A Feature-Focused Neural Network based Bi-Fidelity Surrogate Model for Combustion Prediction in Aerial Internal Combustion Engines

Energy Assisted Compression Ignition (EACI) has garnered significant attention for its applicability in aerial propulsion, particularly in enabling ignition with low cetane number (CN) jet fuels. A ceramic-based glow plug serves as an ignition assistant (IA) to support combustion in such challenging conditions. Effective control systems are essential for internal combustion (IC) engines to optimize performance, ensure reliability, and manage parameters such as injection timing, pressure, and after-treatment strategies. Training these control systems requires a large volume of pressure-crank-angle data across diverse operating conditions. However, collecting such data experimentally is time-consuming and cost-intensive, especially under low-CN scenarios. While Computational Fluid Dynamics (CFD) simulations offer a scalable alternative, they are also expensive computationally when extended to wide-ranging input conditions. In this study, we proposed a layered bi-fidelity modeling framework using artificial neural networks (ANNs) to combine cost-effective CFD data ("low'' fidelity) with limited experimental measurements ("high'' fidelity). The proposed layered learning structure allows the network to first generalize from abundant low-fidelity CFD data and subsequently specialize using sparse but accurate high-fidelity data through a residual correction strategy. The resulting surrogate achieves physically consistent pressure predictions across a wide range of engine parameters, significantly reducing the dependency on large experimental datasets. The model demonstrates strong generalization, enabling efficient and accurate training of IC engine control systems for complex EACI operations, with potential applications in future aviation propulsion platforms.