University of Minnesota
Computational Fluid Dynamics and Physics-Integrated Machine Learning Frameworks for Modeling, Sampling, and Virtual Control Testing of Multi-Fuel Energy-Assisted Compression Ignition Engines
Abstract
dc:description.abstractMulti-fuel capability is an increasingly important requirement for internal combustion engines used in aviation, transportation, and power generation, as stringent emissions restrictions and growing energy demands drive the adoption of clean renewable fuels alongside conventional ones. When an engine must operate across a wide range of fuels, however, the accompanying variation in fuel properties-particularly the fuel cetane number (CN)-poses significant challenges to reliable combustion. These challenges are compounded under extreme conditions such as high-altitude operation (low temperature and pressure), which can lead to undesirable events such as misfires (ignition failure) if the engine is not designed properly. Ignition assistance technologies, such as glow plugs, spark igniters, and plasma igniters, can be used to enhance combustion under such challenging conditions; this dissertation first investigates the numerical modeling of these strategies using computational fluid dynamics (CFD). The engine control system manages the fuel injection and energy-assistant settings so that the engine operates optimally, avoiding undesirable events such as misfires. A robust control strategy is thus required to function amidst the highly variable and dynamic phenomena introduced in energy-assisted multi-fuel combustion. Such control models are typically trained offline using a large database of experimental data generated under a wide range of parametric conditions. However, experimental data generation is challenging: delays in engine part manufacturing, engine malfunctions, and measurement errors can all lead to insufficient data being produced, making it time-consuming and expensive to conduct the large parametric campaigns needed to adequately train and test the control strategy. Numerical CFD simulations using detailed physics-based solvers can help fulfill this demand for data generation, and can also be coupled with the control strategy for testing, in a similar fashion to real engine tests. However, since CFD simulations are computationally expensive, machine learning (ML) based reduced-order models (ROMs) are developed in this dissertation to complement data generation for control system training and testing. Physics-integrated Gaussian process (GP) based ROMs are developed by incorporating known combustion physics, such as pressure-curve segmentation and misfire information, into the probabilistic learning framework, allowing for accurate prediction with limited training data. A multi-fidelity extension of the GP framework is also developed, which blends CFD and experimental data via transfer learning; an analogous multi-fidelity neural network (NN) based ROM is developed as well. It was observed that incorporating physics and/or blending multiple data fidelities significantly improves prediction accuracy in comparison with traditional, single-fidelity data-driven models. As the number of control-relevant input parameters increases, uniformly sweeping the parametric space via CFD simulations or experiments becomes infeasible. Thus, an iso-surface sampler (ISS) is developed for smart and efficient data sampling in high-dimensional spaces, supplemented by a Bayesian optimization based sampling approach for identifying a diverse set of desirable operating conditions. Finally, once the ROM and/or CFD tool has generated sufficient data for control system training, the resulting feed-forward control strategy (i.e., look-up tables) is validated by using the CFD simulation as a virtual engine (i.e., a digital twin) for control system testing, in lieu of a real engine. The virtual and real engine control test results, for a fuel-switching scenario, show good qualitative agreement, and the CFD- and experiment-trained control models are further cross-validated against each other using GP-based surrogate engines. On the whole, this dissertation demonstrates that physics-integrated, multi-fidelity ROMs, along with smart high-dimensional sampling and CFD-based virtual engine testing, can substantially reduce the data generation and testing burden for developing robust control strategies for multi-fuel energy-assisted compression ignition engines. Further validation across a broader range of operating conditions and engine platforms is a valuable direction for future work.
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Narayanan, Sai Ranjeet
Subjects
dc:subject × 5Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Dc Identifier Citation
- Narayanan, Sai Ranjeet. (2026). Computational Fluid Dynamics and Physics-Integrated Machine Learning Frameworks for Modeling, Sampling, and Virtual Control Testing of Multi-Fuel Energy-Assisted Compression Ignition Engines. Retrieved from the University Digital Conservancy, https://hdl.handle.net/11299/283126.