Reduced-Order Neural Operator Modeling (RONOM)

• RONOM is a hybrid paradigm that combines projection-based ROM and neural operator learning to efficiently simulate parametrized and time-dependent PDE systems.
• It integrates encoder-based dimensional reduction, latent dynamics evolution, and kernel-decoded reconstruction to achieve robustness, discretization invariance, and online efficiency.
• Applications in additive manufacturing, structural mechanics, and fluid dynamics demonstrate significant speedup and accuracy improvements over traditional methods.

Reduced-Order Neural Operator Modeling (RONOM) denotes a family of approaches that combine reduced-order modeling with operator learning to construct low-dimensional surrogates for parametrized, time-dependent PDEs and related dynamical systems. In the most explicit formulation, RONOM lifts a discretized input function into a reduced representation, evolves a latent dynamical system, and decodes back to a function, thereby coupling ROM-style projection with neural-operator flexibility (Dummer et al., 17 Jul 2025). Closely related usages of the term include learning reduced latent vector fields by operator inference (Parish et al., 9 Mar 2026), converting unequal-domain spatio-temporal maps into same-domain coefficient-function maps before operator learning (Meng et al., 2024), and coupling continuous neural-field bases to learned or reduced dynamics (Chang et al., 2023). Taken together, these works suggest that RONOM is best understood as an umbrella paradigm for mesh-flexible, many-query surrogate modeling rather than as a single architecture.

1. Conceptual scope and relation to neighboring paradigms

RONOM occupies an intermediate position between classical projection-based ROM and full neural operators. Classical ROM typically constructs a basis and then evolves reduced coordinates on a fixed discretization, whereas full neural operators such as FNO and DeepONet learn function-space mappings more directly. RONOM retains a reduced representation, but it replaces or augments intrusive reduced dynamics, closure terms, or reconstruction maps with learned operators, often with the explicit goal of improving discretization robustness, parametric generalization, or online speed (Dummer et al., 17 Jul 2025).

A second axis of variation concerns what is being learned. In projection-based latent formulations, the learned object is the reduced vector field $f(a,\mu)$ governing latent dynamics after POD or a related basis construction (Parish et al., 9 Mar 2026). In reduced function-space formulations, the learned object is a same-domain operator obtained after compressing an “extra” spatial or temporal dimension through variable separation (Meng et al., 2024). In semi-intrusive hyper-reduction formulations, the learned object is neither the full solution operator nor the latent ODE alone, but rather reduced residuals, Jacobians, or closure terms embedded inside Galerkin or Newton solvers (Cicci et al., 2022).

This breadth also explains an important terminological point. A common misconception is to equate RONOM with an FNO surrogate trained on reduced data. The literature instead includes structured operator inference, DeepONet- and MIONet-based closure models, kernel-decoded latent ODEs, continuous neural-field bases, and sparse graph operators for Lagrangian systems (Ivagnes et al., 22 May 2025).

2. Mathematical formulations

A canonical projection-based starting point is the semi-discrete dynamical system

$$\dot{x}(t;\mu)=g\big(x(t;\mu),\mu\big), \qquad x\in\mathbb{R}^N,$$

with reduced coordinates $a(t;\mu)=V^\top x(t;\mu)$ for a basis $V\in\mathbb{R}^{N\times K}$, $K\ll N$. Non-intrusive reduced modeling then replaces intrusive Galerkin evolution by a learned latent operator

$\dot{a}(t;\mu)=f\big(a(t;\mu),\mu;\Theta\big),$

which is the core mathematical template in parametric OpInf and NN-OpInf (McQuarrie et al., 2021). In this setting, the operator may be polynomial, neural, or a structured sum of multiple terms.

The general RONOM framework of 2025 casts the approximation itself as an operator factorization,

$$\mathcal{K}^\dagger(u_0,t)=\big(D\circ\mathcal{F}_t\circ E\big)(u_0),$$

where $E$ is an encoder based on regularized projection into a reduced space, $\mathcal{F}_t$ is a latent flow induced by a neural ODE, and $D$ is a kernel-based decoder that reconstructs a function at arbitrary spatial query points (Dummer et al., 17 Jul 2025). This formulation makes reduced modeling and operator learning coexist in a single end-to-end map.

