Comparison of Structure Preserving Schemes for the Cahn-Hilliard-Navier-Stokes Equations with Degenerate Mobility and Adaptive Mesh Refinement

This paper presents and compares decoupled implicit-explicit Discontinuous Galerkin schemes combined with adaptive mesh refinement for solving the Cahn-Hilliard-Navier-Stokes equations with degenerate mobility, evaluating their performance in mass conservation, bound preservation, and energy dissipation against existing literature methods.

The Gist

Imagine you are watching a drop of oil floating in a glass of water. As time passes, the oil might stretch, split into smaller drops, or merge with other drops. This is a multiphase flow. Simulating this on a computer is incredibly hard because the boundary between the oil and water isn't a sharp line; it's a fuzzy, shifting zone where the two substances mix slightly.

This paper is about building better "digital cameras" to film these invisible, fuzzy boundaries without making mistakes.

Here is the breakdown of the research using simple analogies:

1. The Problem: The "Fuzzy" Boundary

In the real world, the line between oil and water is a gradient. In computer simulations, we use a variable called a phase-field (let's call it the "mood meter") to track this.

• Mood = 1: Pure Oil.
• Mood = -1: Pure Water.
• Mood = 0: The messy mix in the middle.

The computer needs to make sure three things happen:

1. Don't lose any oil or water: The total amount must stay the same (Mass Conservation).
2. Don't get crazy numbers: The mood meter should never go above 1 or below -1. If it hits 1.5, the computer thinks there is "more than pure oil," which is physically impossible and breaks the math (Bound Preservation).
3. Don't create energy out of thin air: The system should naturally calm down over time, just like a hot cup of coffee cools down (Energy Dissipation).

2. The Old Way vs. The New Way

For a long time, scientists used FEM (Finite Element Method). Think of this like a smooth, continuous blanket draped over the simulation area. It's fast and looks nice, but if you pull the blanket too hard (simulate a sharp interface), it can tear or stretch into impossible shapes (violating the bounds).

The authors are comparing this to DG (Discontinuous Galerkin) methods. Think of DG as a patchwork quilt made of many small, separate squares.

• The Advantage: Because the squares are separate, you can stretch one square without ruining its neighbor. This makes it much better at handling sharp, jagged edges where the oil meets the water.
• The Challenge: Because the squares are separate, you have to write rules for how they talk to each other so the quilt doesn't fall apart.

3. The "Guardians" (Structure Preserving Schemes)

The paper tests different "guardians" (algorithms) to ensure the simulation stays physically realistic. They tested three main types of guardians:

• The "Limiter" (The Bouncer):
  Some methods (like SIPG-L and SWIP-L) use a bouncer at the door of every grid square. If the "mood meter" tries to go to 1.1 or -1.1, the bouncer gently pushes it back to 1 or -1.

  • Result: This keeps the simulation safe and realistic. The SWIP-L method was found to be the most robust and efficient "bouncer."

• The "ASU" Scheme (The Specialized Architect):
  This is a clever method that uses a low-resolution "sketch" to guide the high-resolution painting. It naturally keeps the numbers within bounds without needing a bouncer.

  • Result: It works great for simple mixing, but it gets very slow and complicated when you add fast-moving water currents (Navier-Stokes).

• The "FEM-L" (The Hybrid):
  This takes the smooth blanket (FEM), cuts it into squares to check for errors, fixes the errors, and stitches it back together.

  • Result: It's fast, but sometimes it accidentally loses a tiny bit of "oil" (mass) when the water moves fast.

4. The "Adaptive Mesh" (The Zoom Lens)

One of the coolest features in this paper is Adaptive Mesh Refinement.
Imagine you are taking a photo of a crowd. You don't need a 4K camera for the empty sky, but you need it for the faces in the crowd.

• The computer starts with a coarse grid (low resolution) everywhere.
• As soon as the oil and water start mixing (the "fuzzy" zone), the computer zooms in and adds more grid squares only in that specific area.
• Where the fluid is just pure water, it stays zoomed out to save computing power.
• Analogy: It's like a security guard who only runs fast when they see a thief, but walks slowly when the store is empty.

