POD and Galerkin Projection
Proper orthogonal decomposition extracts an optimal linear basis from data, and Galerkin projection evolves the equations within it.
Finding the dominant modes
Proper orthogonal decomposition (POD) analyzes a collection of simulation snapshots to find the linear directions, called modes, that capture the most variance. It is mathematically the singular value decomposition of the snapshot matrix. A handful of leading modes usually accounts for the vast majority of the system's energy, giving a compact basis for the state.
Projecting the physics
POD gives a basis but not the dynamics. Galerkin projection supplies them: it substitutes the reduced representation into the governing equations and projects the residual onto the retained modes. The result is a small system of ordinary differential equations for the modal coefficients, derived directly from the original physics rather than fit to data.
The intrusive nature
Galerkin projection is intrusive, meaning it needs access to the equations and often to the solver's operators. This yields a reduced model grounded in physics, but it can inherit instabilities, especially for advection-dominated or turbulent flows where truncated modes still carry important interactions. Stabilization and closure terms are often required.
Strengths and weaknesses
- Optimal linear compression for a given snapshot set
- Physics-based reduced dynamics, not merely a data fit
- Struggles with strongly nonlinear or transport-dominated problems
- Linear basis can need many modes where an autoencoder needs few
Where ML enters
Machine learning augments the classical pipeline in several ways: learning a closure for the neglected modes, replacing the linear POD basis with a nonlinear autoencoder for greater compression, or fitting the reduced dynamics non-intrusively from data when the solver operators are unavailable. POD-Galerkin remains the reference against which these learned reduced models are measured.