Model reduction - Ciallo Wiki

1 Problem Setting and Motivation

1.1 Full-order models and computational bottlenecks

Model reduction begins with a “full-order model” (FOM): a high-fidelity mathematical description with many degrees of freedom, such as a discretized PDE, a detailed multibody dynamics model, or a finely resolved system identification model. While the FOM can be accurate, it often becomes costly to evaluate repeatedly because the state dimension is large, the operators are dense or expensive to apply, and time integration requires many solver steps.

The primary bottlenecks include large linear solves or frequent nonlinear iterations, substantial memory use for storing discretized operators and states, and the need to generate large amounts of high-resolution data to cover different scenarios. In many workflows, the dominant cost is not the single evaluation of the FOM, but the sheer number of evaluations needed for design loops, uncertainty quantification, parameter sweeps, or optimization.

1.2 Accuracy–efficiency trade-offs

A reduced-order model (ROM) targets a balance between computational savings and fidelity. Reduction can significantly lower runtime, but accuracy may degrade if the reduced space is too small, if the dynamics change qualitatively outside the training regime, or if nonlinear effects are not represented well.

The trade-off is typically managed by choosing basis dimension and approximation spaces, selecting reduction strategies appropriate to the system structure (linear, nonlinear, parameterized), and monitoring errors using metrics tied to the intended outputs. Since “acceptable accuracy” depends on the application, the reduction goal is usually task-specific rather than uniformly absolute.

1.3 Types of reduction tasks (simulation, forecasting, control, inverse problems)

Model reduction is used in multiple settings:

Simulation: accelerate time-domain or steady computations while approximating the state evolution.
Forecasting: predict future behavior from partial observations or initial conditions with reduced computational load.
Control: enable fast evaluation inside optimization or feedback loops, where control synthesis may require many model calls.
Inverse problems: estimate parameters or hidden inputs by matching model outputs to observations; reduced models can make iterative inference feasible.

Across these tasks, the same foundational idea—replacing the FOM with a surrogate—takes different forms in terms of what is approximated (states, outputs, operators) and what constraints are imposed.

1.4 Offline–online workflow concepts

Many reduction methods follow an offline–online pattern. In the offline phase, one constructs the reduced representation using training data, basis functions, operator projections, or learned parameters. This stage can be expensive but is done once (or infrequently).

In the online phase, evaluations are performed rapidly by using the precomputed reduced operators and low-dimensional state variables. This separation is crucial for real-time or many-query settings, where the cost of repeated FOM evaluations would otherwise be prohibitive.

2 Mathematical Foundations

2.1 Linear systems and state-space formulations

For linear dynamics, reduction often starts from a state-space representation, typically written as \[ \dot{x}(t)=Ax(t)+Bu(t), \quad y(t)=Cx(t)+Du(t), \] or from a discretized linear system for PDEs. Model reduction in this context seeks a lower-dimensional coordinate \(x_r\) such that \(x \approx Vx_r\), where \(V\) is a basis spanning the reduced subspace.

Projection methods transform the original operators into reduced operators (e.g., \(A_r = V^\top A V\) in symmetric settings), enabling the reduced model to evolve in a smaller space while preserving specified properties such as approximate input–output behavior, dominant modes, or selected invariants.

2.2 Nonlinear systems and projection principles

For nonlinear systems, the same state approximation idea is used, but nonlinear terms introduce additional complexity. If the full dynamics can be written schematically as \[ \dot{x}=f(x,u), \] then substituting \(x \approx Vx_r\) leads to reduced equations involving projected nonlinearities. A naive approach still requires evaluating the nonlinear term at high dimension, which undermines speedups.

Consequently, nonlinear reduction methods emphasize ways to evaluate the reduced nonlinear contribution efficiently—often through hyper-reduction, interpolation of nonlinear terms, or calibrated operator approximations.

2.3 Error metrics and norms

Assessing reduction quality requires defining errors in a mathematically meaningful way. Common choices include norms of the state error, output error, or residual-based measures. Depending on application, one may prioritize relative error in key regions, error in integral quantities, or accuracy of derived outputs (e.g., lift/drag, stresses, or sensor readings).

In practice, errors are computed over time intervals, parameter samples, or ensembles. Robust evaluation often uses both training and testing data to detect whether improvements are genuine or confined to the training set.

2.4 Stability, consistency, and convergence concepts

Reduced models can fail even when approximation accuracy appears good, because projection can alter stability characteristics. Stability notions include boundedness of trajectories, avoidance of spurious oscillations, and persistence of dissipative behavior when expected. Consistency concerns whether the reduced equations accurately reflect the original dynamics within the reduced approximation space.