5. The Verdict: Who Won?

The authors ran simulations of rising bubbles and merging droplets to see which method was the best "digital camera."

• The Winner: The SWIP-L method (the patchwork quilt with the bouncer).

  • It kept the oil/water amounts perfect.
  • It never let the numbers go crazy.
  • It handled the zooming (adaptive mesh) beautifully.
  • It was fast enough to be practical.

• The Runner Up: FEM-L. It's a good choice if you already use the "smooth blanket" method and just want to add a safety net, but it's not as perfect as the quilt method for complex flows.

• The Loser: The standard methods without the "bouncer" (limiters). They often produced "ghost" oil or water (violating mass) or impossible physics (violating bounds), which would ruin any scientific prediction.

Summary

This paper is a guidebook for scientists building simulations of fluids. It says: "If you want to simulate mixing fluids accurately, don't just use the smooth blanket. Use the patchwork quilt with a bouncer (SWIP-L) and let the computer zoom in only where the action is happening." This ensures your digital experiments are as reliable as real-world physics.

Technical Summary

1. Problem Statement

The paper addresses the numerical simulation of multiphase fluid flows using the Cahn-Hilliard-Navier-Stokes (CHNS) system. The core challenge lies in developing numerical schemes that simultaneously satisfy three critical physical properties, often referred to as "structure preservation":

1. Mass Conservation: The total mass of the phase-field variable must remain constant over time.
2. Bound Preservation: The phase-field variable $\psi$ (representing concentration) must strictly remain within the physical interval $[-1, 1]$. Violations lead to unphysical densities in the coupled Navier-Stokes equations.
3. Energy Dissipation: The system's free energy must decrease monotonically, adhering to the second law of thermodynamics.

While many existing schemes satisfy energy dissipation, few guarantee both mass conservation and bound preservation, particularly when using degenerate mobility functions ($M(\psi) = 1-\psi^2$) and adaptive mesh refinement (AMR). The authors aim to fill the gap in comparative literature by evaluating decoupled implicit-explicit (IMEX) formulations based on Discontinuous Galerkin (DG) methods against standard Finite Element Methods (FEM).

2. Methodology

Mathematical Model

The authors utilize a thermodynamically consistent CHNS model with:

• Degenerate Mobility: $M(\psi) = 1 - \psi^2$, which naturally enforces bound preservation in the continuous setting but poses challenges for discretization.
• Splitting Scheme: The CH and NS equations are solved separately (decoupled) to facilitate implementation in standard simulation packages.
• Adaptive Mesh Refinement (AMR): Grids are refined near interfaces (where $|\psi| \approx 0$) and coarsened in pure phases to optimize computational cost.

Numerical Schemes Compared

The study evaluates several schemes implemented in Dune-Fem using the Unified Form Language (UFL):

1. Standard Continuous Galerkin (FEM):

   • Uses standard Lagrange basis functions.
   • FEM-C: Applies a post-processing "cut-off" limiter ($\bar{\psi} = \text{clip}(\psi)$) to enforce bounds.
   • FEM-L (Limited FEM): A novel approach where the continuous solution is projected to a DG space, limited element-wise, and projected back to the continuous space using a mass-lumped $L^2$ projection. This aims to preserve mass while enforcing bounds.

2. Discontinuous Galerkin (DG) Schemes:

   • SIPG (Symmetric Interior Penalty Galerkin) & SWIP (Symmetric Weighted Interior Penalty): Both use piecewise linear polynomials.
   • Limiters (SIPG-L, SWIP-L): A scaling limiter is applied element-wise to enforce bounds without violating mass conservation.
   • ASU (Acosta-Soba Upwinding): A specialized structure-preserving scheme using an auxiliary discontinuous piecewise-constant variable for upwinding and mobility treatment.

