Data-Driven ROM Architecture

• Data-driven ROM architecture is a framework that leverages observed or simulated data to construct low-dimensional surrogate models for simulating complex dynamical systems.
• It integrates techniques like POD compression, neural networks, and implicit time integration to overcome instability and closure issues seen in classical projection methods.
• The approach delivers significant computational speed-ups and improved accuracy, making it viable for applications involving fluid flows, multiphysics phenomena, and nonlinear systems.

A data-driven reduced-order modeling (ROM) architecture refers to a mathematical and algorithmic framework that leverages observed or simulated data to construct efficient, low-dimensional representations of complex dynamical systems, typically partial differential equations (PDEs) or high-dimensional ODEs. Unlike classical ROMs that rely primarily on the projection of governing equations onto low-dimensional subspaces, data-driven architectures infer surrogate dynamical models directly from data, often bypassing or augmenting intrusive projection steps. These methods span approaches based on neural networks, operator regression, filtering and closure modeling, and hybrid methodologies that inject physics-based constraints into the data-driven learning process.

1. Core Principles and Motivation

Data-driven ROM architectures are motivated by the need to simulate, predict, and control high-dimensional nonlinear systems—such as fluid flows, multiphysics phenomena, or structural mechanics—with computational cost orders of magnitude lower than full-order models (FOMs). Classical ROMs, usually built by Galerkin projection onto Proper Orthogonal Decomposition (POD) modes, often become unstable or inaccurate in the presence of strong nonlinearity or when important physical effects are not well-resolved in the reduced basis. Data-driven alternatives address these limitations by:

• Inferring the temporal evolution laws or closure terms directly from data, rather than exclusively from truncated governing equations.
• Enabling stable, accurate long-term prediction and robust handling of nonlinearities without access to intrusive full-model operators.
• Allowing non-intrusive application to legacy solvers and black-box codes, since only snapshot data are required.
• Providing flexibility for enhancing models with neural networks, regularization strategies, or embedding integration schemes for stability (Xie et al., 2018, Mohebujjaman et al., 2018, Pawar et al., 2019).

2. Canonical Algorithmic Frameworks

Several computational motifs define modern data-driven ROM architectures:

• POD-based compression: From a matrix $U$ of snapshots $u(t_j) \in \mathbb{R}^{D_h}$, the leading $r$ eigenmodes $\{\varphi_1, \ldots, \varphi_r\}$ are computed, yielding a reduced representation $u(t) \approx \Phi_r a(t)$, with $\Phi_r \in \mathbb{R}^{D_h \times r}$ (Xie et al., 2018).
• Learning reduced dynamics: Instead of direct Galerkin projection, the time evolution $a_{t+1}$ or its derivative $\dot{a}$ in reduced space is inferred via a neural network, regression, or another data-driven map: $\dot{a} = f_r(a)$, often trained to reproduce snapshot trajectories with high fidelity (Xie et al., 2018, Pawar et al., 2019).
• Embedding integration schemes: To ensure stability, the learned dynamics are structured in the manner of implicit linear multistep time integrators (e.g., Adams–Moulton) or via residual learning with explicit time-discretization formulas (Xie et al., 2018, Pawar et al., 2019).
• Supervised learning objectives: Networks are fit by minimizing mean-squared errors in either rollout predictions, local truncation residuals, or closures against filtered FOM data. For example, for a $K$-step implicit scheme,

$u(t_j) \in \mathbb{R}^{D_h}$0

and MSE is minimized over all time steps (Xie et al., 2018).

Technique	Compression	ROM Dynamics	Stability Mechanism
POD–Galerkin	POD/SVD	Projected ODE	None (often unstable)
LMNet-ROM (Xie et al., 2018)	POD	NN in coeff. space	Implicit LM multistep
DNN/NIROM (Pawar et al., 2019)	POD	Feedforward NN	Residuals/time-discret.
DDF-ROM (Xie et al., 2017Mohebujjaman et al., 2018)	POD	Quadratic closure	Physical constraints
CDDF-ROM (Mohebujjaman et al., 2018)	POD	Quadratic closure	Dissipativity, energy

3. Stability, Accuracy, and Closure Strategies

Embedding Time Discretizations

Implicit integration schemes, such as the Adams–Moulton family, encoded in the learning objective, confer A-stability properties to the reduced model (Xie et al., 2018). This is a key distinction from both "black box" time series models and explicit projection ROMs, which may suffer from secular energy drift, amplified instability, or blow-up.

Data-driven Closure and Filtering

In projection-based ROMs, truncation induces energy transfer between resolved and unresolved modes, manifesting as closure problems. Data-driven filtering frameworks first apply spatial filters, yielding non-closed reduced models; explicit quadratic or higher-order surrogate closure maps are then inferred by regression against FOM snapshot-derived true stresses. Imposing physical constraints such as dissipativity (linear part negative semi-definite) and nonlinear energy conservation yields ROMs (e.g., CDDF-ROM) that are stable and accurate even in severely data-limited regimes (Mohebujjaman et al., 2018).