Convergence relates to how errors decrease as the reduced dimension increases or as discretization parameters refine. For some classes of problems—especially linear symmetric systems—stronger guarantees are attainable, while for general nonlinear systems, stability and convergence often require additional conditions or empirical validation.

2.5 Parameterized models and reduced representations

2.5.1 Parametric dependence and sampling in parameter space

In many applications, the model depends on a parameter vector \(\mu\) (material properties, geometry parameters, boundary conditions, operating conditions). A parametric FOM can be written as \(x(t;\mu)\) governed by operators or forcing terms that vary with \(\mu\). ROMs can be constructed to approximate the solution manifold across parameter space.

Training typically involves sampling parameter points to capture representative behavior. The sampling design influences both basis quality and online accuracy, with better coverage generally improving generalization. When parameter dependence is complex, methods such as adaptive sampling, greedy selection, or space-filling designs are used to focus where the reduced model needs improvement.

3 Projection-Based Reduced-Order Modeling

3.1 Proper Orthogonal Decomposition (POD)

3.1.1 Snapshot generation and training data

POD relies on collecting snapshots: samples of the system state (or relevant fields) over time, parameter variations, or operating conditions. Snapshot generation requires choices about time resolution, which trajectories to simulate, and how to cover parameter and initial-condition diversity.

The quality of the snapshot set is often decisive. If snapshots miss dominant behaviors, the reduced basis will not include the corresponding modes, and later evaluations can show systematic bias.

3.1.2 POD basis construction and truncation

Given snapshot data, POD constructs an orthonormal basis that optimally represents the snapshots in a least-squares sense. Mathematically, this is tied to eigenvalue or singular value decompositions of the snapshot correlation structure (or an equivalent formulation). Truncation retains only the leading modes, typically ordered by their energy contribution.

Truncation controls complexity: keeping more modes improves fidelity but increases reduced dimension and associated cost in online evaluations. Selecting an appropriate cutoff thus determines the ROM’s balance of accuracy and speed.

3.1.3 Energy capture and dimension selection

Energy capture refers to how much of the snapshot variance or squared magnitude is explained by the retained modes. A common selection strategy uses a threshold on cumulative energy (or an analogous criterion), ensuring that the reduced basis captures the most influential directions in the data.

For practical deployment, energy criteria are sometimes supplemented with task-based validation (e.g., output error or residual accuracy) because variance in the state does not always translate directly to accuracy in the quantities of interest.

3.2 Galerkin projection

In Galerkin projection, the reduced approximation \(x \approx Vx_r\) is substituted into the governing equations, and the residual is forced to be orthogonal to the same space spanned by \(V\). This yields reduced equations for \(x_r\).

For linear systems, Galerkin projection produces reduced operators that preserve certain structural properties when the original problem and basis are compatible (for instance, symmetry under appropriate inner products). For nonlinear systems, Galerkin projection still requires efficient handling of nonlinear terms, which motivates later techniques such as hyper-reduction.

3.3 Petrov–Galerkin projection

3.3.1 Choosing test spaces for improved accuracy

Petrov–Galerkin projection generalizes Galerkin by using a different test space \(W\) from the trial space \(V\). The reduced system is obtained by enforcing orthogonality of the residual to \(W\), leading to equations involving \(W^\top\) projections.

This flexibility can improve accuracy, stability, or approximation of certain behaviors, especially when the trial space alone is insufficient. The choice of test functions can be guided by theory for specific problem classes, numerical experiments, or moment-matching concepts.

3.4 Interpolation and constrained bases

Some projection approaches incorporate interpolation constraints or structured basis design to better match key solution features. Constrained bases can enforce exactness for selected values, moments, boundary-related behaviors, or conservation laws (when such constraints are formulated in compatible ways).

Interpolation-based ideas can also be used to avoid purely data-driven spaces. By anchoring the reduced representation to known properties or selected solution snapshots, the ROM may improve extrapolation within limits and reduce bias in regions critical to the application.

4 Reduced Models for Nonlinear Dynamics

4.1 Challenges: nonlinearity and closure

Nonlinear reduction faces a central difficulty: projecting the state does not automatically yield a cheap reduced representation of nonlinear terms. The reduced equations may require computing nonlinear contributions in the full space, losing efficiency.

