Volume Term Adaptivity for Discontinuous Galerkin Schemes

This paper introduces "v-adaptivity," a novel strategy for high-order discontinuous Galerkin schemes that dynamically switches between weak form and flux-differencing volume discretizations at each Runge-Kutta stage based on entropy indicators to enhance robustness, efficiency, and accuracy in solving time-dependent compressible flow problems.

The Gist

Imagine you are trying to simulate a complex physical event on a computer, like a storm swirling over a city or a shockwave from an explosion. To do this, scientists use a powerful mathematical tool called the Discontinuous Galerkin (DG) method. Think of the computer screen as a giant jigsaw puzzle made of thousands of tiny tiles (cells). The computer calculates what happens inside each tile to predict the future of the whole system.

However, there's a catch. To get a perfect, stable, and accurate picture, the computer has to choose between two different ways of calculating what happens inside each tile:

1. The "Cheap & Fast" Way (Weak Form): This is like taking a quick, rough guess. It's very fast to compute, but sometimes it's a bit "wobbly." If the simulation gets too chaotic (like a shockwave), this method might make the numbers explode, causing the whole simulation to crash.
2. The "Expensive & Safe" Way (Flux-Differencing): This is like doing a super-detailed, rigorous check. It's mathematically rock-solid and never crashes, even in chaos. But it's computationally heavy—like hiring a team of 100 accountants to check a single receipt. If you use this method for every tile in every step of the simulation, it takes forever to run.

The Problem

For years, scientists had to choose: Speed or Safety.

• If you wanted speed, you risked the simulation crashing.
• If you wanted safety, you waited days for the result.

The Solution: "Volume Term Adaptivity" (v-adaptivity)

This paper introduces a brilliant new strategy called v-adaptivity. Think of it as a smart traffic controller for your computer simulation.

Instead of forcing every single tile to use the same method, the computer now has a "smart switch" that decides, in real-time, which method to use for each specific tile at every single moment.

Here is how the analogy works:

1. The "Traffic Controller" (The Indicator)

The computer constantly monitors the "traffic" inside each tile. It asks a simple question: "Is this area calm, or is it about to crash?"

• If the area is calm (smooth flow, no shockwaves): The controller says, "No need for the heavy accountants! Let's use the Fast & Cheap method." This saves a massive amount of time.
• If the area is chaotic (shockwaves, turbulence, potential instability): The controller screams, "Danger! Switch to the Safe & Detailed method immediately!" This prevents the simulation from crashing.

2. The Two Types of "Smart Switches"

The paper proposes two different ways for this controller to make its decision, depending on what you care about most:

• The "Safety First" Switch (Robustness):

  • The Rule: "If the fast method creates even a tiny bit of 'mathematical mess' (entropy) that the safe method wouldn't, switch to the safe method immediately."
  • The Result: This makes the simulation incredibly robust. It can handle extreme chaos without crashing, often lasting longer than simulations that only use the safe method. It's like having a safety net that only deploys when you actually start to fall.

• The "Efficiency First" Switch (Speed):

  • The Rule: "If the fast method creates a little bit of mess, but it's within a 'tolerable' limit, keep using the fast method! Only switch to the safe method if things get truly out of control."
  • The Result: This is the speed demon. It allows the computer to run the fast method almost everywhere, only paying the "tax" of the expensive method when absolutely necessary. In some tests, this made the simulation 3 times faster than using the safe method everywhere, while still producing accurate results.

Why This Matters

Imagine you are driving a car across a country.

• Old Way: You drive a heavy, armored tank everywhere. It's safe, but you burn a gallon of gas every mile and arrive very slowly.
• Old Way 2: You drive a lightweight sports car everywhere. It's fast, but if you hit a pothole (shockwave), you crash.
• New Way (v-adaptivity): You drive the sports car on the smooth highway. The moment you approach a bumpy road or a construction zone, the car instantly transforms into an armored tank. Once the road is smooth again, it turns back into a sports car.

The Bottom Line

This paper gives scientists a way to have their cake and eat it too. They can run simulations that are both incredibly fast and incredibly stable. By intelligently swapping between a "quick guess" and a "rigorous check" depending on the local conditions, they can simulate complex phenomena like supersonic flight, exploding stars, or turbulent weather much more efficiently than ever before.

It's a shift from "one size fits all" to "the right tool for the right job, right now."

Technical Summary

1. Problem Statement

High-order Discontinuous Galerkin (DG) schemes are powerful tools for solving time-dependent partial differential equations (PDEs), particularly hyperbolic conservation laws like the compressible Euler and Navier-Stokes equations. However, a fundamental trade-off exists between robustness and computational efficiency:

• Weak Form (WF) Volume Terms: Standard DG schemes use a weak formulation for the volume integral. This is computationally cheap (requiring one flux evaluation per quadrature node) but often lacks entropy stability, leading to instabilities in the presence of shocks or under-resolved features.
• Flux-Differencing (FD) Volume Terms: To ensure entropy stability (and thus robustness), schemes often employ a flux-differencing (FD) form using entropy-conserving two-point fluxes. While robust, this approach is significantly more expensive. For tensor-product elements, the cost scales quadratically with the polynomial degree ($p$), and for non-tensor elements (e.g., triangles), the cost is even higher due to the need for flux evaluations between all node pairs.

