Reduced-order models of incompressible flow compress a detailed simulation into a few dozen patterns, called modes, that are extracted from saved simulation states. The standard extraction method is Proper Orthogonal Decomposition (POD), which is essentially principal component analysis applied to flow fields. These compressed models can run thousands of times faster than the original solver. They also fail in a specific way. Velocity and pressure must stay compatible with each other, and this compatibility is measured by an "inf-sup constant", a number that must stay above zero for the pressure to be well defined. The usual repair, supremizer enrichment, adds one extra velocity field for each pressure mode. That field is built so that it pushes the inf-sup constant up. Because supremizers are computed from the pressure modes, they are only as good as the set of saved simulations. For problems whose geometry or physics varies with parameters, that set is usually chosen by a greedy loop that repeatedly adds the parameter where an error estimate is largest.
We argue that this greedy loop can stop improving in exactly the way evolutionary search does. The estimate it chases tracks velocity error, not whether the reduced velocity–pressure pair stays compatible. The snapshot set can therefore keep concentrating where velocity is hard to capture while leaving whole regions of parameter space where the reduced model loses stability. The proposal is to make the reduced inf-sup constant itself a first-class selection signal. That constant is cheap to evaluate for any parameter once the reduced matrices exist. We then test, with ablations, whether it is this stability signal or just generic diversity in the choice of parameters that prevents the failure.
Core Claim
For a parameter vector $\mu$, let $\beta_{r}(\mu)$ be the inf-sup constant of the reduced pair, meaning the enriched reduced velocity space and the reduced pressure space:
$$\beta_{r}(\mu) = \min_{q \in Q_m} \max_{v \in V_r \oplus S_m(\mu)} \frac{b(v, q; \mu)}{\|v\|_V \, \|q\|_Q}.$$
Here $b(\cdot,\cdot;\mu)$ is the parameter-dependent coupling between velocity and pressure, $Q_m$ is the span of $m$ pressure POD modes, $V_r$ is the span of $r$ velocity POD modes, and $S_m(\mu)$ is the span of the supremizers. After an offline affine or empirical-interpolation decomposition, computing $\beta_r(\mu)$ is a generalized eigenvalue problem of size roughly $(r+m) \times m$. It therefore costs almost nothing per parameter, and $\beta_r$ can be mapped over the whole training set at every greedy step.
This quantity is distinct from what standard a posteriori bounds use. Those bounds divide a residual by the stability constant of the full model, and that constant says nothing about whether the reduced saddle-point system is solvable. Our claim is that the map $\mu \mapsto \beta_r(\mu)$ is typically rugged, with isolated pockets of near-zero values. We further claim that error-driven POD-greedy leaves these pockets unvisited, and that a selection rule which treats "maximize error estimate" and "minimize $\beta_r$" as two objectives closes them at no extra cost in full simulations.
Relation to Prior Work
Three groups of work motivate the proposal, supply its methods, or constrain its design.
The first group establishes supremizer enrichment on POD bases and exposes its parametric weakness. Schneier's error analysis of supremizer pressure recovery proves stability and pressure error bounds for momentum-equation recovery with supremizer-enriched POD. Those bounds depend on the first principal angle between the velocity and supremizer spaces. The proof is confined to a single, non-parametrized setting, and the parametric case, where the stability constant becomes $\beta(\mu)$, is left explicitly open. We study that parametric regime and use its principal angle as a diagnostic. The POD-Greedy-Galerkin introduction by Siena, Girfoglio and Rozza describes the standard loop: select the parameter that maximizes the estimate $\Delta_N(\mu)$, then enrich with supremizers. We keep that loop and change only the selection rule. The shallow-water optimal-control work uses aggregated spaces to keep a coupled system solvable, and its authors state that their parameter snapshots were chosen heuristically and that an adaptive rule remains open. That is exactly the gap we target. The geometric applications with reduced cut finite elements show that geometric parameters make $b$ genuinely parameter-dependent, which is the regime where $\beta_r(\mu)$ should vary most.
