to increase the network quality. In the following, the layout and training of a neural network is described, which predicts some combustion characteristics(mfb-10 %, mfb-25 %, mfb-50 %, mfb-75 % and mfb-90 % points), pressure characteristics (IMAP, peak pressure and its position as well as the pressure at exhaust opening) and NO emissions according to the data set.This is followed by a presentation of the statistical accuracy of the network and a detailed view of an exemplary map area. Finally, the integration of the network in GT-Suite and the coupling with FKFS RapidCylinder is shown and an exemplary load jump is considered. The paper thus represents a first proof-of-concept of how neural networks can be used in powertrain development.