Abstract
We present a hybrid Reduced Order Model (ROM) that leverages recent developments in Physics-Informed Neural Networks (PINNs) to solve inverse problems for the Navier–Stokes equations (NSE). The proposed approach assumes the availability of simulated data for fluid dynamics fields. A POD-Galerkin ROM is constructed by applying Proper Orthogonal Decomposition (POD) to snapshot matrices and performing a Galerkin projection of the NSE onto the resulting reduced basis. We then derive a POD-Galerkin PINN ROM by employing deep neural networks to approximate reduced outputs, with time and model parameters as inputs. These networks incorporate the POD-Galerkin reduced equations directly into their loss functions. This framework enables the approximation of unknown physical constants and boundary conditions. We demonstrate the model's efficacy through two primary cases: steady flow around a backward step and unsteady turbulent flow around a surface-mounted cubic obstacle.
Introduction
Simulating fluid flow using the Navier-Stokes Equations (NSE) is computationally intensive, particularly for turbulent flows requiring millions of degrees of freedom. This computational burden becomes a critical bottleneck in "many-query" scenarios such as optimization, uncertainty quantification, and inverse problems—where researchers must infer unknown input parameters (like viscosity or inlet velocity) from observed output data.
While traditional Reduced Order Models (ROMs) reduce dimensionality by projecting physics onto a low-dimensional space, they are often "intrusive," requiring significant modifications to underlying solver codes, and can struggle with the non-linearities of turbulence. Conversely, pure data-driven machine learning often ignores the governing physical laws. This paper introduces a middle ground: the POD-Galerkin PINN ROM, a framework that is computationally efficient, data-driven, and strictly constrained by the laws of physics.
Methodology
The POD-Galerkin PINN ROM is developed through a four-step integration of classical fluid dynamics and modern deep learning:
- Dimensionality Reduction: Proper Orthogonal Decomposition (POD) is applied to "snapshots" of high-fidelity simulation data to identify a low-dimensional basis that captures the flow's essential features.
- Galerkin Projection: The Navier-Stokes equations are projected onto this reduced basis. This transforms complex partial differential equations (PDEs) into a simplified system of Ordinary Differential Equations (ODEs) or Differential-Algebraic Equations (DAEs).
- PINN Integration: Rather than using a traditional numerical integrator, a Deep Neural Network (DNN) is trained to predict the coefficients of the reduced basis over time or parameter space.
- Physics-Informed Loss: The network's loss function includes the residual of the reduced equations. If the network's predictions violate the projected physics, the loss increases. This allows the network to treat unknown physical constants (such as viscosity $\nu$) as trainable weights, optimizing them alongside the network parameters during training.
Experimental Setup and Datasets
The framework was validated using high-fidelity data generated from the open-source CFD library OpenFOAM. Three benchmark cases were evaluated:
- Steady Flow past a Backward Step: A laminar case used to validate 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) involving over 1.2 million cells, designed to test the limits of the model's efficiency.
Key Insights
Hierarchical Approximation
The strength of this approach lies in its two-level hierarchy. POD reduces spatial complexity, while the PINN handles the functional mapping. By training the PINN on the reduced equations rather than the full-order PDEs, the authors avoid the massive computational overhead typically associated with 3D PINNs.
Parameters as Trainable Weights
In the context of inverse problems, unknown physical constants are treated as weights within the neural network. As the network minimizes the physics-based loss to match observed data, it naturally converges on the true physical value of the unknown parameter.
Robustness to Incomplete Data
The model successfully reconstructed full pressure fields and lift/drag coefficients even when training data lacked pressure history, provided velocity data and the underlying physics (the reduced equations) were available.
Results
- Computational 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 unknown physical viscosity with an error of less than 1%.
- Reconstruction Capability: For the cylinder case, the model accurately reconstructed the time-history of lift and drag coefficients despite significant uncertainty in the initial output data.
- Training Efficiency: Because the PINN operates on the reduced system, it can be trained on a single GPU in minutes to hours. Training a standard PINN on the full 3D mesh of the cubic obstacle would be computationally prohibitive.
Limitations and Future Work
Despite its success, the framework has an "offline" cost associated with running initial high-fidelity simulations and training the PINN. Additionally, while 10–20 POD modes are sufficient for parameter estimation, accurately reconstructing complex outputs like lift and drag requires a significantly higher number of modes (60+). Future research will focus on extending this methodology to multi-physics problems, such as Fluid-Structure Interaction (FSI), where the benefits of reduced-order modeling are even more pronounced.
Summary updated