Other RONOM variants alter the reduction stage itself. RO-NORM uses variable separation to rewrite unequal-domain maps such as $$\dot{x}(t;\mu)=g\big(x(t;\mu),\mu\big), \qquad x\in\mathbb{R}^N,$$0 or $$\dot{x}(t;\mu)=g\big(x(t;\mu),\mu\big), \qquad x\in\mathbb{R}^N,$$1 as same-domain coefficient-function mappings. Its two central separated forms are

$$\dot{x}(t;\mu)=g\big(x(t;\mu),\mu\big), \qquad x\in\mathbb{R}^N,$$2

after which a same-domain neural operator on a manifold is trained on the coefficient functions rather than directly on the full unequal-domain fields (Meng et al., 2024).

Continuous-basis formulations use a related but distinct ansatz. LiCROM represents deformations by

$$\dot{x}(t;\mu)=g\big(x(t;\mu),\mu\big), \qquad x\in\mathbb{R}^N,$$3

where $$\dot{x}(t;\mu)=g\big(x(t;\mu),\mu\big), \qquad x\in\mathbb{R}^N,$$4 is a continuous neural field evaluated at arbitrary points of the reference domain and $$\dot{x}(t;\mu)=g\big(x(t;\mu),\mu\big), \qquad x\in\mathbb{R}^N,$$5 contains reduced coordinates (Chang et al., 2023). This separates spatial discretization from latent dynamics and makes the basis itself discretization-independent.

RONOM also includes additive correction models. In the PPE-ROM closure setting for parametrized incompressible flow, the reduced equations are augmented as

$$\dot{x}(t;\mu)=g\big(x(t;\mu),\mu\big), \qquad x\in\mathbb{R}^N,$$6

with the closure $$\dot{x}(t;\mu)=g\big(x(t;\mu),\mu\big), \qquad x\in\mathbb{R}^N,$$7 learned by a modified MIONet and the eddy-viscosity coefficients $$\dot{x}(t;\mu)=g\big(x(t;\mu),\mu\big), \qquad x\in\mathbb{R}^N,$$8 learned by a modified DeepONet (Ivagnes et al., 22 May 2025). In the continuous-memory CD-ROM formulation, the reduced dynamics become

$$\dot{x}(t;\mu)=g\big(x(t;\mu),\mu\big), \qquad x\in\mathbb{R}^N,$$9

so that unresolved effects enter through a learned non-Markovian memory state rather than an instantaneous correction (Menier et al., 2022).

3. Architectural families

The current literature contains several recurrent architectural patterns.

Formulation	Core learned object	Representative papers
Projection-based latent dynamics	Reduced vector field $a(t;\mu)=V^\top x(t;\mu)$0 in latent coordinates	(McQuarrie et al., 2021, Parish et al., 9 Mar 2026)
Reduced function-space operator	Same-domain operator after variable separation or projection	(Meng et al., 2024, Dummer et al., 17 Jul 2025)
Physics-preserving reduced solver surrogate	Reduced residuals, Jacobians, or closure terms inside Galerkin/Newton loops	(Cicci et al., 2022, Ivagnes et al., 22 May 2025, Menier et al., 2022)
Continuous or sparse reduced representation	Neural-field basis, graph sparse stepper, or nonlinear resolution-invariant basis	(Chang et al., 2023, Viswanath et al., 2024, Mou et al., 17 Feb 2026)

Within latent-dynamics RONOM, NN-OpInf is notable for enforcing local operator structure by construction. It uses additive compositions of heterogeneous operator blocks, including skew-symmetric energy-preserving terms, SPSD dissipative terms, gradient operators derived from learned potentials, and unstructured MLP residuals. Training uses hard parameterizations for structure, hybrid L-BFGS and ADAM optimization, max-abs normalization, and small ensembles, with online complexity comparable to quadratic P-OpInf for structured matrix-output operators (Parish et al., 9 Mar 2026).

