Topology-Preserving Mesh Adaptation for Sharp-Interface Multiphase PFEM
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.
Tom: We're looking at the paper: "Topology-Preserving Mesh Adaptation for Sharp-Interface Multiphase PFEM".
Title and authors: Jane: So the title itself points to a solution for keeping those interfaces sharp while using this Particle Finite Element Method framework. Who are the authors and what’s their background?
Tom: We have Félix Ruyffelaere, Michel Henryb, Jonathan Lambrechtsa, and Jean-François Remaclea. They are all from universities like UCLouvain and Grenoble Alpes.
Lu: I'm interested in how they framed this problem because the text mentions historical challenges with Eulerian formulations versus methods like Volume-of-Fluid or Level-Set which have their own inherent limitations when dealing with sharp interfaces <ref:2607.09446#pg2>.
Meng: If the method relies on geometric reconstruction techniques, like the alpha shape method mentioned elsewhere, does that mean we still have to deal with point clouds and remeshing after every step? That sounds computationally expensive to run in a production pipeline.
Lalam: The authors are focused on extending PFEM, which is already robust for things like breaking waves and fluid-structure interactions, but they are specifically struggling with internal interfaces in multiphase flows <ref:2607.09446#pg3>.
The paper's summary: Tom: So to summarize what the paper does, they present a fully Lagrangian framework based on PFEM that handles an arbitrary number of immiscible phases. They solve the incompressible Navier-Stokes equations where material points match the mesh nodes.
Jane: That’s a big shift from Eulerian methods because it’s Lagrangian and material points stay with their fluid phase, which is inherently good for tracking interfaces. But they still have to deal with those topological changes dictated by the grid size.
Lu: The core contribution here is this dynamic mesh adaptation strategy that preserves sharp geometric interfaces without relying on classical constrained triangulation <ref:2607.09446#pg1>. They assign a protecting node-empty disk to each interface segment, which forces that edge to be part of the Delaunay triangulation.
Meng: Decoupling the physics from the grid size sounds promising for incorporating sub-grid models, but I need to know how they handle those topological changes when they are governed by these new rules versus traditional methods.
Lalam: The protecting-ball criterion is derived from a duality between Voronoï diagrams and Delaunay triangulations, ensuring every interface edge belongs to the Delaunay triangulation by enforcing a diametral empty-disk condition <ref:2607.09446#pg1>.
The paper's improvements: Tom: They detail two main operations for this preservation: first, they split interface segments that are larger than a baseline resolution smin, and second, they filter out surrounding nodes encroaching on the protecting-ball while allowing bulk nodes to be removed.
Jane: That’s clever—it suggests a way to manage topology changes systematically. They also include mesh adaptation for the bulk fluid itself using Chew’s refinement algorithm guided by a sizing field s(di).
Lu: The sizing field s(di) is defined specifically to maintain resolution matched to the local interface size within a band of width dmin surrounding those interfaces <ref:2607.09446#pg1>. This smooth transition in mesh quality sounds like it keeps the element quality high while still being computationally efficient.
Meng: So, they are trying to maintain a high resolution exactly where the physics demands it—at the interface—while letting the bulk adapt based on this sizing field. But can we trust that this doesn't introduce new artifacts when we are dealing with sixteen phases <ref:2607.09446#pg3>?
Lalam: The validation shows they successfully preserve fluid filaments and robustly recover sharp boundaries through successive remeshing steps, and they show excellent mass retention, attributing volume loss to interface coarsening <ref:2607.09446#pg3>.
Conclusion: Tom: So, the main implication is that this methodology minimizes the dependence of the computed physics on the computational mesh by separating baseline resolution from topological transition length scales. They’ve shown they can control fluid topology independently of the global mesh size <ref:2607.09446#pg1>.
Jane: It really shows a systematic way to incorporate sub-grid models that account for unresolved physical mechanisms without ruining the interface structure through standard constrained triangulation, which is what they were worried about earlier.
Lu: It allows them to explicitly enforce topological changes, like an edge flip, while remaining fully controlled and mesh-independent by setting the minimum resolution smin based on physical criteria instead of just the global mesh size <ref:2607.09446#pg3>.
Meng: For production use, I see a system that can handle a large number of phases, like sixteen distinct fluid phases as they tested, which proves the framework is agnostic to how many phases we have <ref:2607.09446#pg1>. But I still see the massive challenge in implementing this reliably in three dimensions.
Lalam: The authors acknowledge that while chaotic mixing phases can cause a rapid surge in node count, their method provides a systematic framework for incorporating those sub-grid models that account for unresolved physical mechanisms <ref:2607.09446#pg1>. It’s a strong foundation for future work in this area.
