Adaptive mesh refinement for global stability analysis of transitional flows

This paper introduces a novel adaptive mesh refinement framework that utilizes three independent, error-optimized meshes for the base flow, direct, and adjoint solutions to enhance the accuracy of global stability analysis for transitional incompressible flows, as validated by the two-dimensional flow past a circular cylinder.

The Gist

Imagine you are trying to predict exactly how a leaf will flutter as it falls from a tree. To do this, you build a super-computer simulation. But here's the catch: if your simulation is too "blurry" (low resolution), the computer might invent fake wind gusts that don't exist, causing the leaf to spin wildly when it should just drift gently. This is a problem known as numerical noise, and in fluid dynamics, it can trick scientists into thinking a flow is unstable when it's actually calm, or vice versa.

This paper introduces a clever new way to fix this problem using Adaptive Mesh Refinement (AMR).

The Problem: The "One-Size-Fits-All" Map

Traditionally, when scientists simulate fluid flow (like water around a pipe or air around a cylinder), they use a grid, like a map made of tiny squares.

• The Old Way: You have to decide beforehand how detailed your map needs to be. If you make the whole map super-detailed, you waste massive amounts of computer power on empty space. If you make it too coarse, you miss the important details where the action happens (like the swirling wake behind a cylinder).
• The Risk: In stability analysis (predicting when a smooth flow turns turbulent), getting the detail wrong is fatal. A blurry map might trigger a "fake" transition to turbulence, ruining the experiment.

The Solution: A Smart, Shape-Shifting Grid

The authors propose a technique where the computer grid acts like a smart, shape-shifting camera lens. Instead of keeping the whole picture equally sharp, it zooms in only where the action is happening and stays blurry where nothing is going on.

But here is the twist in this paper: They realized that different parts of the physics need different "zoom levels."

Think of the simulation as a three-act play:

1. Act 1: The Base Flow (The Stage Setup). This is the steady, calm flow before anything goes wrong. It needs a specific level of detail to look right.
2. Act 2: The Direct Perturbation (The Disturbance). This is a tiny ripple introduced to see how the flow reacts. The "ripples" might live in a different part of the flow than the steady stage setup.
3. Act 3: The Adjoint (The Detective). This is a mathematical tool that tells you where the flow is most sensitive to changes. It's like a detective looking for the "weak spot" in the system. The detective might need to focus on a completely different area than the ripple.

The Innovation:
Instead of using one grid for all three acts, the authors created three independent, custom-tailored grids.

• They use a "Spectral Error Indicator" (SEI) as their quality control inspector. This inspector looks at the math and says, "Hey, this area is too blurry; we need more pixels here," or "This area is fine, we can save pixels there."
• The computer refines the grid specifically for the Base Flow, then resets and refines a new grid specifically for the Direct Perturbation, and finally a third grid for the Adjoint.

The Analogy: The Three Photographers

Imagine you are photographing a complex event with three different photographers:

1. Photographer A (The Base Flow) is taking a wide shot of the whole stadium. They need a grid that captures the crowd and the field evenly.
2. Photographer B (The Direct Solution) is zooming in on a specific player running a sprint. They need a super-sharp grid just around that player's legs and the track.
3. Photographer C (The Adjoint/Detective) is looking for the one person in the crowd who could cause a riot. They need a grid focused entirely on the VIP section, ignoring the rest.

In the past, scientists forced all three photographers to use the same camera settings and the same lens. The result was either a blurry sprint or a wasted photo of the empty VIP section.

This paper says: "Let's give each photographer their own custom lens and focus."

The Results: Sharper Pictures, Less Waste

The team tested this on the classic problem of flow past a circular cylinder (like wind hitting a pole).

• The Test: They started with a coarse, blurry grid. The computer couldn't find the correct "unstable" frequency (the rhythm at which the flow starts to wiggle).
• The Refinement: They let the AMR technique zoom in on the wake behind the cylinder.
• The Outcome: With the new, custom grids, they found the correct frequency with incredible precision (accurate to 9 decimal places!).
• The Bonus: They achieved this high precision using half the number of grid points compared to a traditional, uniformly high-resolution grid. This means they saved a huge amount of computer time and energy.

Why This Matters

This is a game-changer for transitional flows (flows that are on the edge of becoming turbulent).

• Safety: In engineering (airplanes, bridges, pipelines), we need to know exactly when a smooth flow will turn chaotic. If our math is too blurry, we might design a bridge that fails because we missed a subtle instability.
• Efficiency: By only refining the grid where it's absolutely necessary, we can solve much more complex 3D problems that were previously too expensive for computers to handle.

In short, this paper teaches us that to understand the delicate dance of fluids, we don't need to look at everything with the same intensity. We need to know where to look, and this new "smart grid" technique does exactly that, acting as a magnifying glass that moves to the most interesting parts of the flow.

Technical Summary

1. Problem Statement

Global stability analysis of transitional flows requires solving three distinct sets of equations: the non-linear base flow, the linear direct perturbation, and the linear adjoint perturbation.

