TL;DR
This paper introduces a new model reduction technique called MP-LSVT, designed to tackle complex, multi-scale problems with extreme stiffness, like reacting flows. By preserving the governing equation structure and using a novel least-squares approach with variable transformations, MP-LSVT achieves significantly improved stability, accuracy, and computational efficiency compared to existing methods, even achieving over 200x speedup on a 3D simulation.
What Problem Does This Paper Address?
Simulating complex multi-physics problems, especially those involving reacting flows, is computationally very expensive. These simulations often suffer from "extreme scale disparity" and "extreme stiffness," meaning they involve phenomena happening at vastly different time and length scales, leading to very stiff systems of equations. This stiffness makes it difficult for standard numerical methods and reduced-order models (ROMs) to achieve stable and accurate results, often leading to spurious oscillations or outright instability. Existing projection-based ROMs, like Galerkin and LSPG, struggle with these challenges, particularly when dealing with transport-dominated phenomena like convection. The authors aim to develop a more robust and efficient ROM framework that can handle these difficult problems.
Methods
The core of the paper is the Model-form Preserving Least-Squares with Variable Transformation (MP-LSVT) formulation. Key aspects include:
- Variable Transformation: Allows for the selection of an arbitrary, but complete, set of solution variables, offering flexibility beyond traditional conserved variables. This transformation is designed to preserve the structure of the governing equations.
- Least-Squares Minimization: Instead of standard Galerkin projection, MP-LSVT minimizes the residual of the fully-discrete FOM equations. This approach, combined with a specific test basis construction, leads to a symmetrized and discrete consistent ROM.
- Two Levels of Scaling: To handle the extreme stiffness, two scaling matrices are employed. One scales the solution variables for constructing POD bases, and another scales the FOM residual to ensure all equations contribute equally to the minimization problem.
- Local Stability Limiters: To mitigate spurious oscillations, especially in reacting flows, physical realizability limiters are enforced on temperature and species mass fractions.
- Hyper-reduction: For further computational efficiency, the authors integrate hyper-reduction techniques (specifically gappy POD) to approximate non-linear terms using sparse sampling.
- Linear Stability Proof: The paper provides a theoretical proof for the linear stability of least-squares-based ROMs for linear time-invariant systems.
Datasets
The authors evaluate their method on two configurations of a generic laboratory-scale single injector combustor:
- 2D Planar Reacting Injector: A simplified representation used for detailed accuracy and robustness assessments of different ROM formulations.
- 3D Reacting Injector: A full three-dimensional representation, used to evaluate computational efficiency and predictive capabilities, particularly with hyper-reduction.
The FOM simulations for these cases generate snapshots of solution variables (pressure, temperature, species mass fractions, etc.) over time, which are then used to construct the Proper Orthogonal Decomposition (POD) bases for the ROMs.
Key Insights and Ideas
- Preserving Equation Structure is Key: The MP-LSVT approach emphasizes preserving the structure of the governing equations through variable transformation. This is crucial for achieving discrete consistency and better stability, especially in stiff systems.
- Least-Squares for Symmetrization and Stability: Minimizing the residual in a least-squares sense, rather than just projecting it, leads to a more stable and well-conditioned ROM system. This is a significant departure from standard Galerkin projection.
- Limiters are Essential for Reacting Flows: For problems with stiff chemistry, like reacting flows, simple limiters on physical quantities (temperature, species mass fractions) are not just helpful but necessary to prevent spurious oscillations that can lead to unrealistic results (e.g., negative temperatures, excessive heat release) and terminate simulations.
- Hyper-reduction Doesn't Always Kill Accuracy: The study shows that hyper-reduction, while drastically speeding up computations, can maintain good accuracy for reconstruction error, especially when combined with explicit time integrators.
- Future-State Prediction Limitations: Even with advanced ROMs, predicting dynamics far beyond the training data remains challenging, particularly for chaotic systems with sharp gradients. This limitation stems more from the static nature of POD bases than the ROM formulation itself.
Key Results
- Improved Stability and Accuracy: MP-LSVT ROMs demonstrated significantly improved stability and accuracy compared to Galerkin and LSPG methods across various test cases, including a 2D reacting flow problem. Galerkin ROMs were often unstable, while MP-LSVT remained stable even with a small number of modes.
- Computational Efficiency: In a 3D reacting flow simulation, MP-LSVT with hyper-reduction achieved over a 319x speedup in wall clock time and a 638x speedup in CPU time compared to the FOM, using an explicit time integrator and a physical time step 10 times that of the FOM, while maintaining less than 2% reconstruction error.
- Linear Stability: A theoretical proof demonstrated the linear stability of MP-LSVT ROMs for linear time-invariant systems. Numerical analysis confirmed improved conditioning and stability over other methods.
- Effective Limiters: The species mass fraction limiter was shown to be effective in mitigating spurious oscillations and improving the accuracy of heat release rate and flame propagation speed predictions in reacting flow problems.
Limitations and Future Work
- Future-State Prediction: The paper acknowledges that predicting dynamics far beyond the training period remains challenging, especially for chaotic systems with sharp gradients. This is attributed to the limitation of static POD bases.
- Future Work Suggestions: The authors suggest leveraging adaptive bases or non-linear manifold ROMs to improve future-state prediction capabilities. Further exploration of explicit constraint preservation (like enforcing conservation) is also proposed.
- Hyper-reduction Sensitivity: For implicit time integration schemes, the hyper-reduced MP-LSVT ROMs showed sensitivity to physical time step size, with errors increasing for larger steps. This is an area for further investigation.
Summary updated