Félix Ruyffelaere, Michel Henryb, Jonathan Lambrechtsa, Jean-FranŇois Remaclea
UCLouvain - iMMC · Univ. Grenoble Alpes, Inria
cs.CE, cs.CG
Submitted: 2026-07-10
Updated: 2026-10-08
DOI: 10.1016/j.cma.2026.119477
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 95/100
The gist: This paper presents a robust, fully Lagrangian framework based on the Particle Finite Element Method (PFEM) capable of simulating multiphase flows with an arbitrary number of immiscible phases.
Key concepts
- Fully Lagrangian Framework
- This method treats material points as nodes in the computational grid. This approach ensures that physical properties and conservation laws are naturally satisfied at every point, making it ideal for tracking fluid interfaces and complex material interactions without needing to track individual fluid parcels separately.
- Protecting-Ball Criterion
- This criterion uses the mathematical relationship between Voronoï diagrams (which define fluid regions) and Delaunay triangulations (the mesh structure). It enforces a condition where interface edges must belong to the Delaunay triangulation, effectively guaranteeing that sharp boundaries are accurately represented in the simulation.
- Dynamic Mesh Adaptation
- The algorithm adjusts the mesh size locally based on physical needs. It uses a sizing field that varies near interfaces, ensuring high resolution precisely where fluid interfaces exist. This allows the simulation to maintain accuracy for sharp features while optimizing computational resources elsewhere.
Terminology
Summary
This paper presents a robust, fully Lagrangian framework based on the Particle Finite Element Method (PFEM) capable of simulating multiphase flows with an arbitrary number of immiscible phases. p. 0 Topology-Preserving Mesh Adaptation for Sharp-Interface Multiphase PFEM
Félix Ruyffelaerea,∗
, Michel Henryb
, Jonathan Lambrechtsa
, Jean-François Remaclea
aUCLouvain - iMMC, Avenue Georges Lemaître 4, 1348 Louvain-la-Neuve, Belgium
bUniv. Grenoble Alpes, Inria, 38000 Grenoble, France
Abstract
This paper presents a robust, fully Lagrangian framework based on the Particle Finite Element
Method (PFEM) capable of simulating multiphase flows with an arbitrary number of immiscible
phases.Where the paper says it
The core contribution is a dynamic mesh adaptation strategy that naturally preserves sharp geometric interfaces without relying on classical constrained triangulation. p. 0 Interface-capturing methods can suffer from numerical diffusion, while interface-tracking
methods can allow the underlying mesh resolution to prematurely dictate topological changes. To
address these limitations, we introduce a dynamic mesh adaptation strategy that naturally preserves
sharp geometric interfaces without relying on classical constrained triangulation. A protecting,
node-empty disk is assigned to each segment of the discretized interface, ensuring that the edge
is part of the Delaunay triangulation. Our approach decouples the interface physics from the grid
size, allowing the integration of sub-grid physical models to properly govern topological changes
independently of the user-defined mesh size.Where the paper says it
Where the paper says it
Framework and Governing Equations
The method is built upon a fully Lagrangian framework where material points coincide with the mesh nodes. The governing equations involve solving the incompressible Navier-Stokes equations on a domain containing multiple fluid phases. p. 3 Governing equations
To dictate the motion of the particles, in the first step of our method, the incompressible
Navier-Stokes equations are considered in a domain with multiple fluid phases. Let Ω ⊂ R
d
, with
d = 2 in this work, be a domain containing n distinct and immiscible phases.Where the paper says it
-
The evolution of the velocity field and pressure field is governed by Equation 1. p. 3
Where the paper says it
The evolution of the velocity field u(x, t) and the pressure field p(x, t) is governed by: ρ du dt = ∇ · σ + ρg − γκnδΓ ∇ · u = 0 (1) where ρ is the density, σ is the Cauchy stress tensor, g is the gravitational acceleration vector. The last term of Equation 1, which we will denote fγ, accounts for the forces acting on the (d − 1)- dimensional interfaces under surface tension γ, with κ the curvature, n the unit normal and δΓ the Dirac distribution supported on the interface.
-
Coupling conditions at the interface require that the jump of a quantity across the interface is defined as [ϕ] = ϕ1 − ϕ2. p. 4
Where the paper says it
The physical properties within the domain are defined piecewise for each phase k: ρ(x, t) = ρk and µ(x, t) = µk if x ∈ Ωk(t) (4) Since the phases are immiscible and transported by the flow, and no mass transfer occurs across Γ, these properties are constant along the trajectories. At the interface Γ(t), defining the jump of a quantity ϕ across the interface as [ϕ] = ϕ1 − ϕ2, the coupling conditions are given by [u] = 0 on Γ [σ] · n = −γκn on Γ (5) where γ is the surface tension coefficient between the two phases around Γ, assumed constant along each interface, so that Marangoni effects are not considered, and κ = ∇ ·n is the local interface curvature.
