💻 computer science

Unfitted Multi-Level hp Refinement for Localized and Moving Solution Features

This paper introduces an unfitted multi-level $hp$-refinement strategy that utilizes superimposed, independently positioned overlay meshes on a fixed base discretization to efficiently and accurately resolve localized and moving solution features while avoiding the computational overhead of traditional mesh modifications and remeshing.

Original authors: Jan Niklas Schmäke, Martin Ruess

Published 2026-04-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jan Niklas Schmäke, Martin Ruess

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to take a high-resolution photograph of a busy city street. Most of the street is calm and uniform, but right in the middle, there is a chaotic, fast-moving parade.

The Old Way (Conventional Methods):
To get a sharp picture of the parade, traditional computer simulation methods would try to redraw the entire map of the street every time the parade moves. They would tear up the grid lines, stretch them, and rebuild the whole neighborhood just to focus on the parade. This is slow, messy, and computationally expensive. It's like trying to take a photo by constantly rebuilding the camera lens every time the subject moves.

The New Way (This Paper's Method):
The authors, Jan Niklas Schmäke and Martin Ruess, propose a smarter approach called "Unfitted Multi-Level hp Refinement."

Think of it like using a magnifying glass over a fixed map.

The Fixed Map (Base Mesh): You have a standard, low-resolution map of the whole city. It stays exactly the same. You don't need to redraw it.
The Magnifying Glass (Overlay Mesh): When the parade (or a crack, or a sharp temperature spike) appears, you simply place a smaller, high-resolution "overlay" map on top of just that specific area.
The Magic Trick (Superposition): Instead of cutting the map, you just layer the two. The computer adds the details from the magnifying glass to the general map. If the parade moves, you just slide the magnifying glass to the new spot. You don't have to rebuild the map underneath.

Key Features Explained Simply

1. "Unfitted" Means "Free to Move"
In older methods, the magnifying glass had to snap perfectly to the grid lines of the map underneath. If the parade moved slightly off-grid, the whole system had to struggle to fit the pieces together.
In this new method, the overlay can be placed anywhere. It doesn't matter if it cuts across the map's grid lines. It's like placing a sticker on a map; the sticker doesn't need to align with the map's borders. This makes it incredibly easy to move the focus around without breaking the system.

2. "Multi-Level hp" Means "Zooming In and Sharpening"
The method doesn't just zoom in; it zooms in and gets smarter.

h-refinement: Making the grid smaller (zooming in).
p-refinement: Making the math inside the grid more complex (sharpening the focus).
The paper shows you can stack these layers. You can have a coarse map, a medium zoom, and a super-sharp zoom all working together at the same time, each layer adding more detail only where it's needed.

3. The "Ghost" Boundary
To make sure the math works smoothly where the magnifying glass meets the regular map, the authors use a clever trick: they tell the computer that the "magnified" part must fade to zero right at the edge of the magnifying glass. This ensures the final picture looks seamless, with no jagged edges where the two layers meet.

What the Experiments Showed

The authors tested this idea on three specific scenarios:

The Broken Stick: They simulated a metal bar that snaps under pressure. The break creates a sudden, sharp jump in the data. The new method caught this break perfectly using less than half the computer memory (degrees of freedom) compared to the old "fitted" methods.
The Sharp Corner: They looked at a square with a sharp corner where forces get concentrated (a singularity). Again, the new method achieved the same accuracy as the old method but used significantly fewer resources because it didn't waste effort refining parts of the square that didn't need it.
The Moving Heat Source: This was the big test. They simulated a heat source (like a laser in 3D printing) moving in a circle.
- Old Way: Would require constantly remeshing the whole area as the laser moved.
- New Way: The laser just "rode" on top of the moving overlay. The result was nearly perfect, matching a super-detailed reference model, but using less than 1% of the computer power.

The One Catch: Tiny Overlaps

The paper does note one potential issue. If the "magnifying glass" overlaps the map so slightly that it's almost invisible (a tiny sliver), the math can get a bit unstable, like trying to balance on a very thin needle. However, in the moving examples they tested, this didn't cause major problems, and they suggest it can be managed with standard computer tricks.

The Bottom Line

This paper presents a way to simulate complex, moving, or broken things without constantly rebuilding the entire simulation. By using "floating" overlays that can be placed anywhere and moved freely, the method saves massive amounts of computing power while keeping the results highly accurate. It's a more efficient way to focus your computational energy exactly where the action is happening.

1. Problem Statement

Accurate finite element analysis of problems with localized features (singularities, sharp gradients, discontinuities) and moving sources (e.g., in additive manufacturing or crack propagation) is computationally challenging.

Limitations of Conventional Methods: Standard adaptive strategies (conforming $h$ -refinement) require repeated mesh generation, topology modification, and state variable projection. This introduces significant algorithmic overhead and limits robustness in transient simulations.
Limitations of Existing Superposition Methods: While "refinement-by-superposition" (overlaying meshes) avoids remeshing, existing fitted multi-level $hp$ methods impose strict topological constraints. They require overlay boundaries to align perfectly with underlying element boundaries and often necessitate complex deactivation rules to ensure linear independence and continuity. This restricts the flexibility of placing refinement zones and increases inter-level coupling.

