A multi-fidelity, reduced-order modeling framework for the development of digital twins of combustion systems
(2025)
Files
OzdenAysu_PhDThesis_final_v2.pdf
Open Access - Adobe PDF
- 26.42 MB
Details
- Authors
- Supervisors
- Abstract
- Climate change is intensifying global challenges such as wildfires, droughts, and rising temperatures. To mitigate these effects, reducing greenhouse gas and pollutant emissions from energy systems is essential. Since most emissions originate from combustion-based processes, developing cleaner and more efficient combustion technologies is a key priority. Designing advanced combustion systems is, however, complex, costly, and time-consuming. Traditional approaches rely on extensive experiments and high-fidelity simulations, which are limited by accessibility, cost, and computational demands. Although Computational Fluid Dynamics (CFD) tools have advanced significantly, predicting detailed combustion phenomena remains computationally expensive. To address these challenges, this research explores data-driven reduced-order models (ROMs) that can replicate the behavior of full-order simulations at a fraction of the cost. While ROMs accelerate design and optimization, they typically require large, high-fidelity datasets for training. Multi-fidelity modeling offers a solution by combining accurate but expensive high-fidelity data with inexpensive low-fidelity data, achieving reliable predictions with reduced computational effort. The multi-fidelity reduced-order modeling framework developed in this work integrates Proper Orthogonal Decomposition (POD) for dimensionality reduction, Procrustes Manifold Alignment (PMA) to create shared low-dimensional manifolds across fidelities, and CoKriging, an extension of Gaussian Process Regression, for accurate state prediction. To optimize data usage, incremental sampling algorithms were designed to minimize both prediction error and model uncertainty. The framework was first tested on a methane–hydrogen furnace under MILD (Moderate and Intense Low-oxygen Dilution) combustion conditions, showing that uncertainty-based sampling strategies were most effective. The model achieved accuracy comparable to single-fidelity models while halving the number of required training samples. The framework’s adaptability was further validated using ammonia combustion in a stagnation point reverse-flow combustor. Despite reduced accuracy in extrapolation cases, the model successfully captured key physical trends outside the training range. A hierarchical clustering approach was later introduced to improve training sample distribution, enhancing both prediction accuracy and computational efficiency. Finally, a multi-level multi-fidelity model was developed by integrating experimental data as high-fidelity inputs, and 3D/2D RANS simulations as mid- and low-fidelity datasets. Incorporating experimental data improved model stability and accuracy, confirming the framework’s robustness. However, sensitivity to dataset selection highlighted the need for careful data linkage and sampling strategies. Overall, this research demonstrates that multi-fidelity reduced-order models offer a powerful and flexible approach to balance accuracy, cost, and efficiency in combustion modeling. Their ability to adapt across fuels and configurations makes them valuable tools for accelerating the design of more sustainable energy systems.
- Affiliations
Citations
Özden, A. (2025). A multi-fidelity, reduced-order modeling framework for the development of digital twins of combustion systems.