Within semi-intrusive RONOM, Deep-HyROMnet learns reduced residual vectors and Jacobian matrices after Galerkin projection. The reduced Newton step is preserved, but the expensive nonlinear reduced operators are replaced by DFNN-plus-CAE surrogates, so online evaluation becomes independent of the full-order dimension $a(t;\mu)=V^\top x(t;\mu)$1 (Cicci et al., 2022). This is conceptually distinct from end-to-end solution regression: the solver remains a reduced PDE solve.

Within basis-learning RONOM, Neural-POD replaces SVD-based linear POD with sequential residual minimization using neural networks to construct nonlinear, orthogonal, resolution-invariant basis functions in infinite-dimensional spaces. The basis functions can then be used either in Galerkin ROM or as a “plug-and-play” trunk representation in operator learning, including DeepONet-style settings (Mou et al., 17 Feb 2026).

Sparse Lagrangian RONOM introduces yet another pattern. GIOROM evolves a sparse graph using an IO–GINO–IO operator stack, then recovers reduced coefficients by least squares in a learned continuous basis and reconstructs full states at arbitrary query points. In that design, the reduced state is implicit in the sparse graph dynamics and the basis-based reconstruction rather than in an explicitly integrated latent ODE (Viswanath et al., 2024).

4. Discretization invariance and error control

Discretization invariance is a recurring claim in RONOM, but the mechanisms differ. In the general RONOM framework, mesh flexibility is achieved by a regularized projection-based lifting and a kernel decoder, and the resulting discretization error is bounded by

$a(t;\mu)=V^\top x(t;\mu)$2

where $a(t;\mu)=V^\top x(t;\mu)$3 is the projection-discretization coefficient error and $a(t;\mu)=V^\top x(t;\mu)$4 is the latent ODE time step (Dummer et al., 17 Jul 2025). This is significant because it provides a ROM-style decomposition of discretization error inside a neural-operator pipeline.

RO-NORM addresses discretization and unequal-domain issues through basis-based dimensional compression before operator learning. Its same-domain approximator, NORM, is discretization-invariant on manifolds, while the unequal-domain encoder-decoder uses precomputed spatial or temporal bases. The paper reports that separate space/time reductions exhibit faster singular value decay than overall reductions, and that online reconstruction yields consistently smaller errors than offline reconstruction in increase-domain tasks (Meng et al., 2024). This indicates that reduction is not merely a preprocessing convenience; it changes approximation behavior.

Continuous-basis methods pursue invariance through basis evaluation rather than explicit projection theory. LiCROM uses an implicit neural field $a(t;\mu)=V^\top x(t;\mu)$5 that is evaluated at arbitrary points, allowing training across different mesh resolutions, connectivity, and element types, as well as runtime mesh swapping, cutting, and hole punching (Chang et al., 2023). GIOROM achieves related behavior through radius-graph integral operators and a neural-field decoder; it is explicitly described as discretization invariant and reconstructs high-resolution states from sparse graphs (Viswanath et al., 2024).

Taken together, these results indicate that discretization invariance in RONOM is typically not “discretization-free.” It is usually obtained through an explicit lifting, basis evaluation, graph integral approximation, or kernel reconstruction stage. This also clarifies a frequent misunderstanding: invariance to resolution does not by itself guarantee robustness to large changes in geometry, boundary conditions, or physics. The AM and RO-NORM studies both note that substantial shifts in geometry, boundary conditions, or physical regime can require retraining or additional conditioning signals (Yaseen et al., 2023).

5. Applications and empirical evidence

A prominent application is additive manufacturing in MOOSE-based direct energy deposition. In the 2023 study centered on a thermo-mechanical DED model, an FNO ROM was trained on 470 usable simulations after discarding cases with no melt pool formation, with an 80%/10%/10% train/validation/test split. On 47 held-out test samples, bead-volume prediction achieved RMSE $a(t;\mu)=V^\top x(t;\mu)$6 and $a(t;\mu)=V^\top x(t;\mu)$7 for FNO versus RMSE $a(t;\mu)=V^\top x(t;\mu)$8 and $a(t;\mu)=V^\top x(t;\mu)$9 for a DNN baseline; maximum melt pool temperature achieved RMSE $V\in\mathbb{R}^{N\times K}$0 and $V\in\mathbb{R}^{N\times K}$1 for FNO versus RMSE $V\in\mathbb{R}^{N\times K}$2 and $V\in\mathbb{R}^{N\times K}$3 for DNN. The FNO used 4 Fourier layers, 5 retained modes per layer, GeLU activations, and 50 modes for 200-step time-series prediction; training time was 276.85 s versus 69.32 s for the DNN, while both surrogates were essentially instantaneous at inference (Yaseen et al., 2023).

