Differential Equations
Finite Element Method
Цели урока
- Derive the weak formulation of the Poisson problem
- Assemble the P1 stiffness matrix and understand its sparsity
- Implement 1D FEM and verify second-order convergence
- Understand the difference between h- and p-refinement and the LBB condition
Предварительные знания
- Integration by parts in 2D (Green's formula)
- Linear algebra (sparse systems)
- L^2 and H^1 spaces
How does Ansys solve elasticity problems in 3D objects with millions of degrees of freedom in seconds?
- Ansys, COMSOL, and Abaqus build FEA solvers on FEM for aerospace and automotive engineering
- Volvo and BMW crash tests are simulated in LS-DYNA with 4 million finite elements
- FEM solvers for the Schrodinger equation underpin quantum chip design
- Physics-Informed Neural Networks embed the FEM residual in a neural network loss function
From Engineering Intuition to Mathematical Theory
FEM was invented by engineers and justified by mathematicians later. In 1956 Turner, Clough, and Martin applied the method to structural analysis at Boeing. Clough coined the term 'finite element' in 1960. The mathematical foundation via Sobolev spaces was established by Babuska and Aziz in 1972, who proved optimal error estimates. Today FEM underpins all computational mechanics and increasingly neural PDE solvers as well.
Weak Formulation and Sobolev Spaces
Ansys, with 2.4 billion dollars in 2023 revenue, builds all its FEA solvers on the finite element method. The key is the weak formulation: instead of requiring u in C^2, it suffices to have square-integrable first derivatives. This opens the door to piecewise polynomial functions on arbitrary meshes.
The weak form is not merely a mathematical convenience - it reflects the physics directly. In structural mechanics the weak form is the principle of virtual work. Discontinuous material coefficients (joining two different materials) are handled automatically.
The Lax-Milgram theorem guarantees existence and uniqueness of the weak solution when the bilinear form a(u,v) = int(nabla u cdot nabla v) is bounded and coercive. For the Poisson problem this follows from the Poincare inequality.
Why is the weak formulation of the Poisson problem preferred for FEM?
Discretization and the Stiffness Matrix
After deriving the weak form, the domain is partitioned into elements (triangles, tetrahedra) and the solution is represented as a linear combination of hat basis functions phi_j supported on a few elements each. The key product is a sparse stiffness matrix.
Why is the stiffness matrix K sparse?
Higher-Order Elements and Adaptive Meshes
Linear P1 elements are just the beginning. P2 elements (quadratic, with midpoint nodes on edges) give O(h^2) in the H^1 norm and O(h^3) in L^2. Spectral elements with high polynomial degrees on large elements underpin high-accuracy solvers like Nektar++ and Nek5000 for turbulence. Adaptive meshes refine where the error gradient is largest.
| Element type | Degree p | Nodes per element (2D) | L^2 order |
|---|---|---|---|
| P1 (linear) | 1 | 3 | O(h^2) |
| P2 (quadratic) | 2 | 6 | O(h^3) |
| P3 (cubic) | 3 | 10 | O(h^4) |
| Spectral (hp-FEM) | high | many | exponential |
The LBB condition (Ladyzhenskaya-Babuska-Brezzi): for incompressible flow (Stokes, Navier-Stokes) the velocity and pressure spaces cannot be identical. Inf-sup stable pairs are required: P2/P1 (Taylor-Hood) or P1-bubble/P1.
What advantage does p-refinement have over h-refinement for smooth solutions?
FEM in Industry and Machine Learning
Physics-Informed Neural Networks (PINNs, Raissi et al. 2019) embed the PDE residual in the loss function and train without a mesh. For problems with sharp gradients (shocks, cracks) classical FEM with adaptive refinement remains more accurate. A natural hybrid is FEM-nets: a neural network parametrizes the solution on a finite-element mesh.
FEM in Crashworthiness: Ford and Tesla
Virtual crash test of an automobile
A Volvo XC90 crash test in LS-DYNA uses a model with 4 million finite elements, nonlinear elastoplastic material laws, and contact between body panels. Simulating 100 ms of impact takes 8 hours on 512 cores. Without FEM each vehicle generation would require more than 200 physical crash tests instead of 15.
Neural Operators (FNO, DeepONet) learn to solve PDEs across an entire family of right-hand sides without re-solving each time. FNO inference is 1000x faster than FEM at comparable accuracy - reshaping structural optimization and fast weather prediction.
Why does the PINN approach struggle with sharp gradients compared to adaptive FEM?
Connections to Other Areas
FEM sits at the intersection of functional analysis, numerical mathematics, and engineering simulation.
- Sobolev Spaces — Related topic
- Iterative Solvers (CG, Multigrid) — Related topic
- Neural Operators (FNO) — Related topic
- Adaptive Meshes — Related topic
Итоги
- The weak formulation reduces smoothness requirements to H^1 and admits piecewise polynomial basis functions
- The stiffness matrix K is sparse: K_ij is nonzero only for neighboring nodes, enabling O(N) solvers
- P1 elements give O(h^2) in L^2; p-refinement achieves exponential convergence for smooth solutions
- Neural Operators (FNO) accelerate FEM by 1000x for structural optimization and weather prediction
Вопросы для размышления
- Why can the same velocity and pressure spaces not be used in the Stokes problem (LBB condition)?
- How do adaptive meshes address the stress concentration singularity at a crack tip?
- What do FEM and convolutional neural networks have in common from the perspective of local operations?
Связанные уроки
- de-24-reaction-diffusion — Reaction-diffusion PDEs are a direct application of FEM
- de-27-schrodinger — FEM solves the Schrodinger equation on arbitrary potentials
- de-23-pde-bvp — Boundary value problems are the foundation of the weak formulation