Reduced-Order Model ROM ★
Definition
An acceleration technique that projects a high-dimensional simulation (millions of DOFs) onto a low-dimensional subspace and solves there. A handful of basis functions capture the dominant physical modes, cutting solve time from hours to seconds. One of the most mature AI-acceleration routes in industry.
What it is
Classical reduction (e.g. POD-Galerkin) runs the full-order solver to collect snapshots, extracts an orthogonal basis with POD, then projects the governing equations onto that basis to obtain a low-dimensional system with tens to hundreds of DOFs. Solving that system is so cheap it can run at or near real time.
Key points
- Intrusive vs non-intrusive: intrusive requires modifying solver code to do the projection (better accuracy and conservation); non-intrusive only fits an input-output map (small change, weaker physical consistency).
- Limits of linear subspaces: for convection-dominated flows, shocks and moving boundaries the solution manifold is not low-dimensional and linear, so plain POD degrades and nonlinear reduction (autoencoder + latent dynamics) is needed.
- Relation to neural operators: a neural operator is "learned reduction", but it learns the map from data instead of doing an explicit projection.
- In production, teams usually keep the full-order model as a checker: ROM drives sweeps and optimization, critical cases are recomputed at full order.
Common pitfalls / boundaries
- ROM error has no a priori bound; you need full-order-vs-ROM error monitoring and an adaptive fallback.
- A change of operating conditions or geometry can invalidate the basis, requiring online updates or a fresh offline run.