Stabilized weighted reduced order methods for parametrized advection-dominated optimal control problems governed by partial differential equations with random inputs

Fabio Zoccolan, Maria Strazzullo, Gianluigi RozzaJ. Num. Math 2025 · DOI

TL;DR

This paper introduces a novel approach for efficiently solving complex optimal control problems governed by partial differential equations with random parameters. By combining stabilized numerical methods with reduced-order modeling techniques, the authors achieve significant speedups while maintaining high accuracy, making real-time analysis of such systems feasible.

What Problem Does This Paper Address?

Many real-world engineering and scientific applications involve optimal control problems where the underlying physics are described by partial differential equations (PDEs). These problems often become computationally intractable when they are "parametrized," meaning they depend on a set of input parameters that can vary. This is especially true when these parameters are random, introducing uncertainty that needs to be quantified. Traditional high-fidelity numerical methods, like the Finite Element Method (FEM), become prohibitively expensive when trying to explore the full range of parameter variations or when rapid evaluation is needed. Furthermore, advection-dominated problems (where a "flow" component is strong) are prone to numerical instabilities, further complicating their analysis. This paper aims to bridge this gap by developing efficient and stable reduced-order models (ROMs) for these challenging problems.

Methods

The authors employ a multi-pronged approach:

  • Stabilized Finite Element Method (FEM): To combat numerical instabilities in advection-dominated problems, they use the Streamline Upwind Petrov-Galerkin (SUPG) technique. This is applied within an "optimize-then-discretize" framework, meaning the optimality conditions are derived first, and then stabilized and discretized.
  • Reduced Order Modeling (ROM): To accelerate computations, they build ROMs. This involves creating lower-dimensional "reduced spaces" that capture the essential behavior of the high-fidelity solutions.
  • Weighted Proper Orthogonal Decomposition (wPOD): A key innovation is the use of wPOD, a modification of the standard Proper Orthogonal Decomposition (POD). wPOD accounts for the probability distribution of the random parameters, giving more weight to more likely scenarios and thus improving accuracy.
  • Offline-Online Paradigm: The ROM approach is split into two phases:
  • Offline Phase (Expensive): This involves generating many high-fidelity solutions (snapshots) for different parameter values, building the reduced spaces, and computing parameter-independent components. This is done only once.
  • Online Phase (Cheap): Once the offline phase is complete, new parameter-dependent solutions can be computed very rapidly using the pre-computed reduced spaces and components.
  • Uncertainty Quantification (UQ): The use of wPOD and various quadrature rules (Monte-Carlo, Quasi Monte-Carlo, tensor rules, sparse grids) allows for the effective quantification of uncertainty arising from random parameters.
  • Two Stabilization Strategies: The paper explores two ways to apply SUPG:
  • Offline-Only Stabilization: SUPG is applied only during the offline phase.
  • Offline-Online Stabilization: SUPG is applied in both offline and online phases.

Datasets

The paper does not introduce new datasets. Instead, it uses two well-established benchmark problems for numerical simulations:

  • Graetz-Poiseuille Problem: This models heat conduction in a pipe, with parameters affecting diffusion and geometry.
  • Propagating Front in a Square Problem: This models the movement of a front within a square domain, with parameters influencing the advection field.

Both problems are studied in both steady and unsteady (parabolic) settings, with various probability distributions assigned to the parameters to test the robustness of the methods.

Key Insights and Ideas

  • wPOD is Crucial for Accuracy with Randomness: The weighted POD (wPOD) significantly outperforms standard POD when dealing with random parameters. By accounting for the probability distribution, wPOD builds more representative reduced spaces, leading to much lower errors.
  • Offline-Online Stabilization is Key: While Offline-Only stabilization showed poor results, the Offline-Online stabilization strategy proved highly effective, yielding relative errors very close to the projection errors (the theoretical best achievable with a given reduced space).
  • Monte Carlo Shines for Accuracy: Among the various wPOD strategies tested, the Weighted Monte-Carlo method consistently delivered the best accuracy, often achieving results nearly 100 times better than standard POD.
  • Sparse Grids Offer a Good Balance: Smolyak sparse grid techniques generally performed better than their tensor product counterparts, offering a good balance between accuracy and computational cost, especially when dealing with higher-dimensional parameter spaces.
  • Stabilization is Non-Negotiable for Advection-Dominated Problems: The numerical results clearly demonstrate that without proper stabilization (like SUPG), advection-dominated problems lead to unstable and inaccurate solutions, even with advanced ROM techniques.

Key Results

  • Significant Speedup: The proposed methods achieve substantial speedups, with computational times for reduced solutions being orders of magnitude faster than high-fidelity FEM solutions (e.g., a speedup-index of around 10⁴ for steady problems and 10² for parabolic problems).
  • High Accuracy Achieved: With Offline-Online stabilization and wPOD (particularly Weighted Monte-Carlo), relative errors for state, control, and adjoint variables were reduced to very low levels (e.g., down to 10⁻⁷ for the state variable in some cases).
  • Robustness Across Problems: The methods demonstrated effectiveness on both the Graetz-Poiseuille and Propagating Front problems, in both steady and unsteady regimes, and across different parameter distributions.

Limitations and Future Work

  • Offline-Only Stabilization Inadequacy: The paper highlights that the Offline-Only stabilization strategy is insufficient for achieving good accuracy in these advection-dominated problems.
  • Computational Cost of Offline Phase: While the online phase is very fast, the offline phase (snapshot generation and POD construction) remains computationally intensive, especially for complex problems or high-dimensional parameter spaces.
  • Future Directions:
    • Developing strongly-consistent stabilization techniques that can handle geometrical parametrization without fluctuations.
  • Exploring other stabilization techniques for the online phase, such as Online Vanishing Viscosity or Online Rectification, which might offer computational advantages.
  • Extending the framework to boundary control and non-linear PDE constraints.
  • Comparing the "optimize-then-discretize" approach used here with the "discretize-then-optimize" approach for a more comprehensive understanding of numerical strategies.

Summary updated

BibTeX
@article{zoccolan2025stabilized,
  author = {Fabio Zoccolan and Maria Strazzullo and Gianluigi Rozza},
  title = {{Stabilized weighted reduced order methods for parametrized advection-dominated optimal control problems governed by partial differential equations with random inputs}},
  journal = {J. Num. Math},
  year = {2025},
  doi = {10.1515/jnma-2023-0006},
}
Figure