An optimisation-based domain-decomposition reduced order model for the incompressible Navier-Stokes equations

Ivan Prusak, Monica Nonino, Davide Torlo, Francesco Ballarin, Gianluigi Rozza

TL;DR

This paper introduces a novel method that combines domain decomposition and reduced-order modeling to efficiently solve complex fluid dynamics problems governed by the Navier-Stokes equations. By breaking down the problem into smaller parts and creating simplified models, they significantly reduce computational costs while maintaining accuracy, making it easier to simulate and optimize fluid flows.

What Problem Does This Paper Address?

Simulating fluid dynamics, particularly with the incompressible Navier-Stokes equations, is computationally very expensive. This is a major bottleneck for many engineering and scientific applications, from designing aircraft to understanding blood flow. Traditional methods require immense computing power, limiting the scope and speed of simulations, especially when dealing with parameter-dependent problems (where physical properties or boundary conditions change). Existing approaches to reduce this computational burden, like domain decomposition (breaking the problem into smaller, manageable pieces) and reduced-order modeling (creating simplified, faster models), have their own limitations. This paper aims to bridge this gap by developing a unified framework that leverages the strengths of both techniques to achieve significant computational savings without sacrificing accuracy.

Methods

The core of this work lies in combining two powerful techniques:

  • Optimization-based Domain Decomposition (DD): The physical domain is split into subdomains. Instead of directly solving the full problem on each subdomain and trying to match solutions at the interfaces, an optimization problem is formulated. This optimization minimizes the "coupling error" at the interfaces, effectively allowing each subdomain to be solved independently. This is achieved by introducing a Lagrange multiplier (represented by 'g') that acts as a control variable on the interface.
  • Reduced-Order Modeling (ROM): Once the optimization problem is set up, a reduced-order model is built for it. This involves:
  • Proper Orthogonal Decomposition (POD): This technique is used to identify the most important "modes" or patterns in the high-fidelity solutions (obtained from standard numerical methods). These modes capture the essential physics of the problem with far fewer degrees of freedom.
  • Galerkin Projection: The reduced basis functions derived from POD are used to project the original, complex equations onto a much smaller-dimensional space, creating a computationally inexpensive ROM.
  • Lifting Function and Velocity Supremiser: Special techniques are employed to handle non-homogeneous boundary conditions (where the solution is prescribed) and to ensure the stability of the reduced-order model for the saddle-point nature of the Navier-Stokes equations.

The novelty lies in applying ROM to the optimization problem that arises from DD, rather than just to the original fluid dynamics problem.

Datasets

The paper uses two standard fluid dynamics benchmarks for evaluation:

  • Backward-facing step flow: A common test case for fluid dynamics simulations, involving flow over an obstacle.
  • Lid-driven cavity flow: Another standard benchmark, simulating fluid motion within a closed box driven by a moving lid.

These are not new datasets but rather established test cases used to validate the proposed methodology.

Key Insights and Ideas

  • Synergy of DD and ROM: The paper demonstrates that combining DD with ROM is more than just additive; it creates a powerful synergy. ROM applied to the DD optimization problem leads to significant computational gains.
  • Reduced Optimization Cost: The ROM approach not only reduces the cost of solving the fluid dynamics equations but also drastically cuts down the number of iterations required by the optimization algorithm used to enforce interface conditions.
  • Parameter-Dependent Problems: The framework is particularly well-suited for problems where physical parameters (like viscosity or inlet velocity) vary, as the ROM can efficiently provide solutions for a range of parameters after an initial offline training phase.
  • Adjoint Variables Require More Modes: A key observation is that the adjoint variables (used in the optimization and sensitivity analysis) often require a higher number of POD modes to be accurately represented compared to the state variables (velocity and pressure). This is attributed to their slower decay in singular values.

Key Results

The authors present numerical results for both the backward-facing step and lid-driven cavity flow test cases.

  • Significant Speedup: The combined DD-ROM approach leads to a substantial reduction in computational cost, both in terms of problem dimensions and the number of optimization iterations. For instance, they report reducing the dimension of the optimization problem by more than 10 times and using fewer iterations.
  • High Accuracy: The reduced-order models achieve accuracy comparable to the high-fidelity (full-order) solutions, with relative errors often in the range of a few percent.
  • Parameter Sweep Efficiency: The ROM allows for rapid exploration of parameter spaces. For example, they show results for different Reynolds numbers and viscosity values, demonstrating the model's ability to capture the physics across these variations.
  • Continuity at Interface: The numerical results show that the solutions obtained from different subdomains are continuous at the interface, indicating successful coupling.

Limitations and Future Work

  • High Reynolds Number Challenges: The authors note that the optimization solver can diverge for higher Reynolds numbers, particularly in the backward-facing step case. They suspect this is related to bifurcation phenomena and the "Coanda effect," suggesting that further stabilization techniques might be needed.
  • Non-Uniqueness of Solutions: For certain flow regimes (like higher Reynolds numbers), the problem might exhibit non-unique solutions, which adds complexity to the optimization process.
  • Future Work: The authors plan to extend this methodology to:
  • More complex time-dependent problems.
  • Multi-physics problems (e.g., Fluid-Structure Interaction).
  • Problems with more than two subdomains.
  • Scenarios where different types of numerical models are used in different subdomains.

Summary updated

Abstract

The aim of this work is to present a model reduction technique in the framework of optimal control problems for partial differential equations. We combine two approaches used for reducing the computational cost of the mathematical numerical models: domain-decomposition (DD) methods and reduced-order modelling (ROM). In particular, we consider an optimisation-based domain-decomposition algorithm for the parameter-dependent stationary incompressible Navier-Stokes equations. Firstly, the problem is described on the subdomains coupled at the interface and solved through an optimal control problem, which leads to the complete separation of the subdomain problems in the DD method. On top of that, a reduced model for the obtained optimal-control problem is built; the procedure is based on the Proper Orthogonal Decomposition technique and a further Galerkin projection. The presented methodology is tested on two fluid dynamics benchmarks: the stationary backward-facing step and lid-driven cavity flow. The numerical tests show a significant reduction of the computational costs in terms of both the problem dimensions and the number of optimisation iterations in the domain-decomposition algorithm.
BibTeX
@article{prusak2023optimisation,
  author = {Ivan Prusak and Monica Nonino and Davide Torlo and Francesco Ballarin and Gianluigi Rozza},
  title = {{An optimisation-based domain-decomposition reduced order model for the incompressible Navier-Stokes equations}},
  journal = {Comput. Math. Appl},
  year = {2023},
  doi = {10.1016/j.camwa.2023.09.039},
}
Figure