Solving compressible Navier-Stokes equations using the feature-enhanced neural network

This study extends the feature-enhanced neural network (FENN) to successfully solve forward and parametric problems involving compressible viscous flows governed by the Navier-Stokes equations, overcoming the limitations of existing physics-informed neural network methods that have previously failed in this challenging scenario.

The Gist

Imagine you are trying to predict how wind blows around a car, an airplane wing, or even a spinning baseball. In the world of physics, this is described by a very complicated set of rules called the Navier-Stokes equations. Think of these equations as the "rulebook" for how fluids (like air and water) move.

For decades, solving these rules has been like trying to navigate a maze in the dark. Traditional computer methods work, but they are slow, expensive, and require building a detailed 3D grid (like a mesh) around the object, which is a lot of work.

Enter Artificial Intelligence (AI). Scientists have been trying to teach AI to solve these rules directly, skipping the messy grid. This approach is called Physics-Informed Neural Networks (PINNs). It's like teaching a student the rules of the game so they can predict the outcome without needing to watch every single play.

The Problem: The AI is Getting Stuck

While this AI method worked great for simple scenarios (like water flowing smoothly or air moving without friction), it hit a wall when things got complicated. Specifically, it failed miserably when trying to simulate compressible viscous flows.

Let's break that down with an analogy:

• Inviscid/Incompressible: Imagine blowing on a piece of paper. The air moves, but it doesn't squish together, and it doesn't stick to the paper much. The AI could handle this.
• Compressible Viscous: Now imagine a supersonic jet engine. The air is being squished (compressed) and it's sticky (viscous) due to friction. It's chaotic, turbulent, and complex. When the researchers tried to use the standard AI on this, it was like giving a calculator to a toddler and asking them to solve advanced calculus. The AI just couldn't figure out the physics; it got lost in the math.

The Solution: The "Feature-Enhanced" Tutor

The authors of this paper, Jiahao Song, Wenbo Cao, and Weiwei Zhang, decided to give the AI a "cheat sheet" or a set of training wheels. They developed a new method called FENN (Feature-Enhanced Neural Network).

Here is how it works, using a creative metaphor:

The "Feature" is like a Compass.
Imagine you are trying to teach a robot to walk around a house.

• Standard AI (PINN): You tell the robot, "Here are the rules of walking. Figure out where the walls are." The robot stumbles, bumps into things, and gets confused about the distance to the walls.
• Feature-Enhanced AI (FENN): You give the robot a compass that always points to the nearest wall. You also tell it, "Hey, the closer you get to the wall, the more careful you need to be."

In this paper, the "compass" is a specific piece of data: the shortest distance from any point in the air to the surface of the airplane wing.

By feeding this distance directly into the AI's brain before it tries to solve the complex math, the AI suddenly understands the geometry of the problem. It's no longer guessing where the wing is; it knows exactly how close it is at every single point.

What Did They Achieve?

The researchers tested this new "compass-equipped" AI on four different challenging scenarios:

1. Different Speeds: From slow to fast (but not supersonic yet).
2. Different Shapes: Three different types of airplane wings (NACA airfoils).
3. Different Angles: Tilting the wing up and down.
4. The "Big Angle" Test: They tilted the wing so far (20 degrees) that the air actually separated and created swirling vortices (turbulence). This is the "hard mode" of aerodynamics.

The Result:

• The old AI (PINN) failed. It produced garbage results, unable to predict the pressure or the swirling air.
• The new AI (FENN) succeeded. It accurately predicted how the air moved, where the pressure was high or low, and even captured the swirling vortices created by the steep angle.

The "Superpower": One Training, Infinite Answers

The paper also showed off a second superpower. Usually, if an engineer wants to know how a wing performs at 5 degrees, 10 degrees, and 15 degrees, they have to run the simulation three separate times.

With FENN, they treated the angle of the wing as just another input variable, like a dial they could turn. They trained the AI once, and suddenly, the AI could instantly predict the airflow for any angle between -5 and +5 degrees. It's like training a chef once to make a soup, and then being able to instantly adjust the recipe for any amount of salt or pepper without starting over.

The Bottom Line

This paper is a breakthrough because it's the first time a method like this has successfully solved the "hard mode" of fluid dynamics (compressible, sticky air) using AI.

In summary:

• The Problem: AI was too dumb to handle complex, sticky, squishy airflows around wings.
• The Fix: They gave the AI a "distance sensor" (a feature) so it could "see" the wing clearly.
• The Win: The AI can now solve these complex physics problems accurately and can even predict many different flight angles in a single go.

This is a huge step forward. It suggests that in the future, we might be able to design better airplanes and cars using AI that is fast, cheap, and incredibly smart, rather than relying on slow, expensive supercomputers.

