Introduction: The Discontinuity Dilemma
High-order finite element methods (HO-FEM) are essential for high-fidelity fluid dynamics simulations, offering superior accuracy and low numerical error. However, these methods produce data as high-degree polynomials that are inherently discontinuous at element boundaries. This poses a significant challenge for modern topological analysis tools, such as those in the Topology ToolKit (TTK), which typically assume linear, continuous data. When raw HO-FEM data is processed, these discontinuities appear as artificial "cliffs," generating a vast amount of spurious topological noise that obscures physical features like vortices or contour trees.
Transformation Methodologies
To bridge this gap, this research evaluates three primary data transformation strategies designed to convert discontinuous HO-FEM data into a format suitable for topological analysis:
- Sampling to a Grid: Data is projected onto a regular, equispaced grid. While simple and fast, this method acts as a crude filter and often ignores the underlying mesh structure.
- Mesh Subdivision: Each high-order element is subdivided into smaller linear elements. Values at shared boundaries are averaged to enforce continuity, though this can still leave artifacts of the original mesh.
- Line-SIAC (L-SIAC) Filtering: This advanced technique uses Line-Smoothness-Increasing Accuracy-Conserving kernels to convolve the data. It mathematically "heals" discontinuities, increasing both the smoothness and accuracy of the field before analysis begins.
Topological Stability and Insights
A critical finding of this study is that simply increasing data resolution does not necessarily improve topological clarity. In fact, higher sampling rates on raw discontinuous data can exacerbate the "staircase" effect, leading to an increase in low-persistence noise features.
The researchers utilized persistence curves to identify "plateaus"—ranges where the number of identified features remains stable despite changes in the persistence threshold. Methods that produce wider, more stable plateaus, such as L-SIAC, allow for more robust feature selection and simplification.
Key Results
The comparative analysis across synthetic and real-world datasets (including 2D cylinder flow and 3D counter-rotating vortices) reveals significant differences in segmentation quality and computational cost.
In the 3D vortex study, the L-SIAC method successfully identified five consistent vortex structures. In contrast, the sampled method produced 45 fragmented segments, many of which were numerical artifacts requiring extensive manual filtering. However, this accuracy comes at a cost: while sampling takes only microseconds, L-SIAC filtering can take hundreds of milliseconds per point, positioning it as a high-fidelity tool for deep analysis rather than rapid previewing.
Conclusion and Future Directions
This work demonstrates that sophisticated filtering is often necessary to extract meaningful structural features from discontinuous simulation data. While L-SIAC provides the highest topological fidelity, its computational overhead suggests a need for more efficient implementations. Future research will focus on extending these methods to vector and tensor fields and developing "topology-aware" simulations that adaptively apply expensive filters only where necessary.
Summary updated