Integration of Neural Networks in GT-Suite and Coupling with an RL-Agent for Automation of Control and Regulation Tasks for a commercial H2 engine

Due to the change in mobility to lower CO2 emissions, there is a turnaround in the automotive industry. The share of classic combustion engines with petrol or diesel fuel will continue to decline. As a result, the development capacities for ICEs will also decrease. Therefore, development process becomes an even greater challenge. The IFS has been researching in the application of artificial intelligence in the development process since 2019. Various successes have already been achieved here, e.g. in virtualizing the combustion chamber or the stack of a fuel cell with the help of neural networks. Since the majority of all AI projects fail, the application of current research projects has to be carried out at an early stage. In the course of the paper, first the combustion chamber of a hydrogen engine is modelled with neural networks. In contrast to the petrol engines used in the previous papers, the combustion speed of hydrogen combustion is much higher. Despite that, very good predictions can be made with the neural networks. The connection of the networks to the 0D-/1D-simulation software, GT-Suite, is shown as an example. Here, two different possibilities are examined via the FMU interface. In the first variant, an FMU of the neural networks is generated and then integrated into the software. This makes it possible to generate a user-friendly variant that can also be sent to a calculation cluster. This was used to do a benchmark with regard to the computing speed and prediction quality. The computing time could be accelerated by a factor of 4. In the second variant, an FMU is made from the simulation and coupled with the networks in Python. Since this also allows direct coupling with Python, an OpenAI Gym environment is set up for coupling with an RL agent. Following this, an RL agent for throttle control was trained.