POD-Galerkin reduced order models and physics-informed neural networks for solving inverse problems for the Navier-Stokes equations

Saddam Hijazi, Melina Freitag, Niels Landwehr

This paper presents a hybrid framework that combines the physical rigor of classical Reduced Order Models (ROMs) with the flexible, data-driven power of Physics-Informed Neural Networks (PINNs). By integrating these two worlds, the authors create a "POD-Galerkin PINN ROM" capable of solving complex inverse problems in fluid dynamics—such as inferring unknown viscosity or boundary conditions—at a fraction of the computational cost of traditional simulations.

TL;DR

The authors propose a novel Reduced Order Model that uses Physics-Informed Neural Networks to solve the reduced-order equations of fluid flow. This approach allows for massive speedups (up to $10^6$) in solving inverse problems, enabling researchers to accurately estimate unknown physical parameters in turbulent 3D flows using limited data.

What Problem Does This Paper Address?

Simulating fluid flow using the Navier-Stokes Equations (NSE) is computationally expensive, especially for turbulent flows which require millions of degrees of freedom. This becomes a bottleneck for "many-query" scenarios like optimization, uncertainty quantification, or inverse problems (where you have the output data but need to find the input parameters, like viscosity or inlet velocity).

While standard Reduced Order Models (ROMs) help by projecting the physics onto a low-dimensional space, they can be "intrusive"—requiring deep changes to the underlying solver code—and often struggle with the non-linearities of turbulence. On the other hand, pure "black-box" machine learning often ignores the underlying physics. This paper seeks a middle ground: a model that is fast, data-driven, yet strictly constrained by the laws of physics.

Methods

The authors develop the POD-Galerkin PINN ROM through a multi-step process:

  1. Dimensionality Reduction: They use Proper Orthogonal Decomposition (POD) on "snapshots" of high-fidelity simulation data to find a low-dimensional basis that captures the essential features of the flow.
  2. Galerkin Projection: They project the Navier-Stokes equations onto this reduced basis, turning complex partial differential equations (PDEs) into a much simpler system of Ordinary Differential Equations (ODEs) or Differential-Algebraic Equations (DAEs).
  3. PINN Integration: Instead of using a traditional numerical integrator, they train a Deep Neural Network to predict the coefficients of the reduced basis.
  4. Physics-Informed Loss: The "secret sauce" is the loss function. It doesn't just look at data labels; it includes the residual of the reduced equations. If the network's predictions violate the projected physics, the loss increases. This allows the network to "learn" unknown physical constants (like viscosity $\nu$) as trainable weights during the optimization process.

Datasets

The models were trained and evaluated using high-fidelity data generated from the open-source CFD library OpenFOAM. Three benchmark cases were used:

  • Steady Flow past a Backward Step: A laminar case used to test basic parameter identification.
  • 2D Flow around a Circular Cylinder: A turbulent case (using RANS modeling) with incomplete data, testing the model's ability to recover missing pressure history.
  • 3D Flow around a Surface-Mounted Cube: A complex, high-dimensional turbulent case (using LES modeling) with over 1.2 million cells, testing the limits of the model's efficiency.

Key Insights and Ideas

  • Two Levels of Approximation: The genius of the approach is the hierarchy. First, POD reduces the spatial complexity. Second, the PINN handles the functional mapping. By training the PINN on the reduced equations rather than the full PDEs, the authors avoid the massive computational overhead usually associated with PINNs.
  • Parameters as Trainable Weights: In an inverse problem, the unknown physical constant (e.g., viscosity) is treated just like a weight in the neural network. As the network minimizes the physics-based loss, it naturally converges on the true physical value of that constant.
  • Robustness to Incomplete Data: The model successfully reconstructed full pressure fields and lift/drag coefficients even when the training data was missing the pressure history, provided the velocity data and the underlying physics (the reduced equations) were available.

Key Results

  • Massive Speedup: The online prediction phase achieved speedup factors of up to $10^6$ compared to the Full Order Model (FOM).
  • High Accuracy: In the 3D turbulent cube case, the model identified the unknown physical viscosity with an error of less than 1%.
  • Reconstruction: For the cylinder case, the model accurately reconstructed the time-history of lift and drag coefficients despite significant uncertainty and incomplete output data.
  • Efficiency in Training: By using the reduced system, the authors could train the PINN on a single GPU in a matter of minutes to hours, whereas training a PINN on the full 3D mesh would be practically impossible.

Limitations and Future Work

  • Offline Cost: While the "online" (prediction) phase is nearly instantaneous, the "offline" phase (running the initial high-fidelity simulations and training the PINN) is still significant.
  • Mode Sensitivity: The authors noted that while 10-20 modes are enough for inverse parameter estimation, accurately reconstructing complex outputs like lift and drag requires a much higher number of modes (60+).
  • Data Dependency: The model still relies on having a representative set of initial snapshots. If the initial data doesn't capture the relevant physics, the ROM will struggle.
  • Future Directions: The authors suggest extending this to more complex multi-physics problems, such as Fluid-Structure Interaction (FSI), where the benefits of reduced-order modeling are even more pronounced.

Summary updated

Abstract

We present a Reduced Order Model (ROM) which exploits recent developments in Physics Informed Neural Networks (PINNs) for solving inverse problems for the Navier–Stokes equations (NSE). In the proposed approach, the presence of simulated data for the fluid dynamics fields is assumed. A POD-Galerkin ROM is then constructed by applying POD on the snapshots matrices of the fluid fields and performing a Galerkin projection of the NSE (or the modified equations in case of turbulence modeling) onto the POD reduced basis. A POD-Galerkin PINN ROM is then derived by introducing deep neural networks which approximate the reduced outputs with the input being time and/or parameters of the model. The neural networks incorporate the physical equations (the POD-Galerkin reduced equations) into their structure as part of the loss function. Using this approach, the reduced model is able to approximate unknown parameters such as physical constants or the boundary conditions. A demonstration of the applicability of the proposed ROM is illustrated by three cases which are the steady flow around a backward step, the flow around a circular cylinder and the unsteady turbulent flow around a surface mounted cubic obstacle.
BibTeX
@inproceedings{hijazi2023pod,
  author = {Saddam Hijazi and Melina Freitag and Niels Landwehr},
  title = {{POD-Galerkin reduced order models and physics-informed neural networks for solving inverse problems for the Navier-Stokes equations}},
  booktitle = {Adv. Model. Simul. Eng. Sci},
  year = {2023},
  url = {https://openreview.net/forum?id=lKQkIJU41r4},
}

Loading related content...

Figure