3. Time Discretization:

   • Implicit-Explicit (IMEX) Euler schemes are used.
   • The nonlinear potential $W'(\psi)$ is treated via the Eyre decomposition (convex-concave splitting) to ensure unconditional energy stability.

Implementation Details

• Solver: GMRES with DUNE-ISTL.
• Adaptivity: Uses an indicator function $H(\psi) = (1-\psi^2)^4$ to trigger refinement/coarsening.
• Coupling: The Incremental Pressure Correction Scheme (IPCS) is used for the Navier-Stokes part.

3. Key Contributions

1. Comprehensive Comparative Study: The paper provides the first thorough comparison of DG and FEM schemes specifically for CHNS with degenerate mobility and AMR, focusing on the trade-offs between the three structure-preserving properties.
2. Novel FEM-L Scheme: The authors propose and validate the FEM-L scheme, which successfully enforces bound preservation for standard FEM while maintaining mass conservation through a specific projection-limiting-projection workflow.
3. Theoretical Analysis:
   • Proved coercivity for the SWIP formulation with degenerate mobility, addressing the singularity issues when $\psi \to \pm 1$.
   • Demonstrated that standard FEM and DG schemes without limiters violate the maximum principle, leading to unphysical densities.
   • Showed that the ASU scheme naturally preserves bounds without explicit limiters but struggles with higher-order extensions.
4. UFL Implementation: All schemes are formulated in UFL, making them accessible for implementation in various FEM-based software packages (e.g., FEniCS, Firedrake).

4. Results

The schemes were tested on four scenarios: an accuracy test with manufactured solutions, pure Cahn-Hilliard (no flow), rotating merging bubbles, and a rising bubble benchmark.

• Convergence:
  • Standard FEM and DG schemes (SIPG/SWIP) achieved $O(h^2)$ convergence in $L^2$ and $O(h)$ in $H^1$.
  • The ASU scheme (using piecewise constant variables) achieved $O(h)$ convergence.
• Mass Conservation:
  • DG Schemes (SIPG-L, SWIP-L) and ASU: Excellent mass conservation (deviations $\approx 10^{-14}$ to $10^{-15}$).
  • FEM-L: Good mass conservation, though slightly higher deviation than DG due to the projection steps.
  • FEM-C: Significant mass loss ($O(10^{-4})$) due to the aggressive cut-off post-processing.
  • Standard FEM/SIPG/SWIP (unlimited): Mass is conserved numerically, but the solution violates physical bounds.
• Bound Preservation:
  • ASU, SIPG-L, SWIP-L, and FEM-L: Successfully kept $\psi \in [-1, 1]$.
  • Unlimited Schemes: Exhibited clear overshoots/undershoots (e.g., $\psi > 1.02$), which would cause unphysical densities in coupled fluid simulations.
• Energy Dissipation: All tested schemes satisfied the energy dissipation law.
• Computational Cost:
  • FEM is the fastest but fails bound preservation.
  • DG schemes are generally slower due to higher degrees of freedom but offer superior physical consistency.
  • SWIP-L was found to be the most robust and efficient among the DG schemes due to better matrix conditioning.
  • ASU was the slowest in coupled settings due to the mixed formulation and additional variables.

5. Significance and Conclusion

The paper concludes that structure preservation is non-negotiable for accurate multiphase flow simulations, particularly when training machine learning models or reconstructing images from experimental data where unphysical values break algorithms.

• Recommendation: The SWIP-L (Symmetric Weighted Interior Penalty with Limiter) scheme is recommended as the most robust choice for general multiphase flow problems, offering a balance of mass conservation, bound preservation, and computational efficiency.
• Alternative: For applications with moderate advection where FEM is already established, FEM-L is a viable, improved alternative to standard FEM.
• Limitations: The ASU scheme, while theoretically sound for bounds, is computationally expensive and difficult to extend to higher-order polynomials.

The work significantly advances the state-of-the-art by providing a clear roadmap for selecting numerical schemes that respect the underlying physics of phase-field models, ensuring that simulations remain physically valid even under complex adaptive meshing and fluid coupling.