Solving optimal control problems governed by random Navier-Stokes equations using low-rank methods

Peter Benner, Sergey Dolgov, Akwum Onwunta, Martin Stoll

Abstract

Many problems in computational science and engineering are simultaneously characterized by uncertainty, nonlinearity, nonstationarity, and high dimensionality. Existing numerical techniques for such models typically require considerable computational and storage resources. This is particularly true for optimization problems governed by time-dependent Navier-Stokes equations with uncertain inputs. In this paper, we approximate the solution using the low-rank Tensor Train (TT) decomposition and present a numerically efficient algorithm to solve the optimality equations directly in the low-rank representation. We demonstrate that the solution of the vorticity minimization problem admits a representation with ranks that depend modestly on model and discretization parameters, even for high Reynolds numbers. This approach opens the way for reduced-order modeling of stochastic optimal flow control with moderate costs at all stages.

The Challenge: The Curse of Dimensionality in Fluid Dynamics

In engineering, controlling fluid flow—such as minimizing vorticity (turbulence) behind an obstacle—is a critical task. However, real-world conditions are rarely deterministic; for instance, the velocity of an incoming fluid might fluctuate randomly. Modeling these uncertainties using the Navier-Stokes equations typically involves the Stochastic Galerkin Finite Element Method (SGFEM).

While robust, SGFEM suffers from the "curse of dimensionality." By combining spatial dimensions, time steps, and multiple random variables, the number of unknowns explodes. A problem manageable in a deterministic setting becomes prohibitively expensive to store or solve, even on supercomputing clusters, as the memory requirements grow exponentially with the number of random parameters.

Methodology: Lifting the Curse with Low-Rank Tensors

To overcome these computational barriers, this research utilizes a suite of low-rank numerical techniques designed to compress high-dimensional data without losing critical information:

  1. Tensor Train (TT) Decomposition: Instead of storing a massive high-dimensional tensor of all possible solutions, the solution is represented as a chain of small, interconnected tensors. This is analogous to matrix factorization but extended to many dimensions, drastically reducing storage needs.
  2. AMEn-Picard Iteration: To handle the nonlinearity of the Navier-Stokes equations, the authors employ Picard iteration. The resulting massive linear systems are solved using the Alternating Minimal Energy (AMEn) algorithm, which operates directly on the compressed TT format.
  3. Custom Preconditioning: Solving the "saddle-point" systems common in fluid dynamics is notoriously slow. The authors developed a specialized preconditioner based on the Schur complement, optimized for the TT format to ensure rapid convergence.
  4. Karhunen-Loève Expansion (KLE): This technique represents random inflow as a series of independent random variables, allowing the stochastic problem to be treated parametrically.

Experimental Setup

The method was evaluated using a classic computational fluid dynamics (CFD) benchmark: the backward-facing step domain. This geometry is standard for studying flow separation and reattachment under various conditions.

The backward-facing step domain used for numerical experiments, showing boundary conditions and coordinate system.

The experiments varied the Reynolds number (via viscosity $\nu$), the number of random variables ($m$ up to 8), and the regularization parameter ($\beta$). The researchers tested both distributed control (acting on the entire domain) and boundary control (acting on a specific wall, labeled $\Upsilon_{cnt}$ in the figure).

Key Insights and Results

The study yielded several breakthrough observations regarding the efficiency of low-rank methods in stochastic control:

  • Natural Compressibility: The solutions to these complex stochastic control problems possess naturally low "TT-ranks." This confirms that high-dimensional fluid data contains significant redundancy that the TT format can exploit.
  • Massive Memory Savings: In the largest test cases—involving over $10^{15}$ potential unknowns—the TT format achieved a memory saving of nine orders of magnitude. This allows simulations that previously required supercomputers to run on a standard desktop.
  • Linear Scaling: Unlike traditional methods where costs grow exponentially, the computational cost of this low-rank approach grows only linearly or logarithmically with the number of random variables.
  • Control Performance: The algorithm successfully minimized vorticity despite uncertainty, effectively smoothing out the eddies typically created by the backward step.

Limitations and Future Outlook

While the results are promising, the authors noted that boundary control remains more challenging than distributed control, as it leads to higher TT-ranks at high Reynolds numbers. Furthermore, extremely turbulent flows (very high Reynolds numbers) may eventually test the limits of current compression techniques. Future work may involve integrating semi-smooth Newton algorithms to handle hard state constraints, such as velocity limits, within this low-rank framework.

Summary updated

Abstract

Many problems in computational science and engineering are simultaneously characterized by the following challenging issues: uncertainty, nonlinearity, nonstationarity and high dimensionality. Existing numerical techniques for such models would typically require considerable computational and storage resources. This is the case, for instance, for an optimization problem governed by time-dependent Navier-Stokes equations with uncertain inputs. In particular, the stochastic Galerkin finite element method often leads to a prohibitively high dimensional saddle-point system with tensor product structure. In this paper, we approximate the solution by the low-rank Tensor Train decomposition, and present a numerically efficient algorithm to solve the optimality equations directly in the low-rank representation. We show that the solution of the vorticity minimization problem with a distributed control admits a representation with ranks that depend modestly on model and discretization parameters even for high Reynolds numbers. For lower Reynolds numbers this is also the case for a boundary control. This opens the way for a reduced-order modeling of the stochastic optimal flow control with a moderate cost at all stages.
BibTeX
@misc{benner2017solving,
  author = {Peter Benner and Sergey Dolgov and Akwum Onwunta and Martin Stoll},
  title = {{Solving optimal control problems governed by random Navier-Stokes equations using low-rank methods}},
  year = {2017},
  eprint = {1703.06097},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1703.06097},
}
Figure