Proper Orthogonal Decomposition
POD extracts an optimal low-dimensional basis from simulation snapshots, enabling reduced-order models that run far faster than the full system.
Snapshots to modes
Proper orthogonal decomposition collects snapshots of a system's state (velocity fields, temperature distributions, plasma profiles) into a matrix and computes its singular value decomposition. The dominant left singular vectors form an orthonormal basis, the POD modes, that captures the most energy of the ensemble in the fewest directions. It is the discrete, data-driven counterpart of the Karhunen-Loeve expansion.
Optimality
For any target dimension r, the POD basis minimizes the mean-square reconstruction error over the snapshot set among all r-dimensional linear subspaces. The singular values quantify how much each mode contributes, and their decay rate tells you how compressible the dynamics are.
Building a reduced-order model
- Project the governing equations onto the POD basis (Galerkin projection)
- Integrate the small system of r ordinary differential equations
- Reconstruct full-field predictions by expanding in the modes
Limitations
POD is linear: it struggles with strongly convection-dominated or moving-front problems where the effective dimension does not decay. Galerkin reduced models can also become unstable even when the full model is stable, requiring stabilization or closure terms. Nonlinear alternatives such as autoencoders address the linearity but sacrifice interpretability.
Role in UQ workflows
POD reduces the output dimension so that a surrogate need only predict a few modal coefficients rather than a full field, and it reduces simulation cost so that many samples become affordable for uncertainty propagation. In fusion design, reduced-order models built from high-fidelity snapshots let parameter sweeps span operating space cheaply, with full-field runs reserved for verification near the frozen design point.