A Discrete Adjoint Gas-Kinetic Scheme for Aerodynamic Shape Optimization in Turbulent Continuum Flows

This paper presents a discrete adjoint gas-kinetic scheme for turbulent continuum flows, which is rigorously verified against a linearized solver and demonstrated to be highly efficient and accurate for aerodynamic shape optimization across various benchmark cases.

The Gist

Imagine you are a master chef trying to perfect a new recipe for a soufflé. You want it to be light, airy, and rise perfectly. But instead of guessing, you have a magical assistant that can tell you exactly how changing one ingredient (like adding a pinch more salt or baking it 30 seconds longer) will change the final taste.

This paper is about building a super-smart, high-speed "culinary assistant" for airplane designers.

Here is the breakdown of what the researchers did, using simple analogies:

1. The Problem: The "Trial and Error" Trap

Designing a better airplane wing is like trying to find the perfect path through a dense, foggy forest.

• The Old Way: Designers used to guess a shape, test it, see if it was slow or fast, and then guess again. If you had 1,000 parts of the wing to tweak, you'd have to run 1,000 separate tests just to see which direction to move. This is incredibly slow and expensive (like trying to taste every single grain of sand on a beach to find the sweetest one).
• The Goal: They wanted a way to look at the whole forest at once and instantly know exactly which way to step to get to the exit faster.

2. The Solution: The "Shadow Twin" (The Adjoint Method)

The researchers developed a new tool called a Discrete Adjoint Gas-Kinetic Scheme. That's a mouthful, so let's break it down:

• The "Gas-Kinetic" Part (The Microscope): Imagine air isn't just a smooth wind, but a swarm of billions of tiny, bouncy balls (molecules) hitting each other. The "Gas-Kinetic Scheme" is a super-detailed way of simulating how these tiny balls bounce, stick, and swirl around a wing. It's like watching a high-speed movie of every single molecule.
• The "Adjoint" Part (The Shadow Twin): This is the magic trick. Instead of running the simulation forward to see what happens, they run a "Shadow Twin" simulation backward.
  • Analogy: Imagine you are trying to figure out why a car crashed. You could watch the whole crash in slow motion (forward). Or, you could start at the wreckage and rewind the physics to see exactly which tire hit the curb first. The "Adjoint" method is that rewind button. It tells the designer: "If you move this curve on the wing just a tiny bit, the drag will drop by exactly this much."

3. The Innovation: Making it "Turbulent" and "Real"

Previous versions of this "Shadow Twin" tool worked well for smooth, calm air (like a gentle breeze). But real airplanes fly through turbulent air (like a bumpy ride in a storm).

• The Challenge: Turbulence is chaotic. It's like trying to rewind a video of a tornado. It's very hard to calculate the "backward" path because the math gets messy.
• The Breakthrough: The team built a new version of the tool that can handle this chaos. They used a technique called Algorithmic Differentiation (AD).
  • Analogy: Think of AD as a "smart recorder" that watches every single step the computer takes while calculating the wind. When it's time to go backward, the recorder plays the steps in reverse, perfectly undoing every calculation without losing accuracy. This allowed them to include the complex math of turbulence (the Spalart–Allmaras model) without breaking the tool.

4. The Proof: Three Test Drives

To prove their new tool works, they ran three different "driving tests":

1. The Reverse Engineering Test (Turbine Blades):

   • The Setup: They took a perfect turbine blade, messed it up on purpose, and asked the tool: "How do we fix it to look like the original?"
   • The Result: The tool fixed the blade in just 10 steps, recovering the original shape with 99.9% accuracy. It was like a sculptor instantly chiseling a broken statue back to perfection.

2. The Efficiency Test (Lift-to-Drag Ratio):

   • The Setup: They took a standard wing (NACA 0012) and asked: "How can we make it fly higher and faster without using more fuel?"
   • The Result: The tool reshaped the wing to be slightly asymmetrical (like a bird's wing). The result? The lift (upward force) doubled, while the drag (air resistance) stayed the same. It's like getting a sports car that goes twice as fast but uses the same amount of gas.

3. The Shockwave Tamer Test (Reducing Shock Strength):

   • The Setup: At high speeds, air creates a "sonic boom" or a shockwave on the wing, which wastes energy. They asked: "How can we smooth out this shock?"
   • The Result: The tool thinned the front of the wing slightly. This slowed down the air before it hit the shock, making the shockwave much weaker. It's like smoothing out a pothole so your car doesn't bounce as hard.

5. Why This Matters

This paper is a big deal because it combines extreme accuracy (watching the tiny molecules) with extreme speed (the backward shadow method).

• Before: Designing a new plane might take months of supercomputer time.
• Now: With this new tool, engineers can test thousands of design changes in the time it used to take to test just one.

In a nutshell: The researchers built a "time-traveling wind tunnel" that can instantly tell engineers exactly how to tweak a plane's shape to make it faster, quieter, and more fuel-efficient, even when flying through the most chaotic, stormy air.

Technical Summary

1. Problem Statement

Aerodynamic shape optimization is critical for reducing operating costs in aviation, but it faces significant computational challenges, particularly in turbulent continuum flows.