-
The pressure jump is accounted for by modifying the discrete pressure gradient using an auxiliary field p∗. p. 4
Where the paper says it
To account for the surface tension force fγ and to model the subsequent jump of p, we use here a method introduced by the authors of [10], inspired from the Ghost-Fluid-Method [21, 22]. The key idea is to modify the discrete pressure gradient by adding to it an auxiliary field p ∗ , that is constructed explicitly to capture the jump given by Equation 6.
Interface Preservation Mechanism
The preservation of topological features relies on a protecting-ball criterion derived from the duality between Voronoï diagrams and Delaunay triangulations. The algorithm ensures interface edges belong to the Delaunay triangulation by enforcing a strict diametral empty-disk condition. p. 8 By enforcing an empty-disk criterion, each edge is protected through two specific operations:
local edge splitting (b) and volume node filtering (c). Algorithm 1 Interface adaptation
1: VΓ ← interfacial nodes ▷ Extracted from the previously advected mesh
2: EΓ ← interfacial edges
3: VB ← bulk nodes
Phase 1: Edge splitting
4: SplitEdges({ei ∈ EΓ : |ei
| > smin}) ▷ Baseline interface resolution
5: while ∃ei ∈ EΓ such that B(ci
, ri) ∩ VΓ ̸= ∅ and |ei
| > wmin do
6: Eenc ← DetectEncroachments(EΓ, VΓ) ▷ In parallel
7: SplitEdges(Eenc) ▷ In parallel; update VΓ and EΓ
8: end while
Phase 2: Bulk node deletion
9: BΓ ← diametral balls associated with the interfacial edges
10: for all vb ∈ VB do ▷ In parallel
11: if ∃B ∈ BΓ : vb ∈ B then
12: MarkForDeletion(vb)
13: end if
14: end for
9Where the paper says it
-
Edge splitting is performed on interface segments exceeding a baseline resolution smin. p. 8
Where the paper says it
By enforcing an empty-disk criterion, each edge is protected through two specific operations: local edge splitting (b) and volume node filtering (c). Algorithm 1 Interface adaptation 1: VΓ ← interfacial nodes ▷ Extracted from the previously advected mesh 2: EΓ ← interfacial edges 3: VB ← bulk nodes Phase 1: Edge splitting 4: SplitEdges({ei ∈ EΓ : |ei | > smin}) ▷ Baseline interface resolution 5: while ∃ei ∈ EΓ such that B(ci , ri) ∩ VΓ ̸= ∅ and |ei | > wmin do 6: Eenc ← DetectEncroachments(EΓ, VΓ) ▷ In parallel 7: SplitEdges(Eenc) ▷ In parallel; update VΓ and EΓ 8: end while Phase 2: Bulk node deletion 9: BΓ ← diametral balls associated with the interfacial edges 10: for all vb ∈ VB do ▷ In parallel 11: if ∃B ∈ BΓ : vb ∈ B then 12: MarkForDeletion(vb) 13: end if 14: end for 9
-
Surrounding nodes encroaching upon the protecting-ball are filtered out, while nodes within the fluid bulk are permissible to remove. p. 9
Where the paper says it
Algorithm 1 describes the implementation of the strategy illustrated in Figure 3. The first phase focuses on protecting interface edges from mutually encroaching upon one another while temporarily disregarding volume nodes (i.