Empirical Performance

For benchmark fluid flows (e.g., 2D channel flow past a cylinder at $u(t_j) \in \mathbb{R}^{D_h}$1), data-driven architectures (LMNet-ROM) achieve average $u(t_j) \in \mathbb{R}^{D_h}$2 errors roughly 2–3x lower than classical POD–Galerkin ROMs:

• For $u(t_j) \in \mathbb{R}^{D_h}$3, LMNet-ROM: $u(t_j) \in \mathbb{R}^{D_h}$4; GP-ROM: $u(t_j) \in \mathbb{R}^{D_h}$5 (Xie et al., 2018).

Correct long-horizon vortex street evolution, energy, and drag are maintained, whereas standard projection ROMs exhibit pronounced drift. Similar improvements are seen for physically-constrained closure models (CDDF-ROM) (Mohebujjaman et al., 2018).

4. Implementation Pathways and Non-intrusive Learning

A defining strength of modern data-driven ROM architectures is that they require neither the explicit FOM operators nor modifications to legacy codes; only state snapshots are needed from simulation or experiment. This non-intrusive property supports:

• Choice of basis: While most reported work employs POD, other encoders such as autoencoders or dynamic mode decomposition can be substituted.
• Training process: Modal coefficients for training are extracted from FOM snapshots using the reduced basis. Feedforward or residual neural networks with moderate depth and width (e.g., one hidden layer, 128 neurons) suffice for reduced dimensions $u(t_j) \in \mathbb{R}^{D_h}$6; larger/deeper networks may require regularization to avoid overfitting (Xie et al., 2018).
• Online deployment: The learned ODE $u(t_j) \in \mathbb{R}^{D_h}$7 (or its discrete analog) may be integrated forward with arbitrary solvers (implicit or explicit). No operator assembly or lifting/projection to full space is needed until field reconstruction is required.

Examples of online speed-up and computational savings are striking:

• DNS: $u(t_j) \in \mathbb{R}^{D_h}$8 s vs. LMNet-ROM: $u(t_j) \in \mathbb{R}^{D_h}$9 s (speed-up $r$0) (Xie et al., 2018).

5. Generalizability, Hyperparameters, and Extensions

Data-driven ROMs are system-agnostic and extendable:

• Applicability: Any system (ODE/PDE) with observable state snapshots can, in principle, be reduced—elasticity, reactive flows, bio-transport, among others (Xie et al., 2018).
• Choice of multistep order $r$1: Higher $r$2 improves stability and accuracy but increases the training cost; $r$3 (backward Euler) or $r$4 (trapezoidal) often achieves practical balance (Xie et al., 2018).
• Network size tuning: For $r$5, a one-layer, 128-neuron architecture suffices; excessive width or depth risks over-fitting (Xie et al., 2018).
• Extensions:
  • Parametric dependence: Augment $r$6 with system parameters as inputs.
  • Physics-informed regularization: Enforce sparsity or physical constraints to improve robustness to noisy data.
  • Alternative time discretizations: Backward differentiation formulas or custom schemes for stiff or multiscale dynamics (Xie et al., 2018).

6. Comparative Context: Related Approaches and Distinctions

Data-driven ROM architectures are distinguished from both intrusive projection-ROMs and black-box time-series models by their hybrid design: they combine optimal projection-based compression with a data-driven learning layer for system evolution in reduced space. In contrast to purely operator-based approaches, they can:

• Systematically address closure and instability issues inherent to the projection ROM, especially under strong nonlinearity.
• Integrate physical constraints (e.g., energy, dissipativity) by design or learned regularization (Mohebujjaman et al., 2018).
• Bypass the need for full model information and derivatives, enabling generalization across models, geometries, and even experimental settings.

The modular structure, with clear separation of encoder, evolution, and decoder, positions data-driven ROMs as extensible tools for both intrusive and non-intrusive reduced modeling of nonlinear PDEs.

7. Open Challenges and Future Directions

Current research highlights several challenges and frontiers:

• Noise robustness: Noise on snapshots degrades accuracy, especially above $r$7 noise; robust/regularized architectures are needed (Xie et al., 2018).
• Hyperparameter sensitivity: Choices for time-stepping order, network size, and regularization parameters can dictate performance; automated model selection remains an open problem.
• Scalability to high-dimensional parameter spaces: Efficient sampling, embedding of parametric dependence, and transfer learning across parameter regimes are active areas for development.
• Hybrid architectures: Integration of data-driven and physics-informed layers, as well as deployment of alternative encoders (e.g., autoencoders, dynamic mode structures), are expected to remain central topics.
• Extending physical constraint enforcement: Embedding principles of energy conservation and dissipativity across architectural types, including deep network-based models, remains a promising direction for ensuring stable reduced models in increasingly complex domains.

In summary, data-driven reduced-order modeling architectures offer a systematic and flexible framework for constructing fast, stable, and accurate surrogate models of nonlinear dynamical systems directly from data. By embedding stability mechanisms, closure corrections, and leveraging non-intrusive training, they outperform classical projection-based ROMs in long-term accuracy, computational savings, and applicability to turbulent, multiscale, and black-box scenarios (Xie et al., 2018, Mohebujjaman et al., 2018, Pawar et al., 2019).