Discretization Methods: FEA, CFD, and Beyond

Breaking Continuous Physics into Computable Pieces

The physical world is continuous. Stress varies smoothly across a beam. Temperature gradients form continuous fields in a heat exchanger. Velocity profiles transition without jumps through a pipe. But computers cannot solve continuous equations over arbitrary geometries in closed form. The breakthrough of discretization methods — the invention that made computational engineering possible — is a simple idea with profound consequences: break the continuous domain into a finite number of discrete pieces, approximate the governing equations on each piece, and assemble the pieces into a global system that can be solved numerically.

This is the foundation of both FEA and CFD. The physics differ. The equations differ. But the core strategy is identical: discretize, approximate locally, assemble globally, solve.

FEA: Structural Mechanics in Discrete Form

In structural FEA, the domain (the geometry of the part or assembly) is divided into elements — tetrahedra, hexahedra, shells, beams — connected at nodes. The governing equation is equilibrium: the internal forces in the structure must balance the applied external forces. In matrix form, this becomes Ku = F, where K is the global stiffness matrix (assembled from element stiffness matrices), u is the displacement vector at every node, and F is the applied force vector.

The elegance of this formulation is that each element's behavior is characterized by its stiffness matrix, which depends only on the element's geometry, material properties, and the assumed displacement interpolation (shape functions). Assembly is bookkeeping: each element contributes to the global matrix at the rows and columns corresponding to its nodes.

Element types matter

The choice of element type is a modeling decision, not just a meshing convenience.

Beam elements represent slender structures with one dominant dimension. They assume plane sections remain plane and are computationally cheap. Use them for frames, trusses, and structures where bending and axial behavior dominate.

Shell elements represent thin-walled structures with two dominant dimensions. They capture membrane and bending behavior without discretizing through the thickness. A shell mesh of an aircraft fuselage with thousands of elements replaces a solid mesh that would need millions.

Solid elements (tetrahedra, hexahedra) represent the full three-dimensional continuum. No geometric simplifications. Hexahedral elements generally provide better accuracy per element than tetrahedra but are harder to generate for complex geometries. Tetrahedral meshes can be generated automatically for almost any geometry, which makes them the default for complex parts — but they require finer meshes to achieve equivalent accuracy.

Mesh quality determines answer quality

A mesh is not just a discretization — it is an approximation. Every mesh introduces discretization error. The question is whether that error is acceptably small.

Too coarse: the mesh cannot resolve stress gradients, especially at geometric features like fillets, holes, and notches. Peak stress is underestimated. The solution appears smooth because the mesh forced it to be smooth, not because the physics is smooth.

Too fine: accuracy improves but at exponential cost. Doubling the mesh density in 3D increases element count by roughly 8x and solution time by considerably more (matrix assembly and solve scale superlinearly). There is a point of diminishing returns where further refinement changes the answer by fractions of a percent.

Mesh convergence is the practice of running the same problem at increasing mesh densities and plotting a key result (peak stress, total deformation) against element count. When the curve flattens, the solution is mesh-converged. This is not optional — it is the minimum standard for credible FEA.

CFD: Fluid Mechanics and the Turbulence Problem

CFD discretizes the same core idea — continuous equations on discrete cells — but the governing equations are far more challenging. The Navier-Stokes equations describe conservation of mass, momentum, and energy for fluid flow. They are nonlinear, coupled, and in turbulent regimes, they produce chaotic solutions with structures spanning orders of magnitude in length and time scales.

The turbulence modeling hierarchy

Direct Numerical Simulation (DNS) resolves every turbulent eddy, from the largest energy-containing structures down to the smallest dissipative scales. It requires no modeling — it solves the exact equations. It also requires mesh resolution proportional to the Reynolds number raised to the 9/4 power. For an aircraft at cruise Reynolds numbers, DNS would require meshes on the order of 10^16 cells. This is not practical and will not be practical for decades.

Large Eddy Simulation (LES) resolves the large turbulent structures and models the small ones. It reduces cost dramatically compared to DNS but still requires meshes and time steps that are orders of magnitude more expensive than industrial practice allows for most applications.