Mesh Adaptation and Quality Control
To enhance element quality in the fluid bulk, a volume mesh adaptation step is incorporated using Chew’s refinement algorithm guided by a user-defined sizing field. This sizing field s(di) is defined to maintain resolution matched to the local interface size within a band of width dmin. p. 7 First, because the interface representation is
piecewise linear, a user-defined baseline resolution, smin, is enforced by splitting excessively long
edges. While this step helps maintain a critical resolution along the interfaces, it does not by itself
guarantee the preservation of these edges in the subsequent triangulation.Where the paper says it
Validation and Results
The framework was validated against several benchmarks including vortex in a box simulations and a challenging 16-phase Rayleigh-Taylor simulation. Numerical results confirm that the method successfully preserves fluid filaments and robustly recovers sharp boundaries through successive remeshing steps. p. 1 Diffuse interface models regularize the boundary over a finite thickness,
which inherently prevents the strict enforcement of sharp pressure jumps and introduces non-physical
numerical diffusion across the phases [5]. Conversely, while the Level-Set method preserves sharp
boundaries and provides smooth geometrical curvatures, it is fundamentally prone to numerical mass
loss, although new re-initialization techniques try to correct this flaw [6]. Diffuse interface models regularize the boundary over a finite thickness,
which inherently prevents the strict enforcement of sharp pressure jumps and introduces non-physical
numerical diffusion across the phases [5]. Conversely, while the Level-Set method preserves sharp
boundaries and provides smooth geometrical curvatures, it is fundamentally prone to numerical mass
loss, although new re-initialization techniques try to correct this flaw [6].Where the paper says it
Where the paper says it
Conclusion and Implications
The proposed methodology minimizes the dependence of computed physics on the computational mesh by decoupling baseline resolution from topological transition length scales. The framework demonstrates the ability to control fluid topology independently of the global mesh size. p. 0 Our approach decouples the interface physics from the grid
size, allowing the integration of sub-grid physical models to properly govern topological changes
independently of the user-defined mesh size. The capabilities and accuracy of the framework are
validated against standard multiphase benchmarks, closely matching references while maintaining a
remarkably low overall node count. Our approach decouples the interface physics from the grid
size, allowing the integration of sub-grid physical models to properly govern topological changes
independently of the user-defined mesh size. The capabilities and accuracy of the framework are
validated against standard multiphase benchmarks, closely matching references while maintaining a
remarkably low overall node count. study are formulated within a two-dimensional framework, the underlying geometric philosophy and
theoretical guarantees remain fully valid in three dimensions. (0): Solving the governing
equations (FEM).Where the paper says it
Where the paper says it
Where the paper says it
CRediT authorship contribution statement
Félix Ruyffelaere: Conceptualization, Methodology, Software, Validation, Investigation, Visualization, Writing – original draft,Software and Writing – review & editing.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Mesh adaptation strategy decoupling geometry preservation from mesh resolution allows physical models to decide when topological transitions occur
independently of the user-defined mesh size.
This enables simulation of sharp interfaces without relying on classical constrained triangulation. -
The protecting-ball criterion, derived from the duality between Voronoï diagrams and Delaunay triangulations, ensures that
every interface segment ei belongs to the Delaunay triangulation of V
by enforcing a diametral empty-disk condition. This guarantees the recovery of sharp interfaceswithout relying on constrained triangulations.
-
The mesh refinement algorithm incorporates a sizing field s(di) that varies continuously near interfaces, allowing the mesh to maintain a high resolution matched to the local interface size
within a band of width dmin surrounding the interfaces.
This smooth transition enhances element quality while maintaining computational efficiency. -
The framework can explicitly enforce topological changes, such as an edge flip, while remaining
fully controlled and mesh-independent
by setting smin based on physical criteria rather than global mesh size. This capability allows for the reproduction ofdistinct behaviors that align with different models from the literature.
-
The system can handle a large number of phases, demonstrated by simulating 16 distinct fluid phases, showing that the
global PFEM framework is completely agnostic to the number of phases.
This scalability is achieved through dynamic mesh adaptation and robust interface recovery.
Abstract
This paper presents a robust, fully Lagrangian framework based on the Particle Finite Element Method (PFEM) capable of simulating multiphase flows with an arbitrary number of immiscible phases. Interface-tracking methods can sometimes suffer from numerical diffusion or allow the underlying mesh resolution to prematurely dictate topological changes. To address these limitations, we introduce a dynamic mesh adaptation strategy that naturally preserves sharp geometric interfaces without relying on classical constrained triangulation. A node-empty disk is assigned to each segment of the discretized interface, ensuring that the edge is part of the Delaunay triangulation. Our approach decouples the interface physics from the grid size, allowing the integration of sub-grid physical models to properly govern topological changes independently of the user-defined mesh size. The capabilities and accuracy of the framework are validated against standard multiphase benchmarks, closely matching references while maintaining a remarkably low overall node count. We demonstrate the scalability and geometric versatility of the method, in particular with a challenging 16-phase Rayleigh-Taylor simulation.
Related papers
- Constrained Sensing and Reliable State Estimation with Shallow Recurrent Decoders on a TRIGA Mark II Reactor
- DREChanNet: Patient-Relative iEEG Model for Channel-Wise Epileptogenic Zone Localization
- A matrix-free, differentiable PyTorch solver for phase-field fracture: Formulation, benchmarks, and inverse analysis
- BizAgentBench: Unveiling Hidden Challenges in Deploying Long-Horizon Agents for Real-World Business
- Colour me shocked: Exact Molecular Hessians from local MLIPs in O(N) time using sparse differentiation!
- In-context learning as a retraining-free paradigm for constitutive modeling