Feedback Control of Nonlinear PDEs Using Data-Efficient Reduced Order Models Based on the Koopman Operator

Sebastian Peitz, Stefan KlusLecture Notes in Control and Information Sciences 2020 · DOI · arXiv · OpenAlex

Introduction: The Challenge of Real-Time PDE Control

Many critical engineering systems, ranging from atmospheric modeling to chemical reactor design and aerodynamics, are governed by nonlinear partial differential equations (PDEs). Controlling these systems in real-time remains a "grand challenge" because the underlying equations are computationally expensive to solve.

Model Predictive Control (MPC) is the gold standard for such tasks, as it repeatedly solves an optimization problem over a future time horizon. However, for high-fidelity PDE models like the Navier-Stokes equations, a single optimization step can take hours. While traditional Reduced Order Models (ROMs) like Proper Orthogonal Decomposition (POD) offer some relief, they are often difficult to implement for nonlinear systems and require extensive training data across all possible control inputs.

Methodology: Leveraging the Koopman Operator

The authors propose a data-efficient approach using the Koopman operator, a mathematical framework that transforms a finite-dimensional nonlinear system into an infinite-dimensional linear one. This transformation allows researchers to apply powerful linear algebra tools to inherently nonlinear problems.

The Koopman operator framework for approximating nonlinear dynamics as linear systems.

The core innovation lies in the data-efficient construction of these models:

  1. Switched Systems: Instead of modeling every possible control input, the system is modeled for a few constant inputs (e.g., "Full On," "Off"). The controller then treats the problem as a "switching" task.
  2. Bilinear Interpolation: To enable continuous control, the researchers interpolate between the Koopman operators of autonomous systems, creating a fast-to-solve "bilinear" surrogate model.
  3. Extended Dynamic Mode Decomposition (EDMD): This algorithm numerically approximates the Koopman operator from data using a dictionary of basis functions, such as monomials.
  4. Online Adaptivity: An incremental update scheme allows the model to "learn" and improve accuracy on the fly using real-time sensor data.

Evaluation and Datasets

The methodology was validated on two classic benchmarks:

  • 1D Burgers Equation: A fundamental PDE combining advection and diffusion, tested with periodic boundary conditions and distributed control.
  • 2D Navier–Stokes Equations (NSE): Specifically, the "cylinder wake" problem at a Reynolds number of 100. The objective is to control lift and drag forces by rotating the cylinder. The ground truth was established using an OpenFOAM simulation with 22,000 cells.

Key Insights

  • Linearity as a Catalyst: Shifting complexity from dynamics to observables makes the control problem a linear optimization, resulting in massive speedups.
  • Simplicity in Basis Functions: Low-order monomial dictionaries (e.g., order 2) often outperformed higher-order ones, proving more robust and requiring less data.
  • Equation-Free Approach: The method is entirely data-driven, requiring only snapshots of the system's state rather than explicit knowledge of the underlying PDE.

Results: Massive Speedups and Dimension Reduction

The Koopman-based ROMs demonstrated exceptional performance across both benchmarks. For the 1D Burgers equation, the ROM was approximately 100x faster than the full PDE solver. In the more complex 2D Navier-Stokes case, the speedup reached a staggering 75,000x.

Performance comparison and speed-up results for 1D Burgers and 2D Navier-Stokes equations.

Furthermore, the 2D NSE system's 22,000 degrees of freedom were compressed into a Koopman model with a dimension of just 45. Online adaptation also proved highly efficient, reaching high accuracy within 40-50 seconds of operation.

Limitations and Future Outlook

While promising, the bilinear interpolation method assumes a linear dependence on control inputs, which may not hold for all systems. Additionally, the model's accuracy depends on the coverage of the training data; performance may degrade if the system enters unexplored regions of the state space. Future work will focus on testing these methods on physical hardware and exploring more complex, chaotic systems with higher Reynolds numbers.

Summary updated

BibTeX
@inproceedings{peitz2020feedback,
  author = {Sebastian Peitz and Stefan Klus},
  title = {{Feedback Control of Nonlinear PDEs Using Data-Efficient Reduced Order Models Based on the Koopman Operator}},
  booktitle = {Lecture Notes in Control and Information Sciences},
  year = {2020},
  doi = {10.1007/978-3-030-35713-9\_10},
}
Figure