The second group supplies the diagnosis and the method. Multiple-Frequencies Population-Based Training shows that greedy selection on a short-term signal collapses diversity. There, greedy population training falls below plain random search on Brax Humanoid. We read POD-greedy's single-signal selection as an instance of that failure. Strategy-space diversity for batch Bayesian optimization avoids trapping by drawing batches from the Pareto front of two competing objectives. We borrow that idea, with our two objectives being the error estimate and the stability deficit rather than the predicted mean and uncertainty of a Gaussian-process model.
The third group constrains the design. Divergence-guided particle swarm optimization gains two- to eight-fold on multimodal functions but loses accuracy on the unimodal Sphere function, because its repulsive force never switches off. We therefore avoid always-on repulsion in parameter space. The stability objective only reshapes selection where $\beta_r$ actually drops, so on problems where $\beta_r$ is uniformly healthy the method should reduce to ordinary POD-greedy. Finally, Rozza's stabilization studies (high Péclet reduced bases, SUPG reduced order method for optimal control) show that stabilization applied only offline is lost when the online projection is plain Galerkin, and that adding modes past the numerical rank damages conditioning. Both results constrain us. Stability must be checked in the reduced operator actually used online, and enrichment must be judged by conditioning as well as accuracy.
Research Program
The method keeps standard POD-greedy with supremizer enrichment and changes only how parameters are picked. At each iteration we have the current reduced spaces. We evaluate two quantities on the training set $\Xi_{\text{train}}$: the existing error estimate $\Delta_N(\mu)$, and the stability deficit $-\log \beta_r(\mu)$. We then choose a batch of $k$ parameters from the non-dominated front of these two objectives, spread along the front as in the batch Bayesian optimization work. The full model is run at those parameters, their velocity and pressure trajectories are compressed by POD and appended to the bases, and supremizers are recomputed for the new pressure modes. To guard against supremizers that nearly duplicate each other, a new supremizer is kept only if it raises $\min_{\Xi_{\text{train}}} \beta_r$. Otherwise it is dropped and the condition number is logged. This is our answer to the redundancy problem Schneier's principal-angle dependence points to.
The analysis component has two parts. First, we map $\beta_r(\mu)$ across parameter space at each greedy step. This tests whether low-stability pockets exist and whether error-greedy leaves them unvisited. Second, we test whether $\beta_r$ at unvisited parameters is predicted by the distance, in subspace angle, between the pressure modes they would contribute and the current pressure space. If it is, that would give a partial explanation of the pockets.
Baselines
The naive reference is space-filling sampling. We use Latin hypercube samples with the same number of full-model solves, followed by the same POD and supremizer pipeline. This is the method a practitioner would reach for first. It is embarrassingly parallel, and the population-training results suggest random selection can beat greedy selection that has collapsed. It beats our method if it matches our worst-case $\beta_r$ and error at equal cost.
The strongest existing method is standard POD-greedy with supremizer enrichment as described by Siena et al.. We run it with the same training set, the same estimator, the same number of full-model solves, the same tolerance for truncating POD modes, and the same batch size $k$. For batch size above one, it takes the top $k$ parameters by estimate. We give it its strongest form by also running a variant that uses supremizers from every pressure snapshot rather than only from the pressure modes. On the shallow-water benchmark, the relevant reported figures come from the heuristic-sampling ROM in the shallow-water paper. With $N=30$ basis functions it reaches average relative errors of about $10^{-3}$ for state and adjoint velocity, about $10^{-4}$ for elevation and control, and a speedup index of about 30. Beating it means lower worst-case error at $N=30$, with no loss of speedup. For POD-greedy on parametrized Navier–Stokes with supremizers, no reported result was found on Lacuna, so we will generate that baseline ourselves.
Three ablations isolate the key ingredient. The first replaces the Pareto selection with a scalarized score, the error estimate divided by $\beta_r$. The second keeps Pareto batching but swaps the stability objective for a generic diversity objective: the minimum distance to parameters already selected, in the spirit of divergence-guided repulsion. The third keeps our selection rule but removes the supremizer-acceptance test. If the second ablation matches the full method, the gain comes from diversity rather than from stability awareness, and the proposal's central claim fails.
