Computational Fluid Dynamics CFD
Definition
The discipline and engineering practice of solving flow governing equations such as Navier-Stokes numerically, covering meshing, discretization, turbulence modeling and solution strategies. It is the primary target that Physics-AI aims to accelerate or replace in the fluid domain.
What it is
CFD is not one algorithm but a whole pipeline: geometry -> mesh -> discretization (FVM/FEM/spectral) -> turbulence model (RANS/LES/DNS) -> nonlinear / pressure-velocity coupling solve -> post-processing. Every stage carries heavy engineering experience, which is why "replacing CFD with AI" is much harder than it sounds.
Key points
- Turbulence is the core difficulty: DNS cost grows roughly as
Re^3, so engineering relies on RANS/LES approximations whose model error often exceeds discretization error. - AI entry points: turbulence closure, solver acceleration (pressure Poisson), end-to-end surrogates, mesh generation, super-resolution post-processing.
- Verification and validation (V&V) is intrinsic to CFD culture; any AI model entering an engineering workflow must pass it.
- Speedups must be measured on the full pipeline: replacing only the pressure solve while ignoring meshing and preprocessing yields limited end-to-end gain.
Common pitfalls / boundaries
- "AI in seconds vs CFD in hours" comparisons are often unfair (different accuracy, geometry and physical models).
- The statistical nature of turbulence makes instantaneous single-field comparisons weakly meaningful; compare statistics instead (mean field, energy spectrum, force coefficients).