Implicit Velocity Correction Schemes for Scale-Resolving Simulations of Incompressible Flow: Stability, Accuracy, and Performance

This study demonstrates that implicit velocity correction schemes, specifically linear-implicit and sub-stepping methods, significantly enhance the stability and reduce the overall time-to-solution of scale-resolving simulations for complex high Reynolds number flows by up to a factor of eleven, while maintaining high accuracy even with time steps twenty times larger than explicit limits.

The Gist

Imagine you are trying to film a high-speed Formula 1 race, but you are doing it with a camera that is incredibly sensitive. To get a perfect, blur-free picture of the cars zooming around a complex track (with lots of curves and corners), you have to take thousands of photos every second. If you take even one photo too slowly, the image blurs, and you miss the action.

In the world of computer simulations, this "camera" is a mathematical model trying to predict how air flows around a car wing. The "photos" are time steps in the calculation.

The Problem:
Usually, these simulations use a "semi-implicit" method. Think of this as a cautious driver who checks the road every single inch before moving. It's very accurate, but because the track is so twisty and the car is so fast (high Reynolds number), the driver is forced to take tiny, baby steps. This means the simulation takes forever to finish, even on supercomputers. It's like trying to cross a river by hopping on every single pebble; you won't fall in, but you'll be there all day.

The Solution:
The authors of this paper asked: "What if we could take bigger steps without falling in?" They tested two new "implicit" driving strategies that allow the simulation to take much larger steps (time steps) while still staying stable.

Here is how they did it, using simple analogies:

1. The Two New Strategies

Strategy A: The "Sub-Stepping" Method (The Mini-Explorers)
Imagine you need to walk across a wide, muddy field in one big stride. Instead of just guessing where to put your foot, you send out a team of tiny "mini-explorers" ahead of you. They run through the mud in a simulated "fake time" to figure out exactly where the ground is safe. Once they report back, you take your big stride.

• Pros: You can take huge strides (time steps) because the mini-explorers checked the safety first.
• Cons: You have to pay the cost of sending out the explorers every time you move. It takes more effort per step, but you take fewer steps overall.

Strategy B: The "Linear-Implicit" Method (The Predictive GPS)
Imagine you are driving, and instead of looking at the road right in front of you, you use a GPS that predicts the road ahead based on where you were a moment ago. You assume the road won't change too wildly, so you steer based on that prediction.

• Pros: You can drive very fast (huge time steps) because you are trusting the prediction.
• Cons: Your steering wheel (the math) becomes more complex and harder to turn because the road conditions are changing. You have to recalculate your map constantly.

2. The Experiment: The "Imperial Front Wing"

To test these methods, the researchers used a very tricky test case: a Formula 1 front wing. This is a complex shape with curves, gaps, and air rushing over it at high speeds. It's the "Mount Everest" of aerodynamic simulations because the air behaves chaotically there.

They compared the old "baby-step" method against the two new "big-step" methods.

3. The Results: Speed vs. Accuracy

The Good News (Speed):
The new methods were game-changers for speed.

• The "Mini-Explorers" (Sub-stepping) allowed them to take steps 20 times larger than the old method.
• The "Predictive GPS" (Linear-implicit) allowed steps 100 times larger in some cases!
• The Bottom Line: Even though each step took more computer power to calculate, the total time to finish the simulation dropped by up to 11 times. They finished the race in a fraction of the time.

The Catch (Accuracy):
However, taking bigger steps isn't free.

• Small Steps (1x to 20x): The results were almost identical to the slow, cautious method. The simulation still correctly predicted where the air would separate from the wing and where turbulence would start.
• Huge Steps (100x): When they pushed the "Predictive GPS" to take steps 100 times larger, things started to get blurry. The simulation missed some of the subtle details of how the air transitioned from smooth to turbulent. It was like watching the race in slow motion but missing the details of the driver's hand movements.

4. The Takeaway: Finding the Sweet Spot

The paper concludes that you don't have to choose between "slow and perfect" or "fast and wrong." You can find a sweet spot.

• For the beginning of the simulation: When the flow is just starting up and changing rapidly, you can use the "big step" methods to get through the initial chaos quickly.
• For the final analysis: When you need to collect precise statistics (like average drag or lift), you might want to slow down and use smaller steps to ensure the details are perfect.

In Summary:
This paper is like a guide for race car engineers. It tells them: "You don't have to drive at 1 mph to stay safe. You can drive at 50 mph using our new navigation systems, and you'll still get to the finish line with a clear picture of the race—just make sure you don't drive at 100 mph unless you're okay with missing a few details."

It provides a roadmap for making complex weather and aerodynamic simulations faster, cheaper, and more practical for real-world engineering, like designing better cars or planes.

Technical Summary

1. Problem Statement

Scale-resolving simulations (SRS), such as Wall-Resolved Large Eddy Simulation (WRLES), are essential for capturing unsteady flow phenomena like vortex shedding and laminar-turbulent transition. However, their application to high Reynolds number flows over complex geometries is severely limited by computational cost.

• The Bottleneck: Explicit time-stepping schemes are constrained by the Courant–Friedrichs–Lewy (CFL) condition, requiring extremely small time steps ($\Delta t$) to maintain stability, particularly in regions with fine boundary layer meshes and high-order discretizations.
• The Trade-off: While implicit schemes relax CFL restrictions and allow larger time steps, they typically increase the computational cost per step ($T_{\Delta t}$) due to the need to solve non-linear or non-symmetric linear systems. Furthermore, increasing $\Delta t$ can degrade temporal accuracy, potentially obscuring critical flow physics.
• The Gap: There is a lack of systematic comparison between different implicit formulations specifically for complex, high-Reynolds number geometries using high-order spectral/hp element methods.