Another issue is closure: the reduced variables may not uniquely determine the nonlinear term unless additional approximation steps are included. Without a closure strategy, the ROM can become both slow and inaccurate, particularly for strongly nonlinear systems or those with sharp transitions.

4.2 Hyper-reduction techniques

Hyper-reduction aims to evaluate nonlinear terms efficiently by approximating the nonlinear contribution directly in a way compatible with the reduced representation. Instead of using high-dimensional operations, hyper-reduction reconstructs nonlinear effects from a smaller set of evaluations or sampled components.

4.2.1 Empirical interpolation method (EIM)

EIM approximates nonlinear functions (or nonlinear operators) using an expansion built from carefully selected sample points and basis functions. It learns how the nonlinear term behaves across the reduced state manifold and then expresses nonlinear evaluations through a reduced set of interpolation components.

The method typically involves an offline stage to construct interpolation data and an online stage that evaluates the nonlinear term efficiently using precomputed interpolation coefficients.

4.2.2 Discrete empirical interpolation method (DEIM)

DEIM is a related approach that selects a subset of entries (components) of the nonlinear term and uses these sampled values to reconstruct an approximation in the full or reduced context. This can dramatically reduce computational cost by avoiding evaluation of all nonlinear components.

DEIM introduces a trade-off: fewer sampled components mean faster evaluations but potentially larger approximation error. Selection strategies for sample indices are therefore important, often relying on greedy procedures tied to basis construction.

4.2.3 Greedy sampling for nonlinear terms

Greedy algorithms are commonly used to pick snapshot points, interpolation samples, or reduced basis elements in a manner that targets maximal expected error reduction. For nonlinear hyper-reduction, greedy sampling often selects the next sample based on a criterion like residual magnitude in the nonlinear approximation.

This iterative selection can improve representation of nonlinear behavior in parts of the state space that are critical for accuracy while keeping the number of sampled evaluations manageable.

4.3 Operator interpolation and ROM calibration

An alternative to approximating nonlinear terms directly is to interpolate reduced operators across the reduced coordinates or parameters. Operator interpolation can be used when the reduced model contains operators that depend on the current state or parameter values.

Calibration refers to tuning aspects of the ROM—such as scaling, correction terms, or interpolation coefficients—using data. Well-designed calibration can correct systematic biases, but poorly chosen calibration can overfit and reduce reliability outside the training conditions.

4.4 Stability handling in reduced nonlinear models

Stability is a key concern because reduced nonlinear dynamics can generate artifacts such as growing oscillations or unphysical energy production. Stability handling may involve modifying the reduced equations, adding regularization terms, using stabilization in the projection framework, or enforcing constraints tied to dissipative structure.

Some strategies aim to preserve known physical or mathematical properties (e.g., energy balance) in the reduced model, while others rely on monitoring stability indicators during online evaluation and applying corrective damping when necessary.

5 Data-Driven Reduction Methods

5.1 Dynamic Mode Decomposition (DMD)

DMD analyzes time-series data to extract modes and growth/decay rates that approximate the system evolution. It typically treats the dynamics as approximately linear in a lifted space, producing a set of eigenmodes that can reconstruct future states.

A practical concern is snapshot alignment and preprocessing: the choice of sampling times, the handling of noise, and how snapshots are paired (current-to-next) affect the quality of the inferred modes. DMD outputs are often interpretable as coherent structures, though the interpretation depends on the system and data richness.

5.2 Koopman operator-based approaches

Koopman theory reframes nonlinear dynamics by considering a linear operator acting on observables rather than directly on states. Under this view, nonlinear systems can be approximated by linear evolution of appropriately chosen functions of the state.

In practice, Koopman-based methods approximate these observables using bases, kernel methods, or neural architectures. Model reduction then becomes an exercise in learning an effective representation where linear operator dynamics capture the nonlinear evolution well enough for prediction or control.

5.3 Subspace learning and manifold methods

Instead of representing dynamics in a linear reduced space, manifold methods seek lower-dimensional structure in which system states evolve. Subspace learning may construct time-dependent or state-dependent reduced coordinates, capturing nonlinear geometry missed by purely linear POD-type bases.

Common ideas include learning embeddings, building reduced coordinates via autoencoder-style mappings, or approximating invariant manifolds. These approaches can offer improved accuracy when dynamics lie on curved structures, though they require careful training and validation to avoid misleading extrapolation.

5.4 Neural-network surrogates

Neural surrogates learn a mapping from reduced inputs (or raw states) to outputs such as state increments, full states, or residuals. They can be combined with reduction concepts by training networks in a reduced coordinate system, predicting coefficients in a ROM basis, or learning closures for nonlinear terms.