• The Challenge: Designing an accurate computational mesh for these flows is critical. Inadequate resolution introduces numerical noise that can trigger premature transition or mask instability mechanisms. Conversely, uniform high-resolution meshes are computationally expensive.
• The Gap: While Adaptive Mesh Refinement (AMR) is established for turbulent flow simulations, it has not been previously applied to global stability analysis. Existing methods often rely on a single static mesh for all three solution stages, which may not be optimal for capturing the specific features of the base flow versus the eigenmodes (which often have different spatial characteristics, e.g., receptivity vs. instability growth).
• Specific Issue: In complex geometries (like a 180° bend pipe), low-resolution meshes can cause numerical instabilities that prevent the extraction of a stable base flow, leading to divergence.

2. Methodology

The authors implemented a novel workflow using the open-source spectral element code Nek5000. The core innovation is the application of isotropic $h$-refinement (subdividing elements) independently for the three required solution stages.

Numerical Framework

• Discretization: Spectral Element Method (SEM) with $P_N-P_{N-2}$ formulation (polynomial order $N=7$).
• Time Integration: Third-order implicit backward differentiation (BDF) with de-aliasing.
• AMR Strategy:
  • Error Indicator: The Spectral Error Indicator (SEI) by Mavriplis (1989) is used. It estimates truncation and quadrature errors based on the decay of spectral coefficients.
  • Refinement Logic: The mesh is refined based on the time-averaged SEI of the specific solution field being computed.
  • Non-conforming Interfaces: The implementation handles hanging nodes using a parent-to-children interpolating operator, ensuring continuity without introducing instabilities.

The Workflow

The process involves three distinct mesh generation phases:

1. Non-linear Base Flow:
   • The incompressible Navier-Stokes equations are solved using Selective Frequency Damping (SFD) to stabilize unstable steady states.
   • The SEI is calculated on the velocity field ($U$).
   • The mesh is refined iteratively until the base flow converges robustly. This results in Mesh A.
2. Linear Direct Solution:
   • The base flow is spectrally interpolated onto a new grid (Mesh A2).
   • The linearized direct Navier-Stokes equations are solved with spatially uncorrelated noise.
   • The SEI is calculated on the perturbation velocity field ($u'$).
   • The mesh is refined to capture the most unstable eigenmode, resulting in Mesh B.
3. Linear Adjoint (Dual) Solution:
   • The same process is repeated for the adjoint equations to determine receptivity, resulting in Mesh C.
   • Note: The authors emphasize that Mesh B and Mesh C often differ significantly because direct modes (instability growth) and adjoint modes (receptivity) localize energy in different spatial regions.

3. Key Contributions

• First Application of AMR to Global Stability: This is the first study to apply AMR specifically for the global stability analysis of transitional flows, moving beyond its traditional use in turbulence.
• Independent Mesh Design: The framework allows for three independent, optimized meshes (one for the base flow, one for the direct eigenmode, and one for the adjoint eigenmode), minimizing truncation errors specific to each physical phenomenon.
• Code Adaptation: Significant modifications were made to Nek5000 to support AMR within the linear solver and stability tools (KTH framework), specifically handling non-conforming interfaces for eigenvalue calculations.
• Validation Protocol: A rigorous workflow is established for determining the optimal transient time ($T_d$) before starting refinement to avoid over-refining transient noise while preventing numerical transition on coarse grids.

4. Results

The methodology was validated using the 2D flow past a circular cylinder at $Re_D = 50$ (slightly above the critical $Re_{cr} \approx 47$).

• Base Flow: The AMR successfully captured the wake structure. The SEI focused refinement around the cylinder surface and downstream in the wake, matching previous steady-flow AMR results.
• Eigenmode Separation:
  • The Direct Mode (instability) showed skew-symmetry and was concentrated in the near-wake.
  • The Adjoint Mode (receptivity) was highly localized near the cylinder surface (upper and lower sides) and decayed rapidly upstream/downstream.
  • Significance: This spatial separation confirmed the necessity of distinct meshes; a single mesh cannot optimally resolve both.
• Spectral Convergence:
  • Mesh 1 (Coarse, ~300 elements): Failed to capture the most unstable mode correctly and showed discrepancies between direct and adjoint eigenvalues (violating linear algebra theory).
  • Mesh 2 (~1200 elements): Captured the growth rate ($\sigma$) and frequency ($\omega$) with $10^{-4}$ accuracy.
  • Mesh 3 (~3000 elements): Achieved excellent agreement ($<10^{-9}$ difference) with a highly resolved reference conforming mesh (~5000 elements).
• Efficiency: The refined AMR mesh achieved the same accuracy as the reference mesh using roughly half the number of grid points, demonstrating significant computational savings.

5. Significance

• Robustness: The study proves that non-conforming interfaces introduced by AMR do not generate spurious instabilities in stability analysis, a critical requirement for accurate eigenvalue computation.
• Physical Insight: The ability to resolve the distinct spatial structures of direct and adjoint modes provides deeper insight into flow control and receptivity mechanisms.
• Scalability: The computational savings are expected to be even more pronounced in complex 3D transitional flows (e.g., bent pipes), where uniform refinement is prohibitively expensive.
• Future Outlook: The authors suggest future work on handling multiple meshes simultaneously for multiple eigenmodes in the Arnoldi method and developing refinement criteria based directly on eigenvalue quality rather than just error indicators.

In conclusion, this work establishes a robust, efficient, and accurate framework for global stability analysis, demonstrating that adaptive mesh refinement is essential for capturing the nuanced physics of transitional flows without the prohibitive cost of uniform high-resolution grids.