• Computational Cost: Traditional gradient-free methods scale poorly with the number of design variables. Gradient-based methods are more efficient but require accurate sensitivity analysis.
• Adjoint Complexity: Developing discrete adjoint solvers for complex flow solvers is difficult. Conventional finite-volume schemes often decouple inviscid and viscous fluxes (using Riemann solvers and central differencing, respectively), creating structural inconsistencies that complicate adjoint formulation, especially for turbulence models.
• Turbulence Modeling: Manually linearizing turbulence model equations (like Spalart–Allmaras) within a continuous adjoint framework is notoriously challenging and often requires simplifying assumptions (e.g., constant eddy viscosity).
• Gap: While Gas-Kinetic Schemes (GKS) offer a unified treatment of inviscid and viscous fluxes, their application to fully turbulent discrete adjoint shape optimization has remained largely unexplored.

2. Methodology

The authors developed a Discrete Adjoint Gas-Kinetic Scheme (GKS) solver using Algorithmic Differentiation (AD) to address the above challenges.

A. The Primal Solver: Gas-Kinetic Scheme (GKS)

• Foundation: Based on the Bhatnagar–Gross–Krook (BGK) model, which derives numerical fluxes from the integral solution of the kinetic equation.
• Unified Flux: Unlike conventional schemes, GKS computes inviscid and viscous fluxes simultaneously by taking velocity moments of a time-dependent distribution function at cell interfaces. This inherently captures both equilibrium and non-equilibrium states.
• Turbulence Model: The one-equation Spalart–Allmaras (SA) model is integrated to handle fully turbulent flows. The eddy viscosity is coupled directly into the collision time and flux calculations.
• Discretization: Uses a high-order, compact discretization in space and time with enhanced shock-capturing capabilities.

B. The Adjoint Solver Development

• Approach: The discrete adjoint solver was constructed using the backward mode of Algorithmic Differentiation (AD). This ensures exact consistency with the numerical discretization of the primal solver.
• Verification Strategy: The discrete adjoint solver was rigorously verified against a linearized GKS solver developed using the forward mode of AD.
  • Duality-Preserving Principle: The linearized solver (forward mode) and the adjoint solver (backward mode) are mathematically dual. They should share identical eigenvalues, asymptotic residual decay rates, and sensitivity convergence behaviors.
• Implementation: The solver handles the full chain of dependencies, including the flow governing equations, boundary conditions, the SA turbulence model, and the objective function.

C. Optimization Framework

• Parameterization: Shape perturbations are defined using Hicks-Henne hump functions.
• Mesh Deformation: A linear elastic method is used to deform the mesh while preserving quality.
• Optimizer: A steepest descent method with a constant step size updates the design variables.

3. Key Contributions

1. First Fully Turbulent Discrete Adjoint GKS: The paper presents the first implementation of a discrete adjoint solver for GKS that fully incorporates the Spalart–Allmaras turbulence model for continuum flows.
2. Unified Flux Advantage: By leveraging the GKS's unified treatment of inviscid and viscous fluxes, the complexity of formulating the adjoint system (particularly for viscous terms) is significantly reduced compared to conventional upwind schemes.
3. Rigorous Verification: The study provides a comprehensive verification protocol comparing the discrete adjoint and linearized solvers. It confirms:
   • Qualitative: Adjoint turbulence variables exhibit inverse trends compared to primal flow variables (a hallmark of correct adjoint implementation).
   • Quantitative: The solvers exhibit identical sensitivity convergence curves and asymptotic RMS residual decay rates.
   • Accuracy: At full convergence, the relative discrepancy in sensitivity predictions between the two methods is negligible.
4. Algorithmic Differentiation Integration: Demonstrates the effective use of AD (both forward and backward modes) to automate the derivation of complex adjoint equations for kinetic-theory-based solvers.

4. Results

The solver was validated and tested on three benchmark cases:

Case 1: Inverse Design of Turbine Blades (Durham Cascade)

• Objective: Recover a target Mach number distribution on a perturbed turbine blade.
• Result: The objective function (Mach number discrepancy) was reduced by 99.9% within 10 design cycles. The optimized geometry accurately reconstructed the target profile.

Case 2: Lift-to-Drag Ratio Enhancement (NACA 0012 Airfoil)

• Objective: Maximize the lift-to-drag ratio ($L/D$) at a 1° angle of attack.
• Result: After 10 cycles, lift increased by more than a factor of two while drag remained nearly constant. The optimizer produced an asymmetric airfoil shape, demonstrating the solver's ability to handle complex trade-offs in turbulent flow.

Case 3: Shock-Strength Reduction (NACA 0012 Airfoil)

• Objective: Minimize shock strength at $M_\infty = 0.8$ by minimizing entropy generation (using Euler equations with slip walls to isolate shock entropy).
• Result: Entropy generation was reduced by >55%. The pre-shock Mach number decreased from 1.2 to 1.1, resulting in a significantly weaker shock wave. The optimized shape featured reduced thickness near the leading edge to control flow acceleration.

5. Significance

• Computational Efficiency: The adjoint approach allows for sensitivity analysis with a computational cost largely independent of the number of design variables, making it feasible for high-dimensional aerodynamic optimization.
• Robustness for Turbulent Flows: By successfully integrating the SA turbulence model into the discrete adjoint framework, the method overcomes the limitations of continuous adjoint approaches that often struggle with turbulence linearization.
• Future Applicability: The unified nature of GKS combined with the robustness of the discrete adjoint method makes this framework highly suitable for complex, multi-physics aerodynamic design problems involving shocks, viscous effects, and turbulence, potentially extending to hypersonic and rarefied flow regimes in the future.

In conclusion, this work establishes a rigorous, efficient, and accurate framework for aerodynamic shape optimization in turbulent continuum flows, bridging the gap between kinetic theory-based solvers and modern gradient-based design optimization.