A major advantage is flexibility: neural models can capture complex nonlinear dependencies without explicitly deriving reduced operators. The main limitations involve data requirements, generalization risk, and the difficulty of guaranteeing stability or constraint satisfaction.

5.4.1 Physics-informed neural surrogates (ROM perspective)

Physics-informed approaches incorporate governing equations, conservation constraints, or residual losses into neural training. From a ROM viewpoint, this can be interpreted as enforcing structure so that the learned surrogate behaves consistently with the underlying dynamics.

Such methods can reduce the need for extensive labeled data and improve extrapolation within physically plausible regimes, though effectiveness still depends on how constraints are formulated and whether the model class can represent the true dynamics.

5.5 Uncertainty-aware data-driven reduction

Data-driven ROMs may be sensitive to noise, limited sampling, and distribution shift. Uncertainty-aware strategies aim to quantify or mitigate these issues using ensembles, Bayesian methods, probabilistic calibration, or uncertainty quantification in learned coefficients.

Uncertainty estimates help decide when a ROM prediction can be trusted and can guide adaptive sampling or conservative decision-making in downstream tasks like optimization or control.

6 Model Reduction for Partial Differential Equations (PDEs)

6.1 Spatial discretization and reduced basis connection

For PDEs, the FOM typically arises from spatial discretization, producing a large system of ODEs. Reduced basis methods then approximate the solution field as a combination of basis functions in space. In many cases, the basis functions are obtained from snapshot data or from problem structure, then used to represent the discretized state with fewer degrees of freedom.

The connection between spatial discretization and reduced representations is crucial: the ROM must operate consistently with the chosen inner products, discretization schemes, and boundary treatment so that the reduced equations reflect the original PDE behavior.

6.2 Boundary conditions and reduced spaces

Boundary conditions can be challenging in reduced spaces because basis functions may not satisfy constraints exactly. Two common strategies are to construct basis functions that inherently satisfy boundary conditions or to use techniques that enforce them weakly or through augmentation.

Improper handling can lead to systematic errors near boundaries, loss of accuracy for fluxes, and sometimes instabilities. Thus, boundary-aware basis design or constraint enforcement is often essential for reliable PDE ROMs.

6.3 Time integration in reduced models

6.3.1 ROM time stepping and projection consistency

Reduced models still require time integration. The choice of integrator (explicit, implicit, or structure-preserving) affects stability and accuracy. Consistency between the reduced projection and the time-stepping scheme matters: if the ROM uses projections that conflict with the time discretization assumptions, errors can accumulate or oscillate.

In nonlinear cases, time integration also interacts with hyper-reduction: efficient evaluation of nonlinear terms must be compatible with the integrator’s repeated calls within each time step.

6.4 Error estimation in PDE contexts

6.4.1 A posteriori error indicators (conceptual)

PDE ROM quality is often assessed via residual-based or estimator-based approaches that measure how well the reduced solution satisfies the governing equations. A posteriori indicators aim to estimate the error after solving the reduced system, sometimes without requiring access to the full FOM.

Conceptually, these estimators can guide adaptive basis enrichment, prompting the ROM to increase dimension or update sampling where errors are likely largest. While exact guarantees depend on problem class and estimator design, residual indicators are widely used for practical error control.

7 Parameterization and Multi-Query Efficiency

7.1 Reduced models for parametric systems

For parametric PDEs or parameterized dynamical systems, ROMs can be constructed to approximate solutions across a range of parameter values. The reduced representation may involve basis construction from snapshots at several parameter points, along with reduced operators that depend on parameters.

When operator dependence is structured (e.g., affine in parameters), online evaluations can be efficient by reusing precomputed components. When dependence is general, interpolation, empirical techniques, or calibration may be required to retain speedups.

7.2 Training strategies and design of experiments

Training for parametric ROMs commonly uses design-of-experiments approaches to select parameter samples. Strategies include grid-like designs for small dimensions, space-filling methods for moderate dimensions, and adaptive sampling for higher-dimensional or complex regimes.

Effective training seeks coverage of the parameter space while limiting the number of expensive FOM solves. In many workflows, the sampling strategy is adjusted after initial runs based on observed errors or changes in dominant modes.

7.3 Surrogate reuse and transfer across parameter regimes

In multi-query settings, the same ROM basis and reduced operators may be reused across many parameter evaluations. When parameter regimes shift significantly, transfer learning ideas may allow reuse of parts of the ROM while updating basis or calibrations.