RO-NORM provides a broader benchmark suite for unequal-domain predictive learning on complex domains. Across six tasks, it outperformed DeepONet, POD-DeepONet, PCA-Net, and vanilla NORM on most reported metrics. In 2D Burgers’ equations it reached relative $V\in\mathbb{R}^{N\times K}$4 error $V\in\mathbb{R}^{N\times K}$5 versus $V\in\mathbb{R}^{N\times K}$6 for NORM and $V\in\mathbb{R}^{N\times K}$7 for DeepONet; in heat source layout $V\in\mathbb{R}^{N\times K}$8 temperature-field prediction it achieved $V\in\mathbb{R}^{N\times K}$9 versus $K\ll N$0 for NORM and $K\ll N$1 for DeepONet. In addition, the paper reports a 5.6$K\ll N$2 faster training time in a decrease-domain wave case and a 23$K\ll N$3 faster training time, roughly 10$K\ll N$4 lower relative $K\ll N$5 error, and about 20$K\ll N$6 fewer parameters in an increase-domain heat source layout case (Meng et al., 2024).

For latent-dynamics RONOM, NN-OpInf demonstrates the value of structure preservation across several nonlinear and parametric systems. In 1D Burgers’ equation, the skew-symmetric model attains about $K\ll N$7 reproductive error and preserves energy in future-state prediction. In 2D nonlinear convection–diffusion–reaction and nonlinear heat conduction, structured NN-OpInf is reported as 5–10$K\ll N$8 more accurate than polynomial OpInf baselines; in the premixed H$K\ll N$9–air flame problem it attains about 3$\dot{a}(t;\mu)=f\big(a(t;\mu),\mu;\Theta\big),$0 lower test errors at the highest reduced dimension under non-affine parametric dependence; and in 3D hyper-elastic torsion the SPSD-potential model is approximately 10$\dot{a}(t;\mu)=f\big(a(t;\mu),\mu;\Theta\big),$1 more accurate than P-OpInf (Parish et al., 9 Mar 2026).

Semi-intrusive RONOM has been especially effective in nonlinear structural and cardiac mechanics. Deep-HyROMnet learns reduced residuals and Jacobians so that the online Newton solve remains reduced and physics-based. For the clamped rectangular beam, reported online timings include 0.026 s with speedup $\dot{a}(t;\mu)=f\big(a(t;\mu),\mu;\Theta\big),$2 and $\dot{a}(t;\mu)=f\big(a(t;\mu),\mu;\Theta\big),$3 in one test configuration. For the idealized left ventricle with $\dot{a}(t;\mu)=f\big(a(t;\mu),\mu;\Theta\big),$4 and reduced dimension $\dot{a}(t;\mu)=f\big(a(t;\mu),\mu;\Theta\big),$5, the reported online time is 0.1 s with speedup $\dot{a}(t;\mu)=f\big(a(t;\mu),\mu;\Theta\big),$6 and $\dot{a}(t;\mu)=f\big(a(t;\mu),\mu;\Theta\big),$7; on a refined mesh with $\dot{a}(t;\mu)=f\big(a(t;\mu),\mu;\Theta\big),$8, online time is about 0.2 s with speedup $\dot{a}(t;\mu)=f\big(a(t;\mu),\mu;\Theta\big),$9 and $$\mathcal{K}^\dagger(u_0,t)=\big(D\circ\mathcal{F}_t\circ E\big)(u_0),$$0 (Cicci et al., 2022).