Technical Summary

1. Problem Statement

The paper addresses a critical gap in the application of Physics-Informed Neural Networks (PINNs) to fluid mechanics. While PINNs have shown success in solving inviscid flows and incompressible viscous flows, they have historically struggled with compressible viscous flows. These flows are governed by the full compressible Navier-Stokes (CNS) equations, which include continuity, momentum, and energy equations coupled with complex thermodynamic relations. Solving these equations is essential for engineering applications (e.g., aircraft design) but remains a significant challenge for PINNs due to the high complexity, stiffness, and multi-scale nature of the governing equations.

2. Methodology

The authors propose and validate a Feature-Enhanced Neural Network (FENN) to solve the steady, two-dimensional, dimensionless compressible Navier-Stokes equations.

• Governing Equations: The study solves the CNS equations for density ($\rho$), velocity components ($u, v$), and temperature ($T$), incorporating Reynolds number ($Re$), Mach number ($Ma$), Prandtl number ($Pr$), and specific heat ratio ($\gamma$).
• Standard PINN Limitations: Standard PINNs minimize a loss function composed of boundary condition residuals ($\mathcal{L}_{bc}$) and PDE residuals ($\mathcal{L}_{r}$). The authors utilize a volume-weighted PDE loss (introduced in their previous work) to mitigate ill-conditioning caused by non-uniform collocation points. However, even with this improvement, standard PINNs failed to converge accurately for the tested compressible viscous cases.
• Feature-Enhanced Neural Network (FENN):
  • Concept: FENN improves approximation capabilities by introducing specific "features" highly correlated with the flow physics into the network inputs.
  • Feature Selection: The authors selected the shortest distance ($D$) from collocation points to the airfoil wall as the primary feature. This geometric feature was chosen over physical features (like potential flow solutions) because it offers a dominant enhancement effect with significantly lower computational cost.
  • Offline Training: To avoid errors in automatic differentiation (where treating features as independent inputs would yield incorrect partial derivatives), the feature networks are trained offline in a supervised manner. These networks learn the mapping $D(x,y)$, allowing the main network to correctly compute gradients via the chain rule during the main training phase.
  • Network Architecture: The main network uses $tanh$ activation, Xavier initialization, and is optimized using the L-BFGS algorithm.

3. Key Contributions

• First Successful Application: To the best of the authors' knowledge, this is the first time a PINN-like method has successfully solved both forward and parametric problems involving compressible viscous flows.
• Overcoming PINN Failure: The study demonstrates that while state-of-the-art PINNs (even with volume weighting) fail to solve these specific equations, the addition of geometric features (FENN) enables accurate solutions.
• Parametric Capability: The method successfully extends to parametric problems, where the angle of attack ($\alpha$) is treated as an input variable. This allows the network to predict flow fields across a range of angles in a single training run, a task that is computationally expensive for traditional numerical methods.

4. Results

The effectiveness of FENN was validated through five specific test cases:

• Forward Problems (4 Cases):
  • Conditions: Four scenarios involving different airfoils (NACA0012, NACA2418, NACA4424), Mach numbers (0.5–0.7), Reynolds numbers (400–800), and angles of attack (including a high angle of 20°).
  • Performance:
    • Convergence: FENN converged significantly faster than standard PINNs, with boundary and PDE losses approximately one order of magnitude lower.
    • Accuracy: Standard PINNs produced large discrepancies compared to Finite Volume Method (FVM) reference solutions. In contrast, FENN results showed small deviations from the reference.
    • Separated Flow: In Case 2 (NACA0012 at 20° AoA), FENN accurately captured the separation vortices, matching the FVM reference, whereas PINNs failed to resolve the flow physics correctly.
• Parametric Problem:
  • Setup: A parametric study on the NACA0012 airfoil ($Ma=0.4, Re=1000$) with the angle of attack varying from $-5^\circ$ to $5^\circ$.
  • Outcome: FENN successfully generated flow fields for the entire range of angles in a single training session. The pressure coefficients ($C_p$) at specific angles matched the sequential FVM solutions with high fidelity.

5. Significance and Limitations

• Significance: This work bridges a major gap between deep learning-based solvers and high-fidelity engineering fluid dynamics. It proves that PINN-like methods can handle the complexity of compressible viscous flows, opening the door for their use in inverse problems, design optimization, and real-time flow control where traditional mesh-based solvers are too slow.
• Limitations:
  • The current study is limited to subsonic laminar flows.
  • It does not yet address flows involving shock waves or turbulence, which present even greater challenges for neural network solvers.
• Future Work: The authors suggest that advanced techniques will be needed to extend this framework to transonic/supersonic regimes and turbulent flows.

In conclusion, the paper establishes FENN as a robust tool for solving compressible Navier-Stokes equations, significantly outperforming standard PINNs and offering a viable alternative to traditional numerical methods for parametric flow analysis.