Model Order Reduction Methods
Theory: representing solutions with a few modes
Overview
I've heard hours-long analyses can be replaced by reduced models that answer instantly. Is that true?
With conditions, yes. Solutions of many physical problems, however many unknowns they have, can often be expressed as combinations of a few characteristic shapes. POD solves the full model under several conditions, applies a singular value decomposition to those solutions, and extracts shapes (modes) that best represent the variation. Then only modal coefficients need solving, shrinking tens of thousands of unknowns to a handful or dozens. But modes know only the shapes in the training solutions; conditions producing a different kind of variation cannot be represented however many modes are added. So when building a reduced model, the training must cover the range of use, and accuracy outside it must be checked.
Proper orthogonal decomposition (POD)
Reduced model (Galerkin projection)
Example: 1D transient conduction in a unit rod (200 interior nodes, zero temperature at both ends). A Gaussian source (width 0.05) varies as 1 + 0.5 sin 8πt; 500 snapshots from backward Euler (step 0.001) to time 0.5 are used.
From turbulence structure to design tool
POD shares the idea of principal component analysis in statistics; Lumley brought it into fluid dynamics in 1967 to extract large structures recurring in turbulence. Sirovich's snapshot method in the late 1980s eased computation, and it spread to flow control and reduced models. It is now a core technology for reducing coupled structural, thermal and electromagnetic analyses and for digital twins estimating machine states in real time.
Worked examples
Example 1: number of modes and captured energy
| Modes | Uncaptured energy | Max error (training case) | Max error (source moved to 0.6) |
|---|---|---|---|
| 2 | 1.2×10⁻⁵ | 1.1% | 22.5% |
| 4 | 1.8×10⁻⁹ | 0.037% | 19.2% |
| 8 | ≈ 0 | 4.4×10⁻⁶ % | 18.7% |
| 16 | ≈ 0 | 1.3×10⁻¹¹ % | 18.5% |
Example 2: training on three source positions (0.4, 0.5, 0.7)
| Modes | Uncaptured energy | Max error at untrained 0.6 |
|---|---|---|
| 4 | 5.7×10⁻⁵ | 12.1% |
| 8 | 5.7×10⁻⁹ | 0.71% |
| 16 | ≈ 0 | 0.053% |
| 24 | ≈ 0 | 0.0011% |
Error is essentially zero for the training case, yet moving the source slightly misses by 19%.
Training only with a central source makes every extracted mode symmetric about the centre. Moving the source to 0.6 shifts the temperature peak right, and no sum of symmetric shapes can represent a shifted peak, so adding modes doesn't reduce error. This is the biggest pitfall of reduced models: small error on the training case guarantees nothing about accuracy where you want to use it. Training with varied source positions adds modes that represent shifted shapes, and error at the untrained position 0.6 fell sharply. In practice, choose training conditions spanning the edges and interior of the range of use and always validate on conditions not used in training. Some systems also monitor the full model's residual inside the reduced model and fall back to the full model when the range is exceeded.
Building a reduced model
- Define the range of use (loads, boundary conditions, material and geometry parameters).
- Generate training solutions with the full model spanning edges and interior of the range.
- Choose mode count from singular-value decay.
- Check error on conditions not used in training.
- For nonlinear terms, consider hyper-reduction to keep evaluation fast.
“The digital twin failed in winter”
A reduced model estimating plant temperatures in real time matched measurements through summer and autumn but drifted several degrees in winter. All training solutions had been computed with outdoor temperatures above 15 °C, and at low winter temperatures the shape of the temperature field itself differed from the trained shapes. Adding winter conditions to training and rebuilding the modes brought errors within 1 °C. To watch for out-of-range use, the residual of the reduced solution substituted into the full equations is now monitored.
Common mistakes
Mistakes and fixes
| Mistake | Effect | Fix |
|---|---|---|
| Validating only on training data | Large errors outside range | Validate on unseen cases |
| Narrow training range | More modes don't help | Include range edges |
| Choosing modes by energy alone | Local errors remain | Check target quantities |
| Evaluating nonlinear terms at full size | No speed-up | Use hyper-reduction |
| No out-of-range monitoring | Wrong results go unnoticed | Monitor residuals |
I'd like to learn related topics.
Related pages include reduced-order multiphysics, multiphysics digital twins, data-driven multiphysics, staggered coupling and data transfer between non-matching meshes.