The core problem addressed is how to achieve the robustness of FD schemes without incurring their high computational cost across the entire domain.

2. Methodology: Volume Term Adaptivity (v-adaptivity)

The authors propose a novel adaptivity strategy termed v-adaptivity, which dynamically switches the discretization of the volume term at every element and every Runge-Kutta stage. The choice is based on suitable indicators:

A. The Two Discretization Options

1. Weak Form (WF): Uses the standard analytical flux evaluation. It is efficient but potentially unstable.
2. Flux-Differencing (FD): Uses entropy-conserving two-point fluxes. It is robust and entropy-stable but expensive.

B. Adaptation Strategies

The paper introduces two distinct strategies based on the choice of indicator:

1. Robustness-Oriented Strategy (Rigorous Indicator):

   • Goal: Maximize entropy stability.
   • Mechanism: Computes the entropy production rate for both the WF and FD forms. If the WF form produces more entropy dissipation (decay) than the FD form, the WF is used. Otherwise, the FD form is used.
   • Optimization: To avoid the cost of computing the FD term just for comparison, the authors utilize a rigorous entropy inequality. They replace the FD volume term calculation with a surface integral of the entropy potential, allowing the decision to be made without fully evaluating the expensive FD volume term unless instability is detected.

2. Efficiency-Oriented Strategy (Heuristic Indicator):

   • Goal: Maximize computational speed while maintaining stability.
   • Mechanism: Uses the WF form as a predictor. If the entropy production of the WF form stays below a user-defined threshold ($\sigma$), the WF is accepted. If it exceeds the threshold, the scheme switches to the robust FD form.
   • Application: This allows the scheme to use the cheap WF form in smooth regions and only pay the cost of the FD form in critical regions.

C. Integration with Shock-Capturing

The v-adaptivity framework is also combined with Blended DG-FV schemes. In this nested approach:

• An a-priori troubled-cell indicator (based on modal energy) first determines if a cell needs stabilization.
• If stabilization is needed, a blend of DG and low-order Finite Volume (FV) is used.
• Within this framework, the volume term can still be adapted between WF and FD based on entropy production indicators.

3. Key Contributions

• Novel Adaptivity Concept: Introduction of "v-adaptivity," a general framework for switching volume term discretizations locally in space and time, distinct from traditional $h$-, $p$-, or $r$-adaptivity.
• Rigorous Entropy-Stable Criterion: Development of a method to ensure global entropy stability by switching to WF only when it is more dissipative than FD, proven to maintain convergence to entropy solutions.
• Efficiency Heuristics: Proposal of a threshold-based indicator that significantly reduces computational cost by allowing controlled entropy production in non-critical regions.
• Linear Stability Analysis: Demonstration that v-adaptive schemes can remedy linear stability defects inherent in pure entropy-conserving FD schemes, allowing simulations to run longer without crashing.
• General Applicability: The framework is designed to work with general quadrature rules and is particularly beneficial for non-tensor-product elements (e.g., triangles/tetrahedra) where FD costs are prohibitive.

4. Results and Verification

The authors verified the scheme using the Trixi.jl high-order DG code across various test cases in 1D, 2D, and 3D:

• Accuracy & Convergence: The adaptive schemes maintain the formal order of accuracy (e.g., 4th order) in smooth regions (Isentropic Vortex, Density Wave).
• Stability:
  • In linear stability tests (Density Wave), pure FD schemes crashed, while the adaptive scheme remained stable up to the final time.
  • For Burgers' equation and the Sod shock tube, the adaptive scheme captured shocks correctly without the uncontrolled oscillations seen in pure WF schemes.
• Performance Gains:
  • Kelvin-Helmholtz Instability (Triangles): The heuristic indicator made the adaptive scheme ~7x faster than the pure FD scheme while maintaining stability.
  • Taylor-Green Vortex: The adaptive scheme achieved "super-resolution" behavior for enstrophy and reduced hyperbolic volume term costs by roughly 50%.
  • ONERA M6 Wing: Using a shock-capturing indicator, only 0.4% of elements required the expensive FD term, resulting in a **35% speedup** in total RHS computation.
  • Rayleigh-Taylor & Sedov Blast: Significant speedups (up to 3.3x) were observed on uniform meshes where only a small fraction of cells required stabilization.
• Navier-Stokes Limitations: For viscous flows, the speedup is less pronounced (13–25%) because the parabolic terms (gradients) dominate the runtime and are not affected by the volume term adaptivity.

5. Significance

This work provides a practical solution to the long-standing efficiency vs. robustness dilemma in high-order DG methods. By intelligently localizing the expensive entropy-stable discretization only where necessary, v-adaptivity enables:

1. Feasible High-Fidelity Simulations: Making high-order simulations of complex turbulent and shock-dominated flows computationally tractable.
2. Robustness without Penalty: Achieving the stability of entropy-conserving schemes without the associated computational overhead in smooth flow regions.
3. Future-Proofing: The framework is particularly well-suited for unstructured meshes (simplices) where FD costs are highest, opening the door for more efficient high-order methods on complex geometries.

In conclusion, the paper demonstrates that adaptive volume term selection is a powerful tool that enhances both the robustness and efficiency of DG schemes, making them more viable for large-scale, real-world engineering and scientific applications.