A ROM that works only in a narrow range may require multiple local models or a piecewise parameterization strategy. Transfer across regimes can reduce overall cost, provided that validation confirms performance where the ROM is applied.

7.4 Complexity analysis: speedup and storage

A ROM’s efficiency depends on more than runtime. Complexity analysis accounts for:

cost of offline basis construction,
online cost of projecting states and evaluating reduced operators,
storage needs for basis matrices, reduced operators, and interpolation data,
additional overhead for hyper-reduction or neural inference.

Speedup is typically reported as a function of reduced dimension versus full dimension, but memory and initialization overhead may also affect real deployment. For large-scale problems, storage and operator assembly can dominate if not handled carefully.

8 Certification, Validation, and Verification

8.1 Verification: checking numerical correctness

Verification addresses whether the reduced model computations are implemented correctly and solve the reduced equations accurately. This includes checking solver convergence, projection consistency, and numerical stability for the reduced system.

Verification can be performed by comparing ROM outputs to trusted reference solutions for a limited set of states, validating reduced operators, and ensuring that numerical errors do not mask modeling errors.

8.2 Validation: comparing against reference/experiments

Validation evaluates whether the ROM predictions match real-world observations or high-fidelity benchmarks relevant to the task. Depending on availability, reference data may come from experiments, a more accurate simulation, or carefully controlled FOM runs.

Validation should use data not employed in basis construction whenever possible. It also requires defining what constitutes success in the context of the outputs of interest, such as error in observables, not only in full state vectors.

8.3 ROM error bounds and confidence measures

Certification seeks confidence that ROM errors remain within acceptable limits. Some methods provide theoretical bounds under specific assumptions, while others use empirical confidence intervals derived from testing and uncertainty quantification.

Even when strict bounds are unavailable, confidence measures can still guide safe use. For example, a residual-based indicator or learned uncertainty can be used to flag predictions made far from the training distribution.

8.4 Robustness to perturbations and noise

Real data and model inputs may include measurement noise, perturbations, or mismatch in initial conditions. Robust ROMs maintain performance under such perturbations, either through stable reduced dynamics, noise-robust training, or regularization.

Robustness assessments involve testing against perturbed inputs and analyzing whether errors grow gracefully or exhibit brittle behavior. For data-driven ROMs, robustness also includes sensitivity to changes in sampling cadence and data quality.

9 Practical Implementation Considerations

9.1 Pipeline: data collection to deployment

A typical ROM pipeline includes:

collecting snapshot data from the FOM (and potentially experimental or observational data),
preprocessing and selecting relevant state components,
building the reduced basis and reduced operators,
validating offline choices using test data,
integrating the ROM into an online workflow with appropriate solvers and interfaces.

Deployment requires additional attention to numerical parameters, runtime constraints, and consistency between the offline training assumptions and the online inputs encountered in production.

9.2 Choosing basis dimension and regularization

Selecting reduced dimension affects accuracy, runtime, and stability. Too few modes can miss essential dynamics; too many modes can increase cost and may amplify noise. Regularization may be applied during basis construction or in learning models to prevent overfitting and improve conditioning.

Practical selection often relies on combining energy-based criteria with error metrics on validation sets. Regularization strength may require tuning, especially for noisy snapshot data.

9.3 Numerical conditioning and solver considerations

Reduced computations can suffer from numerical conditioning issues, particularly when bases are nearly collinear, when operators are poorly scaled, or when hyper-reduction introduces approximations that amplify errors.

Solver choices in the reduced system—linear solvers, tolerances, and iteration limits—affect both accuracy and stability. Careful scaling and consistent inner-product definitions help mitigate conditioning problems.

9.4 Software toolchains and interfaces

Implementation frequently uses numerical linear algebra libraries for decompositions, projection assembly, and reduced operator evaluation. For PDE ROMs, additional frameworks may handle discretization, basis management, and boundary treatments.

Interfacing a ROM with existing simulation or control software requires standardizing data formats (state, inputs, outputs), ensuring reproducible offline results, and providing stable online evaluation APIs with predictable performance characteristics.

10 Applications and Use Cases

10.1 Structural dynamics and vibrations

In structural dynamics, ROMs can approximate vibration modes and transient responses of mechanical systems. Reduced bases built from modal or time-domain snapshots enable faster evaluation for design optimization, durability studies, or rapid uncertainty exploration.

When nonlinearities such as contact or material nonlinearity are present, hyper-reduction or closure learning may be used to maintain speed without sacrificing critical response features.