Reynolds-Averaged Navier-Stokes (RANS) models all turbulent fluctuations, solving only for time-averaged quantities. This is what most industrial CFD uses. It is affordable. It is also approximate, and the quality of the approximation depends entirely on whether the turbulence model's assumptions match the actual flow physics. A model tuned for attached boundary layers may fail for massively separated flow. A model validated at subsonic speeds may be inappropriate at transonic conditions.

The practitioner's responsibility is not to master turbulence theory but to know that RANS results carry modeling uncertainty that does not decrease with mesh refinement — because the error is in the model, not the mesh. Mesh convergence in CFD proves the numerics are resolved. It does not prove the physics is captured correctly.

Mesh sensitivity in CFD vs. FEA

CFD solutions are generally more mesh-sensitive than FEA solutions. In structural FEA, a factor-of-two change in mesh density typically changes peak stress by single-digit percentages once the mesh is reasonably resolved. In CFD, the same refinement can change drag predictions by tens of percent, particularly in regions of separated flow or complex turbulent structures. This is why CFD validation against experimental data is not a luxury — it is a necessity.

Where Discretization Methods Struggle

Despite decades of maturation, several classes of problems remain challenging for traditional discretization approaches.

Multi-scale problems. Predicting how microstructural features (grain boundaries, fiber orientations, void distributions) influence macroscopic behavior requires resolving phenomena across length scales spanning six or more orders of magnitude. Full discretization at the finest scale is computationally prohibitive. Multi-scale methods exist (FE-squared, homogenization) but add complexity and computational cost that limit their routine use.

Real-time applications. A structural digital twin that ingests sensor data and predicts remaining life in real time cannot wait for a multi-hour FEA solve. Control systems that adapt based on flow conditions cannot wait for a CFD run. The solve time of traditional methods makes them fundamentally unsuitable for applications requiring sub-second response.

High-dimensional parameter spaces. When the design space has dozens or hundreds of parameters, running a full FEA or CFD evaluation at each design point during optimization or uncertainty quantification becomes intractable. A 50-dimensional optimization with 1,000 evaluations per dimension would require 50,000 simulations — at hours per simulation, that is years of compute.

How AI Augments Discretization Methods

AI does not replace FEA and CFD solvers. It changes when and how often you need to run them.

ML surrogate models. Train a machine learning model on a database of FEA or CFD results (inputs: geometry, loads, boundary conditions; outputs: stress, displacement, drag, heat flux). Once trained, the surrogate predicts results in milliseconds. The key insight is that you are not replacing the physics — you are replacing the need to run the solver for every evaluation. The solver generates the training data. The surrogate replaces the solver for predictions within the training domain. This makes optimization over thousands of candidates and Monte Carlo over tens of thousands of samples computationally feasible.

Physics-Informed Neural Networks (PINNs). Standard ML surrogates learn from data alone. PINNs embed the governing equations (equilibrium, Navier-Stokes, heat equation) as constraints in the neural network's loss function. This means the network's predictions must satisfy the physics, even in regions where training data is sparse. The result: better generalization with less data. The limitation: PINNs are currently slower to train than standard surrogates and have difficulty with sharp gradients and complex geometries.

AI-assisted mesh generation. Mesh quality is critical and mesh generation for complex geometries is time-consuming. ML models trained on databases of meshed geometries and their solution quality can predict where finer mesh is needed before running the solver, reducing the iterative mesh-refine-resolve cycle. Some approaches predict optimal mesh density distributions directly from geometry features.

Error estimation. ML models can predict where the current mesh is too coarse — identifying regions where discretization error is high — before running the full high-fidelity simulation. This allows targeted refinement rather than uniform refinement, saving compute without sacrificing accuracy where it matters.

Walk through the three stages. The progression is not simply from worse to better — each stage serves a different purpose and carries different risks. The coarse mesh is appropriate for screening. The fine mesh is necessary for certification. The surrogate is transformative for exploration — but only within the domain established by the fine-mesh training data.

Assessment