Reduced Order Model (ROM)

Theoretical Foundations of ROM

Why Reduce the Order at All

A detailed CAE model (FOM: Full Order Model, 10⁵–10⁸ degrees of freedom) takes minutes to hours per run, which rules it out for real-time applications (digital twins, control) and for the thousands of repeats that optimisation and UQ demand. A reduced order model (ROM) compresses the degrees of freedom to tens or hundreds by finding the low-dimensional subspace the solution actually moves in and solving the equations only inside that subspace. The price of a thousand- to ten-thousand-fold speed-up is that behaviour outside the subspace is in principle inexpressible — designing that trade is what ROM practice really is.

Building the Basis with POD

The standard basis construction is proper orthogonal decomposition (POD). Take the singular value decomposition of the snapshot matrix \( S = [\mathbf{u}_1, \dots, \mathbf{u}_m] \), assembled from FOM solutions collected across varying conditions, and keep the leading \( k \) left singular vectors as the basis \( \Phi \).

$$ S = U \Sigma V^T, \qquad \mathbf{u}(t; \mu) \approx \Phi\, \mathbf{a}(t; \mu), \quad \Phi = U_{[:,1:k]} $$

The mode count \( k \) follows from the cumulative singular-value energy (\( \sum_{i \le k}\sigma_i^2 / \sum_i \sigma_i^2 \ge 99.9\% \), for example). What matters is that how fast the singular values decay decides whether a ROM is viable at all. Diffusion-dominated problems (heat conduction, static structural analysis) pass 99% with a handful of modes, whereas advection-dominated problems (shock waves, a moving flame front) decay slowly and are fundamentally hard to reduce with a linear subspace — the Kolmogorov n-width barrier.

Projection-Based and Data-Driven ROMs

In a working environment where the internal operators of a commercial solver are out of reach, the non-intrusive route (POD plus regression, operator inference) is the realistic mainstay.

Numerical Methods for Construction

Snapshot Design — Where ROM Quality Is Decided

The basis knows nothing beyond the space its snapshots span, so snapshot planning is design of experiments. The essentials: ① cover the parameter range with LHS or similar (the same discipline as the DOE for Kriging), ② for transient problems, include representative phases of the transient (ramp-up, peak, steady state), ③ thin out snapshots whose mutual correlation is too high (watch the rank of the SVD). When mixing quantities of different scale (displacement and stress, temperature and flow velocity), forget per-quantity normalisation — or per-quantity bases — and the large quantity alone will occupy the basis.

The Nonlinear-Term Wall and Hyper-Reduction

When the reduced equations of a Galerkin ROM contain a nonlinear term, evaluating it forces a return to full dimension, and the reduction buys no speed at all. Hyper-reduction (DEIM, ECSW and relatives) cuts that loop by evaluating the nonlinear term at a small set of sample points (elements) and interpolating. The symptom "we built a ROM but only got 2× faster" is almost always hyper-reduction left unapplied. Data-driven ROMs (assuming a quadratic form in operator inference, for instance) avoid the problem entirely, which is another reason they get adopted.

Stability and Structure Preservation

Galerkin projection carries no guarantee that the stability of the original equations survives, and a ROM can diverge where the FOM is stable. Work through the remedies in this order: ① projection in an energy inner product (a symmetric positive-definite weight matrix), ② Petrov-Galerkin (least-squares) projection, ③ a closure term that supplies the missing dissipation. In fluids, the treatment of the pressure term (POD on velocity only, with pressure handled separately) also governs stability. ROM verification must cover not only reproduction of the training cases but boundedness under long-time integration.

Applying It in Practice

The Standard Workflow — Offline/Online Split

1. Offline (expensive, done once) — snapshot plan → FOM runs → POD basis → reduced model construction (projection or regression) → validation
2. Online (cheap, every time) — solve the reduced system for a new parameter or input (milliseconds to seconds)
3. Validation — compare against the FOM on hold-out parameters not used in training (report field L2 error and quantity-of-interest error separately)
4. Monitoring — in operation, check whether the input lies inside the training domain; outside it, warn and fall back to the FOM

Where It Fits — ROM Against Kriging and FNO

All of these get lumped together as "surrogates", but ROM's distinctive value is reduction as a dynamical system — it carries state and responds to arbitrary input. If all you need is a static parameter-to-output map, Kriging is simpler.

Setting Validation and Acceptance Criteria

Acceptance criteria for a ROM are worked backwards from the application. For control, "frequency-response error within ±1 dB in band"; for a twin, "error in the monitored quantity below sensor noise"; for UQ, "error in the statistics (mean, variance) below a tenth of the spread caused by input uncertainty". A small mean field error routinely coexists with a badly missed quantity of interest (a local hot spot), so always include quantity-of-interest error in the report. Confirming that results hold steady when the mode count is varied by a few modes either way (mode-count sensitivity) is a standard item too.

Tools and Implementation

Tool Support

A Minimal Non-Intrusive Recipe (for Commercial CAE Users)

1. Vary the parameters by LHS, run the FOM N times, export the field snapshots
2. Build the POD basis \( \Phi \) by SVD (mode count from the cumulative energy)
3. Compress each snapshot to coefficients \( \mathbf{a} = \Phi^T \mathbf{u} \)
4. Regress the parameter-to-coefficient map with Kriging or similar (POD+Kriging); for dynamic problems, identify the time evolution of the coefficients with operator inference or DMD
5. Validate field error and quantity-of-interest error on hold-out cases

This needs zero solver modification, and it already meets most practical needs (parametric field prediction). Moving on to a projection ROM can wait until conservation of the physics or long-time stability becomes a stated requirement.

Research Frontiers

Advection-Dominated Problems — Past the Linear-Subspace Barrier

Advection-dominated problems such as shocks and moving interfaces have singular values that refuse to decay, the weak point of linear POD. Research runs along three fronts: ① absorb the motion of the solution with a coordinate transformation before reducing (shifted POD, registration methods), ② nonlinear subspaces built on a quadratic manifold, ③ neural ROMs that learn a nonlinear latent space with an autoencoder. "Accuracy on advection problems stops improving no matter how many modes I add" is a limitation of the method itself, and these extensions are the answer to it.

Operator Inference and Closure Modelling

Operator inference — assume the form of the governing equations in reduced coordinates from the physics (linear plus quadratic terms, say) and identify only the coefficients from data — is a compromise that stays non-intrusive while preserving equation structure, and implementations are spreading. Data-driven construction of closure models that account for the truncated modes (the counterpart of an SGS model in LES) is equally active, and is the key to ROMs that stay stable over long times with few modes.

Certified ROMs with Error Bounds (Reduced Basis Method)

For elliptic and parabolic problems, the reduced basis (RB) method comes with an established theory that computes rigorous error upper bounds online. Being able to answer "the error of this ROM prediction is at most ε" carries real engineering weight, and it is the justification for adopting ROMs where safety is at stake (structural health monitoring). Extension to general nonlinear problems is still research, but within the linear range it is production-ready.

Troubleshooting

Symptoms, Causes, and Fixes

Three Conditions for Adoption

Related: Kriging surrogates, polynomial chaos, Fourier neural operators.