Predictive reduced order modeling of chaotic multi-scale problems using adaptively sampled projections

Cheng Huang, Karthik DuraisamyJournal of Computational Physics 2023 · OpenReview

TL;DR

This paper introduces an adaptive reduced-order modeling (ROM) technique that significantly improves predictive accuracy for chaotic, multi-scale fluid dynamics problems. By dynamically updating the model's basis and sampling points in real-time, it overcomes limitations of traditional ROMs, enabling accurate predictions even for unseen conditions and complex phenomena like turbulence and combustion.

What Problem Does This Paper Address?

Simulating complex fluid dynamics problems, especially those involving chaotic, multi-scale, and transport-dominated physics (like turbulent combustion), is computationally expensive and often intractable for many engineering applications such as optimization and uncertainty quantification. Traditional Reduced-Order Models (ROMs) offer a way to speed up these simulations by creating lower-dimensional approximations. However, existing ROMs struggle with:

  • Stability Issues: They can become unstable due to inherent numerical limitations or the truncation of high-frequency dynamics.
  • The Kolmogorov Barrier: For problems with a slow decay of energy across different scales, linear ROMs (like those based on Proper Orthogonal Decomposition - POD) require an impractically large number of basis modes to capture the dynamics accurately.
  • Limited Predictive Capability: Standard ROMs are often trained on a specific set of parameters or conditions. When faced with new, unseen parameters or significant changes in dynamics (e.g., different flow rates or acoustic perturbations), their predictive accuracy plummets, making them more of a reproduction of training data than a true predictor.

This paper aims to address these limitations by developing an adaptive ROM that can accurately predict the behavior of chaotic, multi-scale systems, even when encountering novel conditions or dynamics.

Methods

The core of this work is an adaptive projection-based reduced-order model (ROM) formulation. Key methodological contributions include:

  • Model-Form Preserving Least-Squares with Variable Transformation (MP-LSVT): This forms the foundation for constructing a robust ROM. It uses a least-squares approach with a variable transformation to enhance stability and accuracy.
  • Online Basis Adaptation: Instead of a static basis determined offline, this method dynamically updates the low-dimensional subspace (the "basis") at each time step during the online simulation. This allows the ROM to adapt to newly emerging unresolved dynamics.
  • Non-local Full-State Estimation: To improve predictive capabilities, especially for phenomena exhibiting non-local coherence (like acoustic waves), the method incorporates information from unsampled regions of the domain. This is achieved by estimating full states at unsampled locations using a larger time step, balancing computational cost with accuracy.
  • Adaptive Sampling Point Selection: The locations where the ROM samples information from the full-order model (FOM) are also updated dynamically. This is done to minimize interpolation errors introduced by hyper-reduction, ensuring that the most informative regions are sampled.
  • Predictor-Corrector Approach: The adaptation of the basis and sampling points is handled in a decoupled manner, inspired by a predictor-corrector strategy, to manage computational complexity.

The overall approach aims to minimize offline training, making the ROM construction largely an online process.

Datasets

The paper evaluates its adaptive ROM formulation on two representative reacting flow benchmark problems:

  • 1D Propagating Laminar Flame: This is a well-established benchmark problem used to test ROMs. The authors use FOM simulations of this flame under various acoustic perturbations (different frequencies and amplitudes) as their testing dataset.
  • 2D Reacting Injector: This is a more complex benchmark representing a generic laboratory-scale rocket combustor. FOM simulations of this system, including shear coaxial injection and acoustic perturbations, are used to evaluate the adaptive ROM's capabilities in predicting long-term dynamics and transient changes due to parametric variations.

The paper does not introduce new datasets but utilizes existing benchmark problems to demonstrate the efficacy of its proposed method.

Key Insights and Ideas

  • Dynamic Adaptation is Crucial for Prediction: The most significant insight is that static ROMs, even with extensive offline training, struggle to predict dynamics outside their training regime. Adaptive ROMs that update their basis and sampling in real-time are essential for true predictive capabilities in chaotic systems.
  • Non-local Information Bridges the Gap: Incorporating information from unsampled regions (non-local coherence) is vital for accurately capturing global phenomena like acoustic waves, significantly improving prediction accuracy compared to methods that only consider local information.
  • Balancing Accuracy and Efficiency: The choice of parameters like the non-local estimation rate ($z_s$) involves a trade-off between prediction accuracy and computational efficiency. Higher $z_s$ values reduce computational cost but can decrease accuracy, requiring careful tuning.
  • "No Offline Stage" is Achievable: The adaptive approach significantly minimizes the need for extensive offline training, making ROM construction more accessible and enabling rapid deployment for new problems or parameters.
  • Beyond Reproduction to Prediction: The adaptive ROM moves beyond simply reproducing known dynamics to truly predicting future states and behavior under novel conditions, a critical step for practical engineering applications.

Key Results

  • Significant Error Reduction: The adaptive ROM with non-local information incorporated achieved error reductions of approximately two orders of magnitude compared to adaptive ROMs without this feature, particularly for the 1D laminar flame problem.
  • High Accuracy in Complex Flows: The method demonstrated excellent agreement with FOM solutions for both the 1D laminar flame and the 2D reacting injector problems, achieving prediction errors below 8% for pressure and 15% for temperature in the 2D case.
  • Accurate Parametric Predictions: The adaptive ROM successfully predicted flame dynamics under various acoustic perturbations and even captured phenomena like flame extinction and enhancement due to changes in oxidizer mass flow rates.
  • Efficiency Gains: While not achieving the same level of efficiency as static ROMs (due to the overhead of adaptation), the adaptive ROM still provides substantial computational speedups (e.g., O(22) acceleration in computational time for the 1D flame).
  • Robustness to Initial Conditions: The adaptive ROM's accuracy showed little sensitivity to the initial training window size, indicating flexibility and ease of use.

Limitations and Future Work

  • Computational Cost of Adaptation: While significantly reducing offline costs, the online adaptation process still adds computational overhead compared to static ROMs. The paper notes that current adaptive formulations yield one to two orders of magnitude cost reduction, whereas static ROMs can achieve three to four.
  • Parameter Tuning: The selection of parameters like $z_s$ (the rate for non-local estimation and sampling point updates) is crucial and requires careful tuning to balance accuracy and efficiency.
  • Scalability: The authors suggest that further development is needed for scalable implementation of the adaptive ROM algorithm for large-scale engineering problems, particularly in developing effective load-balancing strategies.
  • Fixed Parameters: The current formulation uses fixed input parameters (reduced space dimension, initial training window, number of sampling points). Future work could explore adapting these parameters over time.
  • Further Investigations: More research is needed on strategies for selecting sampling points and in hyper-reduction.

Summary updated

Abstract

Highlights•An adaptive reduced-order model formulation is derived to enable true predictions of chaotic and convection-dominant physics.•A compact method is formulated to update the basis to account for unresolved scales at every time step.•A full state estimation strategy is developed to incorporate non-local information in ROM adaptation.•The adaptive procedure significantly reduces the offline training cost, making it negligible compared to the online calculations.•Detailed evaluations show significant predictive accuracy in future state and parametric prediction in chaotic flows.
BibTeX
@article{huang2023predictive,
  author = {Cheng Huang and Karthik Duraisamy},
  title = {{Predictive reduced order modeling of chaotic multi-scale problems using adaptively sampled projections}},
  journal = {Journal of Computational Physics},
  year = {2023},
  url = {https://openreview.net/forum?id=Uqylw0cuod},
}
Figure