Local Neural Operators in PDE Analysis

• Local neural operators are formulations that impose locality via patchwise kernels and short-time evolution rules to capture fine-scale PDE dynamics.
• They combine global spectral methods with local differential and integral operators, enabling resolution invariance and efficient solution of transient PDEs.
• Empirical studies show that tuning the local receptive range optimizes accuracy and speed across varied domains, enhancing simulation and control performance.

Searching arXiv for recent and foundational papers on local neural operators to ground the encyclopedia entry. Local neural operators are neural-operator formulations that impose locality on mappings between function spaces, typically by learning short-time evolution rules, locally supported kernels, patchwise operators, or local residual and flux updates instead of a single global input–output map. In the literature, the term covers several related but distinct ideas: local-in-space time-marching operators for transient PDEs on varied domains, localized differential and integral operator layers that preserve resolution transfer, local-in-time surrogates for equation-free computation, and architectures that augment globally spectral models with explicit local branches or local feature extractors (Li et al., 2022, Liu-Schiaffini et al., 2024, Fabiani et al., 5 May 2025). Across these variants, the common objective is to align operator learning with the finite dependence structure, short-range interactions, and multiscale organization of PDE dynamics, while retaining the discretization-agnostic advantages associated with neural operators.

1. Conceptual scope and historical formulations

A central early formulation treats the target of learning not as a domain-specific solver but as a local time-marching operator

$\mathcal{G}:u_t(x) \mapsto u_{t+\Delta t}(x),$

with locality expressed by

$$\frac{\partial u_{t+\Delta t}(x_a)}{\partial u_t(x_b)}=0,\qquad \forall \|x_a-x_b\|>r,$$

for some finite interaction radius $r$. On this view, a local neural operator learns a shift-invariant rule that can be reused across spatial locations and then combined with separate boundary treatment, which is why it was proposed for solving transient PDEs on varied domains rather than only on one fixed geometry (Li et al., 2022).

Subsequent work broadened the meaning of locality. In equation-free analysis, a “local neural operator” may be local in time, local in space, or local in space-time. The local-in-time case learns a short-time solution operator $\mathcal{S}_{\Delta t}$ rather than a long-horizon map; the local-in-space case is a “Gap-Tooth” operator on small spatial subdomains; and the local-in-space-time case is a patch operator that restricts both spatial and temporal support (Fabiani et al., 5 May 2025). In additive manufacturing, locality is introduced through physically defined heat-affected windows of size $11\times11\times11$, although the core operator inside each window remains an FNO with global Fourier convolution on that restricted subdomain (Chen et al., 2023). This distinction is important: “local” may refer either to the learned kernel itself or to the domain on which an otherwise global operator is applied.

The broader operator-learning literature also clarifies what locality cannot mean. Purely pointwise nonlinear maps are not universal for operator learning; a general operator architecture must be both nonlinear and nonlocal. Theoretical analysis of the averaging neural operator shows that even a minimal nonlocal ingredient—a spatial average, equivalent in periodic settings to the zeroth Fourier mode—already suffices for universal approximation on compact subsets of $C^s$ or $W^{s,p}$ spaces (Lanthaler et al., 2023). This places local neural operators within a larger design tension: locality is useful, but complete elimination of nonlocal coupling is not, in general, compatible with universal operator approximation.

2. Architectural realizations of locality

The original Local Neural Operator architecture is organized as a lifting layer, a stack of $n$ blocks, and a projection layer. Each block has two parallel paths: a physical path consisting of convolutional layers in physical space, and a spectral path in which a local subdomain is transformed by a Legendre transform, processed by a learnable linear map, and mapped back by the inverse transform. With GELU activations, the full operator is written as

$$\mathcal{G}_\theta = \mathcal{P}_\textup{projection}\circ\mathcal{B}_n\circ\cdots\circ\mathcal{B}_1\circ\mathcal{P}_\textup{lifting},$$

and locality is enforced through bounded physical stencils and localized spectral windows rather than global Fourier modes (Ye et al., 2023).

A second line of work constructs local neural-operator layers directly at the operator level. “Neural Operators with Localized Integral and Differential Kernels” introduces localized differential operators derived from CNN-like kernels under the correct scaling as $h\to0$, together with localized integral operators based on discrete-continuous convolutions (DISCO). The differential branch gives continuous-limit justification for stencil-like layers, while the DISCO branch preserves operator-learning properties and resolution transfer by evaluating a continuous kernel basis across discretizations and, on the sphere, via group convolution (Liu-Schiaffini et al., 2024).