10.2 Fluid dynamics and flow-field reduction

Fluid problems often have high-dimensional discretizations. ROMs can represent dominant flow structures using POD-like modes, DMD-derived dynamics, or learned embeddings. Hyper-reduction may be required to handle nonlinear advection and turbulence-related effects in a computationally efficient manner.

Key application outputs can include pressure distributions, forces, wake dynamics, or reduced-order control signals, rather than full field reconstruction.

10.3 Thermal and multiphysics systems

Thermal systems and coupled multiphysics models can benefit from ROMs that capture cross-domain interactions with fewer degrees of freedom. Reduced bases may include combined state vectors or separately reduced components with coupling terms approximated in reduced coordinates.

Because multiphysics systems can exhibit complex parameter sensitivities, training strategies that include relevant coupling regimes improve generalization.

10.4 Real-time simulation and rapid design

ROMs enable near-instant evaluations that are otherwise infeasible with full-fidelity models. Real-time simulation supports interactive design exploration, rapid what-if analyses, and responsive visualization.

Achieving real-time performance depends on careful runtime profiling, avoiding costly online operations, and ensuring that online projections and nonlinear evaluations remain within strict time budgets.

10.5 Control and optimization with ROMs

Optimization and control often require repeated objective and constraint evaluations. ROMs can reduce the cost of these evaluations, allowing faster convergence of optimizers and higher controller update rates.

In closed-loop settings, stability and robustness are particularly important because model errors can feed back into control decisions. Practical implementations may incorporate constraint handling, conservative tuning, or online monitoring to mitigate risk.

11 Emerging Directions

11.1 Hybrid physics–data reduction

Hybrid approaches combine structure from physical models with flexibility from data-driven components. For instance, a ROM may use projection for the main dynamics while learning the missing closure or correcting operators using data.

These methods aim to reduce training data needs, improve interpretability, and enhance generalization by respecting known governing relations while compensating for modeling deficiencies.

11.2 Adaptive and online-updating ROMs

Adaptive ROMs update their basis or parameters based on new information encountered during deployment. Online updating can help handle distribution shift, evolving operating conditions, or initially underestimated regimes.

However, updating requires safeguards to prevent drift, preserve stability, and avoid excessive overhead. Practical designs often combine conservative updates with periodic revalidation.

11.3 Scalable algorithms for large-scale problems

As full-order problems grow, scalable reduction must manage large snapshot sets, expensive decompositions, and memory constraints. Techniques such as randomized linear algebra, incremental basis construction, and distributed snapshot management aim to make reduction feasible at larger scales.

Scalability also includes reducing offline computational cost and streamlining operator assembly for high-dimensional discretizations.

11.4 Benchmarking and reproducibility practices

Benchmarking compares methods across standardized problem suites, metrics, and evaluation protocols. Reproducibility practices include publishing datasets or snapshot generation details, specifying hyperparameters, and providing reference implementations.

These practices help the community distinguish genuine performance improvements from artifacts of evaluation choices and enable more reliable method selection.

12 Common Pitfalls (Lessons Learned)

12.1 Overfitting to training snapshots

A ROM can fit training data well yet perform poorly on new conditions. Overfitting occurs when the basis or learned components capture noise, transient artifacts, or unrepresentative behaviors.

Using separate validation sets, monitoring error as a function of reduced dimension, and limiting model complexity help reduce this risk.

12.2 Inadequate basis coverage

If training snapshots do not cover the relevant dynamics—such as missing parameter regimes, insufficient initial-condition variety, or too-short time windows—the reduced space may omit critical modes. This typically leads to systematic underprediction, inability to reproduce transients, or incorrect steady behavior.

Basis enrichment strategies and improved sampling designs can address coverage gaps, but they require additional offline data or computations.

12.3 Hidden instabilities from projection choices

Projection can introduce instabilities not visible in training metrics computed at limited conditions. For nonlinear systems, instabilities may manifest only after long time integration or under parameter extremes.

Stability-aware projection choices, consistent time integration, and residual/energy monitoring can help detect and mitigate such issues early.

12.4 Misleading metrics and uncontrolled extrapolation

Metrics computed solely on state vectors can misrepresent errors in quantities of interest. Conversely, output-based metrics might hide deficiencies in internal dynamics that later affect stability or prediction quality.

Extrapolation beyond the training manifold is another frequent failure mode. Even with good training errors, performance can degrade sharply outside the region where the ROM assumptions hold.