Computational fluid dynamics-based feed-forward control of multi-fuel energy-assisted compression ignition engines for sustainable aviation

Multi-fuel compression ignition (CI) engines are expected to play a vital role in improving the versatility and operational range of aviation propulsion systems. These engines are required to function efficiently across a broad range of fuel specifications, such as low cetane number (CN) sustainable aviation fuels (SAF). Due to challenges in igniting low CN fuels at high altitude conditions, ignition assistants such as glow plugs are used to enhance combustion. In order to achieve the desired performance, the engine control system needs to optimize parameters such as the fuel injection and ignition assistant settings. Data-driven control models such as Gaussian process (GP) regression are popularly used to represent the physical system, and usually require a comprehensive training dataset. Typically, such datasets are generated through physical experiments, which can often be challenging for a broad range of conditions and parameters. Furthermore, testing control models requires physical proximity of the controller and the engine, which is another challenge. A viable solution to these challenges is to utilize physics-based virtual engines to supplement both data generation and model testing. This study proposes a highly practical framework which couples computational fluid dynamics (CFD) with a feed-forward control model for both training and testing of multi-fuel engines. The results from the virtual testing were compared with real fuel-switch experimental results, and high qualitative agreement was observed between the two. Additionally, the feed-forward control models developed from CFD and experimental data are cross-validated against each other, where the GP models trained from the two data sources serve as two engine surrogates. A high degree of similarity in trends and compatibility between the CFD- and experiment-based models are observed, lending confidence in using virtually trained control models in real systems.