TL;DR
This paper provides the mathematical "missing link" required to implement complex, high-performance finite elements—specifically the Arnold–Winther and Mardal–Tai–Winther elements—into automated software frameworks like Firedrake. By generalizing Piola transformation theory, the authors enable these elements to be used on arbitrary unstructured meshes while maintaining their unique properties, such as exact symmetry for stress tensors and robustness in incompressible fluid flow.
What Problem Does This Paper Address?
In finite element analysis, we typically define basis functions on a simple "reference" shape (like a unit triangle) and map them onto the actual "physical" cells of a mesh. For simple scalar functions, this is easy. For vector fields, we use the Piola transform to ensure that physical properties (like flow across an edge) are preserved when moving from the reference to the physical space.
However, some of the most powerful elements in mathematical physics—like the Arnold–Winther (AW) element for elasticity and the Mardal–Tai–Winther (MTW) element for fluid flow—are "Piola-inequivalent." This means the standard mapping "mixes up" their degrees of freedom (DOFs). For example, a tangential moment on the reference triangle might map to a combination of normal and tangential moments on the physical triangle. Because of this complexity, these elements have historically been excluded from high-level, automated finite element software, forcing researchers to write cumbersome, element-specific code.
Methods
The authors build upon a general transformation theory developed by Robert Kirby in 2018. Instead of assuming the mapping is a simple one-to-one correspondence, they treat the transformation as a change of basis.
- The Matrix $M$: They derive a cell-dependent, sparse matrix $M$ that relates the pushed-forward reference basis to the desired physical basis.
- Duality Approach: By looking at how the functionals (the rules for calculating DOFs) transform, they can systematically construct the correct mapping for any element, even those with vertex-based or tangential DOFs.
- Nitsche’s Method: The paper also introduces a way to handle "traction" boundary conditions (forces applied to surfaces) using Nitsche’s method, which is traditionally difficult for these specific types of stress-based elements.
- Software Integration: They implemented these transformations into the Firedrake project, allowing the software’s symbolic code generator to handle these exotic elements automatically.
Datasets
As a theoretical and numerical analysis paper, the authors do not use external "big data" sets. Instead, they evaluate their methods using:
- Manufactured Solutions: Mathematically constructed problems where the exact answer is known, used to verify that the code converges at the theoretically predicted rates.
- The Cantilever Beam: A classic engineering benchmark for linear elasticity.
- Warped Meshes: They specifically test on distorted, non-uniform grids to prove the mapping works on generic, unstructured geometry.
Key Insights and Ideas
- Sparsity is Key: Even though the mapping is complex, the transformation matrix $M$ is very sparse. This means the computational "tax" for using these advanced elements is negligible.
- Kernel Preservation: The authors show that these mappings preserve the "kernel" (the null space) of differential operators. This is vital for Geometric Multigrid solvers, which need to understand the underlying structure of the equations to solve them quickly.
- Unified Framework: The theory isn't just for AW or MTW; it provides a recipe for implementing any $H(\text{div})$ or $H(\text{curl})$ element that doesn't fit the standard Piola mold.
Key Results
- Numerical Verification: The MTW and AW elements showed the expected $O(h^2)$ or $O(h^3)$ convergence rates on unstructured, warped meshes, proving the mapping is correct.
- Incompressibility Robustness: The elements remained stable and accurate even as the material became perfectly incompressible (the "locking" limit), a scenario where standard elements often fail.
- Solver Performance: By using "vertex star relaxation" (a specific type of multigrid smoother), they achieved nearly constant iteration counts for solving massive systems with millions of degrees of freedom.
Limitations and Future Work
- 3D Complexity: While the theory holds for 3D, the authors note that the 3D versions of these elements are "monsters"—the Arnold–Winther element in 3D has a local dimension of 162, making it computationally expensive regardless of the mapping.
- Scaling: The authors noted that global basis functions can become ill-conditioned as the mesh is refined. They suggest a post-hoc rescaling of DOFs to keep the math stable.
- Extension to Other Elements: The authors suggest that this framework could next be applied to the Hu–Zhang elements, which are a more efficient (but still complex) alternative for 3D elasticity.
Summary updated