2. Methodology: Unfitted Multi-Level $hp$ Refinement

The authors propose a novel framework that extends refinement-by-superposition to unfitted, non-matching meshes.

Core Concept

Superposition Strategy: The global approximation space ( $V_h$ ) is constructed by the superposition of a coarse base discretization and multiple independently positioned overlay meshes.
$u_h(\mathbf{x}) = \sum_{k=0}^{L} u_h^{(k)}(\mathbf{x})$
Unfitted Nature: Unlike fitted methods, overlay mesh boundaries do not need to align with the element boundaries of the base or other overlay levels. Overlay meshes can partially cover base elements.
Continuity Enforcement: Global $C^0$ continuity is enforced by applying homogeneous Dirichlet constraints on the artificial boundaries of the overlay domains ( $\Gamma_o^{(k)}$ ). This ensures the overlay contribution vanishes at the edge of the refinement region, preventing discontinuities in the global solution.
Linear Independence: The method relaxes the strict deactivation rules required in fitted hierarchical schemes. While pathological dependencies are theoretically possible, numerical experiments showed no issues, suggesting the constraints for continuity are sufficient for practical applications.

Implementation Aspects

Integration Regions: To handle non-matching meshes, the intersection of element partitions is partitioned into $C^\infty$ integration regions (sub-cells). A dimension-independent algorithm collects unique vertex coordinates from all overlapping meshes to define these sub-regions.
Auxiliary Elements: A "thin superposition layer" is introduced. For each integration region, an auxiliary element aggregates shape functions from all contributing parent elements across different mesh levels. This allows the reuse of standard finite element routines (interpolation, gradient evaluation, matrix assembly) without modifying the core solver.
Moving Features: For transient problems, overlay meshes can be inserted, translated, or removed dynamically. State transfer between time steps is performed via variational projection (e.g., $L^2$ projection) rather than pointwise interpolation to maintain consistency.

3. Key Contributions

Decoupling of Topology: The method removes the requirement for overlay boundaries to match base element boundaries, allowing for arbitrary placement and sizing of refinement zones.
Reduced Complexity: By avoiding complex topological tracking and deactivation rules inherent in fitted multi-level $hp$ methods, the implementation is lighter and more flexible.
Parameter Control: The introduction of a mesh size parameter ( $\alpha$ ) for overlay domains provides an additional degree of freedom to optimize approximation quality and convergence rates.
Dynamic Adaptation: The framework supports moving features without global remeshing, making it highly suitable for transient simulations like additive manufacturing.

4. Numerical Results

The method was validated against benchmark problems with closed-form solutions and a practical moving heat source problem.

1D Elastic Bar with Discontinuous Strains:
- Result: The unfitted method achieved exponential convergence.
- Efficiency: For a target error of $\approx 4.34 \times 10^{-5}\%$ , the unfitted method required 211 unknowns, whereas the fitted method required 436 unknowns (error $2.17 \times 10^{-3}\%$ ). The unfitted approach yielded significantly lower errors per degree of freedom.
2D Singular Corner Problem:
- Result: The unfitted method avoided creating unnecessary elements in regions away from the singularity (unlike the fitted method which bisects entire elements).
- Efficiency: The unfitted approach achieved an error of $1.54783\%$ with 1,177 unknowns, compared to $1.54778\%$ with 3,013 unknowns for the fitted method. Further optimization of the overlay size ( $\alpha=0.2$ ) reduced the error to $0.127\%$ with only 1,160 unknowns.
Small Overlap Conditioning:
- The study analyzed the effect of very small overlaps between superimposed elements. While condition numbers increase as overlap decreases (especially for high polynomial degrees $p \geq 6$ ), the moving-overlay simulations did not encounter configurations where this dominated computational cost.
Traveling Heat Source (Additive Manufacturing Simulation):
- Setup: A heat source moved along a circular path.
- Comparison: A locally refined unfitted mesh (1,841 unknowns) matched a highly resolved reference solution (250,000 unknowns) with high accuracy.
- Accuracy: The refined model captured peak temperatures and gradients with relative deviations of <1%, whereas an unrefined high-order mesh underestimated peaks by 32-71%. The refined model used <1% of the degrees of freedom of the reference mesh.

5. Significance and Conclusion

This work presents a robust and efficient alternative to traditional adaptive meshing for problems with localized or moving features.

Performance: It retains the exponential convergence rates of $hp$-methods while drastically reducing the number of degrees of freedom required for high accuracy.
Flexibility: The "unfitted" nature allows refinement zones to be moved and resized dynamically without remeshing or complex topological updates.
Applicability: The method is particularly significant for transient simulations (e.g., additive manufacturing, crack propagation) where the cost of repeated remeshing and state transfer is prohibitive.
Future Work: The authors suggest future research into automated refinement indicators, systematic sensitivity analysis for small overlaps, and advanced preconditioning techniques for ill-conditioned systems arising from tiny intersections.

In summary, the unfitted multi-level $hp$ method offers a superior error-to-cost ratio compared to fitted multi-level approaches, enabling high-fidelity simulations of complex, evolving physical phenomena with minimal computational overhead.