A third formulation learns local numerical updates rather than solution snapshots. In ideal MHD, the modified Flux Fourier Neural Operator replaces hand-designed numerical fluxes with learned flux functions over shifted local stencils. The method uses separate FNO models for density, velocity components, magnetic-field components, and energy, and advances the conservative state through residual updates such as

$$\frac{\partial u_{t+\Delta t}(x_a)}{\partial u_t(x_b)}=0,\qquad \forall \|x_a-x_b\|>r,$$0

in one dimension, or with both $$\frac{\partial u_{t+\Delta t}(x_a)}{\partial u_t(x_b)}=0,\qquad \forall \|x_a-x_b\|>r,$$1- and $$\frac{\partial u_{t+\Delta t}(x_a)}{\partial u_t(x_b)}=0,\qquad \forall \|x_a-x_b\|>r,$$2-direction fluxes in two dimensions (Kim et al., 2024). Here locality is inherited from the numerical-flux formulation itself.

Hybrid designs retain a global operator backbone but add explicit local modules. Conv-FNO introduces a CNN pre-extractor whose local spatial features are concatenated with the original input before FNO processing, while two resizing schemes are proposed so that the added CNN does not destroy resolution invariance (Liu et al., 22 Mar 2025). LoGlo-FNO instead inserts a parallel local spectral convolution branch defined on patches $$\frac{\partial u_{t+\Delta t}(x_a)}{\partial u_t(x_b)}=0,\qquad \forall \|x_a-x_b\|>r,$$3, where

$$\frac{\partial u_{t+\Delta t}(x_a)}{\partial u_t(x_b)}=0,\qquad \forall \|x_a-x_b\|>r,$$4

and supplements it with a high-frequency propagation branch based on pooled-and-upsampled residuals (Kalimuthu et al., 5 Apr 2025). These models do not reject global structure; they repartition it into local and global channels.

Graph-based surrogates introduce yet another locality mechanism. In graph neural simulators for time-dependent PDEs, nodes represent spatial points and edges represent local neighborhoods, with message passing restricted to $$\frac{\partial u_{t+\Delta t}(x_a)}{\partial u_t(x_b)}=0,\qquad \forall \|x_a-x_b\|>r,$$5. The learned object is the instantaneous derivative

$$\frac{\partial u_{t+\Delta t}(x_a)}{\partial u_t(x_b)}=0,\qquad \forall \|x_a-x_b\|>r,$$6

which is then integrated by explicit Euler,

$$\frac{\partial u_{t+\Delta t}(x_a)}{\partial u_t(x_b)}=0,\qquad \forall \|x_a-x_b\|>r,$$7

Although this work does not use the term “local neural operator” explicitly, it is local in space through neighbor-only message passing and local in time through one-step derivative prediction (Nayak et al., 7 Sep 2025).

3. Mathematical notions of locality

The most explicit formalization of locality in LNOs is through receptive field and receptive range. For a one-dimensional simplification, the relevance from input point $$\frac{\partial u_{t+\Delta t}(x_a)}{\partial u_t(x_b)}=0,\qquad \forall \|x_a-x_b\|>r,$$8 to output point $$\frac{\partial u_{t+\Delta t}(x_a)}{\partial u_t(x_b)}=0,\qquad \forall \|x_a-x_b\|>r,$$9 is defined as

$r$0

and, statistically at random initialization,

$r$1

From this, the paper defines the normalized receptive-field profile $r$2, the maximum receptive range (MRR) $r$3, and the effective receptive range (ERR)

$r$4

For multi-block architectures, the MRR grows with depth and spectral width, and the analysis shows that $r$5 affects the receptive field more strongly than $r$6 or $r$7 (Ye et al., 2023).

The main empirical theorem-like conclusion from this analysis is that an initial receptive range compatible with the learning task is crucial. If the receptive range is too small, LNOs exhibit cutoff, numerical oscillation, and recurrent-rollout failure; if it is too large, accuracy degrades because capacity is spent on irrelevant long-range interactions. The reported error curves are U-shaped as a function of initial ERR, and the compatible $r$8 shifts with Reynolds number, Mach number, and time step because the physical dependence range changes with those parameters (Ye et al., 2023). This suggests that locality is not a static architectural virtue but a task-dependent quantity that must be matched to the PDE’s effective domain of dependence.

A more abstract approximation-theoretic treatment reaches a related conclusion from a different direction. “Local approximation of operators” reduces nonlinear operator approximation to approximation on $r$9 using spherical polynomials and localized kernels. It proves constructive estimates whose constants grow only like $\mathcal{S}_{\Delta t}$0, and it shows that local approximation at $\mathcal{S}_{\Delta t}$1 can be obtained using only samples in a neighborhood $\mathcal{S}_{\Delta t}$2 because the kernel suppresses far-away contributions exponentially (Mhaskar, 2022). In this setting, locality is smoothness-adaptive and neighborhood-based, not tied to any specific neural architecture.

There is also a phase-space notion of operator locality. For operators with bandlimited Kohn–Nirenberg symbols, localized test functions and local subsets of Gabor coefficients yield local approximations of the operator, whereas exact identification of the full operator class requires a globally non-decaying identifier. The result is a rigorous statement that exact global identifiability and local probing are different objectives (Krahmer et al., 2012). In a modern reading, this supports a common design trade-off in local neural operators: practical local approximation is often preferable to exact globally coupled reconstruction.