2. Methodology

The study utilizes the open-source Nektar++ framework, employing a high-order spectral/hp element discretization. The research focuses on velocity correction schemes, which split the incompressible Navier-Stokes equations into pressure Poisson and velocity Helmholtz/Advection-Diffusion-Reaction (ADR) problems.

Three schemes were compared:

1. Semi-implicit (Reference): Treats diffusion implicitly and advection explicitly. Subject to strict CFL limits.
2. Sub-stepping (Semi-Lagrangian): Resolves the advection term by solving an auxiliary unsteady advection equation in pseudo-time ($\tau$) using multiple sub-steps ($N_{\Delta \tau}$) within one physical time step.
3. Linear-implicit: Linearizes the advection operator by approximating $u^{n+1} \cdot \nabla u^{n+1} \approx \tilde{u} \cdot \nabla u^{n+1}$ (where $\tilde{u}$ is extrapolated). This converts the velocity step into a non-symmetric ADR problem.

Benchmark Case:

• Geometry: Extruded Imperial Front Wing (eIFW), a complex multi-element wing configuration with curved surfaces and ground effect.
• Flow Regime: $Re \approx 1.69 \times 10^5$ (based on chord length).
• Physics: Wall-resolved LES (iLES strategy) capturing laminar-turbulent transition.
• Metrics: Stability limits, temporal accuracy (integral forces, surface coefficients, spectral content), and time-to-solution ($T_{sol} = N_{\Delta t} \times T_{\Delta t}$).

3. Key Contributions

• Systematic Comparison: The first comprehensive evaluation of implicit velocity correction strategies (sub-stepping vs. linear-implicit) against a semi-implicit baseline for WRLES on a complex industrial geometry.
• Quantification of Trade-offs: Provides quantitative data on the balance between stability, accuracy degradation, and computational cost.
• Stability Limits: Demonstrates that implicit schemes can extend the stable time step size by 1 to 2 orders of magnitude compared to explicit limits.
• Performance Analysis: Identifies that while implicit schemes have higher per-step costs, they can reduce the total time-to-solution by up to a factor of 11 (specifically for the linear-implicit scheme at large time steps).

4. Key Results

A. Stability

• Semi-implicit: Stable up to $\Delta t \approx \Delta t_{CFL}$ (CFL $\approx 1.4$).
• Sub-stepping: Stable up to $\Delta t = 20 \Delta t_{CFL}$ (CFL $\approx 26$).
• Linear-implicit: Stable up to $\Delta t = 20 \Delta t_{CFL}$ (CFL $\approx 30$). With equal-order approximation spaces ($P_v = P_p = 4$), stability extended to $\Delta t = 100 \Delta t_{CFL}$ (CFL $\approx 166$).

B. Accuracy

• Integral Quantities (Lift/Drag): Both implicit schemes maintained high accuracy (within ~2-5% of reference) even at $\Delta t = 20 \Delta t_{CFL}$.
• Surface Quantities (Transition): Accuracy degradation was first observed in transition-sensitive metrics.
  • At $\Delta t = 20 \Delta t_{CFL}$, the laminar-turbulent transition location shifted upstream on the main plane.
  • At $\Delta t = 100 \Delta t_{CFL}$ (linear-implicit), the transition mechanism was significantly altered, and spectral peaks were no longer resolved correctly.
• Spectral Content: The linear-implicit scheme introduced numerical damping at high frequencies even at the reference time step. At large time steps, the spectral content changed, losing distinct peaks.

C. Computational Performance

• Per-Step Cost:
  • Sub-stepping: ~1.9x slower than semi-implicit (due to pseudo-time sub-steps).
  • Linear-implicit: ~5.2x slower (due to non-symmetric ADR solves and matrix reconstruction).
• Time-to-Solution ($T_{sol}$):
  • Break-even point: Linear-implicit becomes faster than semi-implicit at $\Delta t \approx 5 \Delta t_{CFL}$.
  • Maximum Speed-up:
    • Linear-implicit: Achieved a speed-up of ~8.6x at $\Delta t = 100 \Delta t_{CFL}$.
    • Sub-stepping: Achieved a modest speed-up of ~2x at $\Delta t = 20 \Delta t_{CFL}$. Performance was limited by the linear growth in required pseudo-time sub-steps and increasing pressure solver iterations.

5. Significance and Conclusions

The study provides critical guidance for selecting time integration strategies in large-scale incompressible flow solvers:

1. Implicit schemes are viable: They can drastically reduce time-to-solution for complex geometries without sacrificing the resolution of key flow statistics, provided the time step is chosen carefully.
2. Scheme Selection:
   • Linear-implicit is superior for rapidly advancing through the initial transient phase where high temporal accuracy is less critical, offering the highest potential speed-up.
   • Semi-implicit remains the most efficient and accurate choice for the statistically stationary phase where high-fidelity resolution of transition and turbulence is required.
3. Future Directions: The authors suggest combining these schemes with adaptive time stepping (large steps during transients, small steps for sampling) and utilizing matrix-free formulations to further reduce the overhead of the linear-implicit scheme.

In summary, the paper demonstrates that implicit velocity correction schemes offer a powerful pathway to making high-fidelity WRLES of complex industrial geometries computationally feasible, provided the trade-off between stability gains and accuracy loss is managed via appropriate time-step selection.