Sparse-graph Lagrangian RONOM extends the paradigm beyond fixed-grid PDE surrogates. GIOROM reports a 6.6$$\mathcal{K}^\dagger(u_0,t)=\big(D\circ\mathcal{F}_t\circ E\big)(u_0),$$1–32$$\mathcal{K}^\dagger(u_0,t)=\big(D\circ\mathcal{F}_t\circ E\big)(u_0),$$2 reduction in input dimensionality, reconstruction from sparse graphs to full-order states with negligible extra cost, and lower rollout losses than GNS, GINO, and GAT in most listed benchmarks. The paper further reports that, when sparsity is exploited, the overall speedup versus GNS is about $$\mathcal{K}^\dagger(u_0,t)=\big(D\circ\mathcal{F}_t\circ E\big)(u_0),$$3 with comparable or better accuracy in one 7056-DoF setting simulated using 1776 nodes (Viswanath et al., 2024).

6. Limitations, misunderstandings, and research directions

The first limitation is definitional. RONOM is not a single model class, and the term can refer to at least four distinct practices: learning reduced latent dynamics, learning reduced residuals or Jacobians inside intrusive solvers, learning closure operators for truncated modes, and reducing unequal-domain operator-learning problems before applying a neural operator. Taken together, the literature suggests that the unifying principle is not architectural uniformity, but the placement of operator learning inside a reduced representation.

The second limitation is that reduced-order learning does not automatically imply full-field or multiphysics fidelity. The MOOSE-based AM study trains on bead volume and maximum melt pool temperature rather than on full fields $$\mathcal{K}^\dagger(u_0,t)=\big(D\circ\mathcal{F}_t\circ E\big)(u_0),$$4, $$\mathcal{K}^\dagger(u_0,t)=\big(D\circ\mathcal{F}_t\circ E\big)(u_0),$$5, or $$\mathcal{K}^\dagger(u_0,t)=\big(D\circ\mathcal{F}_t\circ E\big)(u_0),$$6, and explicitly identifies field-level thermo-mechanical prediction as future work (Yaseen et al., 2023). Deep-HyROMnet preserves Galerkin structure, but its accuracy remains tied to POD basis quality and to the representativeness of reduced operator training data (Cicci et al., 2022).

A third limitation is optimization and identifiability. NN-OpInf is non-convex, more expensive to train than polynomial OpInf, sensitive to optimizer choices, and dependent on accurate derivative data or derivative estimates (Parish et al., 9 Mar 2026). Neural-POD inherits the higher offline cost of sequential neural basis training, as well as sensitivity to loss weights, orthogonality enforcement, and data coverage, even though it gains nonlinear expressivity and resolution invariance (Mou et al., 17 Feb 2026). More generally, discretization invariance does not eliminate domain shift: significant changes in geometry, boundary conditions, or physics can still induce failure modes (Dummer et al., 17 Jul 2025).

Current research directions are correspondingly diverse. Structure-preserving latent dynamics, including richer operator libraries such as port-Hamiltonian or metriplectic forms, are an explicit direction in NN-OpInf (Parish et al., 9 Mar 2026). RO-NORM points toward adaptive or learned bases and scalable eigensolvers on complex manifolds (Meng et al., 2024). Neural-POD suggests nonlinear orthogonal bases that can be optimized in norms other than $$\mathcal{K}^\dagger(u_0,t)=\big(D\circ\mathcal{F}_t\circ E\big)(u_0),$$7 and reused in both ROM and operator learning (Mou et al., 17 Feb 2026). The general RONOM framework identifies physics-informed training, stronger handling of temporal super-resolution, and extensions beyond initial-condition input spaces as natural next steps (Dummer et al., 17 Jul 2025). In AM, the listed extensions include full thermo-mechanical outputs, full 3D FNOs, richer scan paths, multi-material deposition, melt-pool dynamics, phase change, and uncertainty quantification (Yaseen et al., 2023).

These directions suggest that the central research problem in RONOM is no longer merely dimensionality reduction. It is the design of reduced operators that remain stable, interpretable, and discretization-robust while preserving enough physical structure to support long-horizon prediction, control, optimization, and many-query scientific computing.