Why not just…?
Why not just sample the parameter space uniformly or randomly? For two or three parameters, this may genuinely suffice, and we expect it to be competitive there. The POD-greedy introduction notes that grid sampling becomes prohibitive beyond that. Low-$\beta_r$ pockets that are small in volume are exactly what random samples miss at a fixed budget. The Latin hypercube baseline, run on four-parameter cases, settles this directly.
Why not just divide the error estimate by the reduced inf-sup constant? This is the most serious threat, because it is one line of code. If the scalarized ablation matches Pareto selection, we will adopt it and report that the useful contribution is the cheap $\beta_r$ signal, not the multi-objective machinery. We expect scalarization to be dominated by whichever term is currently largest in magnitude, and the experiment will show whether that matters.
Why not just add more POD modes or more supremizers? Rozza's SUPG reduced order study shows that adding modes past the numerical rank buys rounding error, not accuracy. Inf-sup stability is a compatibility condition between two spaces, so extra velocity energy does not repair a pressure space that never saw a given regime. The all-snapshot supremizer baseline tests the "more supremizers" version directly.
Why not just recover pressure by a Poisson equation and skip supremizers? Schneier shows that pressure recovered from a Poisson equation stagnates near solid boundaries because of an artificial Neumann boundary condition. We include Poisson recovery on the annular disk only as a reference point.
Proposed Experiments
The primary testbed extends Schneier's eccentric annular disk, discretized with Taylor–Hood elements, by adding parameters for the offset of the inner disk, its radius, and the viscosity. Pushing the inner disk toward the outer wall creates a narrow gap where pressure modes become localized. This region is a deliberately planted candidate low-$\beta_r$ pocket. As a sanity check at the unparametrized setting, we will first confirm that our pipeline reproduces the reported pressure-error scaling of roughly $A_m^{0.993}$ and $A_r^{1.32}$. The second testbed is a finite-element cylinder wake whose Reynolds-number and geometry ranges follow the variation categories in CFDBench. CFDBench's own data are finite-volume, so the snapshots are regenerated with finite elements. The third testbed is the shallow-water spreading Gaussian control problem, where the selection rule replaces the heuristic sampling reported in that paper.
The metrics are as follows. The primary stability metric is the minimum and fifth-percentile $\beta_r$ over a dense held-out parameter set, together with the fraction of held-out parameters where the reduced system is ill-conditioned or its Newton solve fails. We also report maximum and mean relative errors in velocity and pressure, the condition number of the reduced system, and the number of full-model solves. Every method is run with five random training-set draws.
The decisive check is on the annular disk. We ask whether error-driven POD-greedy leaves the narrow-gap region unselected while $\beta_r$ collapses there, and whether our rule, at equal cost, raises the held-out minimum $\beta_r$ and cuts worst-case pressure error while leaving mean error unchanged. A second check guards against the always-on penalty seen with divergence-guided swarms. On a viscosity-only parametrization where $\beta_r$ is uniformly healthy, our method must match plain POD-greedy. If it is worse there, the stability objective is acting as an unwanted repulsion and must be gated more tightly.
References
- Error Analysis of Supremizer Pressure Recovery for POD based Reduced-Order Models of the Time-Dependent Navier-Stokes Equations
- An introduction to POD-Greedy-Galerkin reduced basis method
- POD-Galerkin model order reduction for parametrized nonlinear time-dependent optimal flow control: an application to shallow water equations
- A Reduced Order Cut Finite Element method for geometrically parametrized steady and unsteady Navier-Stokes problems
- Multiple-Frequencies Population-Based Training
- Exploiting Strategy-Space Diversity for Batch Bayesian Optimization
- Divergence-Guided Particle Swarm Optimization
- Reduced Basis Approximation of Parametrized Advection-Diffusion PDEs with High Péclet Number
- A Streamline Upwind Petrov-Galerkin Reduced Order Method for Advection-Dominated Partial Differential Equations Under Optimal Control
- CFDBench: A Comprehensive Benchmark for Machine Learning Methods in Fluid Dynamics