Proper Orthogonal Decomposition POD
Definition
A dimensionality-reduction method that extracts an energy-optimal orthogonal basis from a matrix of simulation snapshots: the fewest basis functions capturing the most field energy. It is the foundation of classical POD-Galerkin reduced-order models.
What it is
Stack N field snapshots into a matrix S (one column per field), take the SVD, and keep the top r left singular vectors as the basis Phi_r. Choose r by an energy criterion (e.g. retain 99.9% of energy); typically tens to hundreds of modes represent a field with millions of DOFs.
Key points
- Optimality: in the Frobenius/energy sense, the POD basis is the best rank-
rlinear representation. - Snapshot design determines basis quality: sampled cases must cover the deployment range of parameters and flow regimes, otherwise the basis "has seen but not learned".
- The energy criterion does not guarantee accuracy: 99.9% energy can still hide large local errors, especially in small-scale-dominated problems.
- Relation to autoencoders: POD is the analytic optimum for linear reduction; an autoencoder learns a nonlinear manifold, which is more expressive but needs training.
Common pitfalls / boundaries
- POD assumes the solution manifold is nearly linear. For convection/transport-dominated problems (traveling waves, moving shocks) snapshot energy decays slowly and you need alignment/registration preprocessing or nonlinear methods.
- Once the basis is fixed, cases outside the training parameter range break it; online basis updates are required.