Singular layer Physics Informed Neural Network method for Plane Parallel Flows

Teng-Yuan Chang, Gung-Min Gie, Youngjoon Hong, Chang-Yeol Jung

Abstract

This paper introduces singular layer Physics-Informed Neural Networks (sl-PINNs), a semi-analytic framework designed to predict the stiff solutions of plane-parallel flows under conditions of small viscosity. Standard PINNs often struggle with the sharp gradients found in boundary layers, where the solution changes abruptly. By leveraging boundary layer analysis, we derive a "corrector" that describes the singular behavior of viscous flow. This corrector is then integrated into the PINN architecture—either through explicit embedding or implicit training. Numerical experiments demonstrate that sl-PINNs provide stable, highly accurate predictions for plane-parallel flows even at viscosities as low as $\epsilon = 10^{-8}$, where traditional PINNs fail to converge.

Introduction

Physics-Informed Neural Networks (PINNs) have emerged as a powerful tool for solving partial differential equations (PDEs) by embedding physical laws directly into the loss function of a neural network. However, a significant challenge remains: the "stiff" problem. In fluid dynamics, as viscosity ($\epsilon$) approaches zero, the transition from the fluid's velocity at a boundary to its velocity in the main channel occurs across an extremely thin region known as the boundary layer.

For standard PINNs, these regions are a primary source of failure. Capturing such sharp gradients requires a prohibitive number of collocation points and iterations. Furthermore, the loss functions associated with these equations often become unbounded as $\epsilon$ decreases, leading to numerical instability. This paper presents sl-PINNs as a solution to these limitations by incorporating asymptotic analysis into the neural network's structural design.

Methodology: The sl-PINN Framework

The core innovation of the sl-PINN is the use of a semi-analytic corrector. Rather than forcing the neural network to learn the entire complex solution from scratch, the problem is decomposed into two manageable parts:

  1. The Inviscid Limit ($u^0$): This represents the solution to the equations in the absence of viscosity.
  2. The Corrector ($\phi$): This is a mathematical function derived from boundary layer theory that specifically accounts for the singular behavior (the sharp "snap") near the boundaries.

Architectural Integration

The authors propose two primary methods for integrating this corrector into the PINN architecture:

  • Explicit Embedding: For components like primary velocity ($u_1$), the mathematical expression of the corrector is derived and baked directly into the network's output layer. This ensures the boundary conditions are satisfied by construction.
  • Implicit Training: For more complex components such as transverse velocity ($u_2$) or vorticity ($\omega$), where explicit formulas are less accessible, a secondary "corrector" network is trained simultaneously with the main PINN. This allows the model to learn the singular behavior implicitly while maintaining the stability provided by the corrector framework.

By restructuring the problem this way, the loss function remains bounded even as $\epsilon \to 0$, allowing the optimizer to converge on a stable minimum where standard methods would diverge.

Numerical Experiments and Results

The sl-PINN was tested using synthetic benchmark problems based on the stationary Navier-Stokes equations (NSE) for plane-parallel flows in a periodic channel. The models were evaluated across a range of viscosities, from $\epsilon = 10^{-3}$ down to $\epsilon = 10^{-8}$.

Performance Comparison

The performance gap between standard PINNs and sl-PINNs is substantial:

  • Accuracy at Extremes: At a viscosity of $\epsilon = 10^{-8}$, standard PINNs failed completely, yielding relative $L^2$ errors exceeding 6900%. In contrast, the sl-PINN maintained a relative error of approximately $1.17 \times 10^{-5}$.
  • Convergence and Efficiency: sl-PINNs demonstrated remarkable efficiency, often reaching minimum loss in fewer than 1,000 iterations. Standard PINNs, conversely, often failed to reach a stable solution even after 20,000 iterations.
  • Vorticity Resolution: The sl-PINN successfully captured the "approximation of identity" behavior of vorticity near the walls, a feature that typically requires massive mesh refinement in traditional numerical solvers.

Discussion and Limitations

While sl-PINNs represent a significant leap forward for high-speed flow simulation, the method currently relies on specific geometric and theoretical prerequisites. The correctors used in this study are tailored for plane-parallel flows in periodic channels; extending this to complex geometries, such as flow over an airfoil, would require the derivation of new asymptotic correctors. Additionally, the current work focuses on stationary flows, leaving the challenge of turbulent, time-dependent flows for future research.

Conclusion

The sl-PINN method demonstrates that the integration of classical asymptotic analysis with modern machine learning can overcome the inherent limitations of standard PINNs in high-Reynolds-number regimes. By using physics not just as a constraint but as an architectural guide, sl-PINNs offer a robust and efficient path for solving singular perturbation problems in fluid dynamics.

Summary updated

Abstract

We construct in this article the semi-analytic Physics Informed Neural Networks (PINNs), called {\em singular layer PINNs} (or {\em sl-PINNs}), that are suitable to predict the stiff solutions of plane-parallel flows at a small viscosity. Recalling the boundary layer analysis, we first find the corrector for the problem which describes the singular behavior of the viscous flow inside boundary layers. Then, using the components of the corrector and its curl, we build our new {\em sl-PINN} predictions for the velocity and the vorticity by either embedding the explicit expression of the corrector (or its curl) in the structure of PINNs or by training the implicit parts of the corrector (or its curl) together with the PINN predictions. Numerical experiments confirm that our new {\em sl-PINNs} produce stable and accurate predicted solutions for the plane-parallel flows at a small viscosity.
BibTeX
@inproceedings{chang2024singular,
  author = {Teng-Yuan Chang and Gung-Min Gie and Youngjoon Hong and Chang-Yeol Jung},
  title = {{Singular layer Physics Informed Neural Network method for Plane Parallel Flows}},
  booktitle = {Comput. Math. Appl},
  year = {2024},
  url = {https://openreview.net/forum?id=9Pfh9eMtzl},
}
Figure