4. Relation to global neural operators

Local neural operators are frequently introduced as a response to limitations of globally spectral or globally parameterized neural operators. FNOs perform global convolutions in Fourier space, which is effective for long-range interactions but can over-smooth local detail; plain CNNs capture local features efficiently but are tied to fixed discretizations. Localized differential and integral kernels were proposed precisely to bridge this gap while keeping operator-learning properties such as prediction at arbitrary resolution (Liu-Schiaffini et al., 2024).

Several papers frame the problem as one of local–global complementarity rather than replacement. Conv-FNO argues that standard FNOs are strong at global modeling but weak at extracting local spatial features from raw inputs, and therefore augments them with CNN-based local feature extraction (Liu et al., 22 Mar 2025). LoGlo-FNO makes the same point in spectral terms: standard FNOs often miss local, high-frequency, and non-dominant spectral components because of spectral bias and explicit high-frequency truncation, so it couples a global Fourier branch to a local patchwise branch that retains all Fourier modes within each patch, plus a high-frequency propagation module and a frequency-sensitive loss (Kalimuthu et al., 5 Apr 2025).

A related but distinct argument is made by adaptive-coordinate methods. ACT attributes part of the apparent nonlocality of neural-operator mappings to fixed Eulerian coordinates: moving waves, vortices, fronts, and interfaces are spatially misaligned in static coordinates, which forces the backbone to connect distant points. By learning residual coordinate transforms

$\mathcal{S}_{\Delta t}$3

with bounded offsets and differentiable resampling, ACT aims to make the operator effectively more local in the transformed representation (Liu et al., 7 May 2026). This suggests that some “nonlocal” burdens can be transferred from the kernel to the coordinate system.

A persistent misconception is that any operator applied on a small patch is automatically local in the strict mathematical sense. The additive-manufacturing FNO is explicitly local only through its heat-affected windows; inside each window, the operator still performs global Fourier convolution (Chen et al., 2023). Conversely, some models that are not labeled LNOs, such as graph neural simulators or Flux FNO, instantiate strong locality biases by learning local derivatives or local fluxes (Nayak et al., 7 Sep 2025, Kim et al., 2024). The term therefore names a family resemblance rather than a single canonical architecture.

5. Empirical performance across PDE classes

The earliest varied-domain LNO results were reported for transient Navier–Stokes. Trained on random flows in a periodic square and then reused as an explicit time-marching scheme on unseen domains, the model achieved lower reported errors than FNO at $\mathcal{S}_{\Delta t}$4 s for viscosities $\mathcal{S}_{\Delta t}$5 and $\mathcal{S}_{\Delta t}$6, while using fewer trainable weights. On a lid-driven cavity at $\mathcal{S}_{\Delta t}$7, the reported mean absolute errors were $\mathcal{S}_{\Delta t}$8 for normalized $\mathcal{S}_{\Delta t}$9 and $11\times11\times11$0 for normalized $11\times11\times11$1, with total hybrid LNO-plus-boundary-treatment time $11\times11\times11$2 s per step versus $11\times11\times11$3 s for implicit FEM. On a cascade of NACA0012 airfoils, the reported relative error rate was $11\times11\times11$4, and the full $11\times11\times11$5–$11\times11\times11$6 s simulation took $11\times11\times11$7 s with LNO plus boundary treatment versus $11\times11\times11$8 s for FEM, a speedup of $11\times11\times11$9 (Li et al., 2022).

The locality analysis paper confirms that performance depends sharply on architectural locality. For 2-D compressible flow on a $C^s$0 periodic grid, a baseline task $C^s$1 failed catastrophically for $C^s$2 with $C^s$3, $C^s$4, and $C^s$5, whereas moderate $C^s$6 values such as $C^s$7, $C^s$8, or $C^s$9 yielded stable errors on the order of $W^{s,p}$0–$W^{s,p}$1 in density and temperature and around $W^{s,p}$2–$W^{s,p}$3 in velocity. For smaller Mach number $W^{s,p}$4, the compatible locality shifted to larger $W^{s,p}$5, and for larger Mach number $W^{s,p}$6, it shifted back to smaller $W^{s,p}$7 (Ye et al., 2023).

Localized-kernel augmentations of FNO and SFNO report substantial improvements on canonical benchmarks. On Darcy flow, adding a differential kernel reduced relative $W^{s,p}$8 error from $W^{s,p}$9 for FNO to $n$0; on turbulent 2D Navier–Stokes, combining differential and local integral kernels reduced 1-step error from $n$1 to $n$2; and on spherical shallow water equations, adding a local integral kernel reduced 5-step autoregressive error from $n$3 to $n$4. The paper summarizes the gains as relative $n$5-error reductions of $n$6–$n$7 on turbulent 2D Navier–Stokes and spherical shallow water benchmarks (Liu-Schiaffini et al., 2024).

Local-update formulations perform particularly well on conservation-law structure. In ideal MHD, Flux NO achieved relative $n$8 error $n$9 at $$\mathcal{G}_\theta = \mathcal{P}_\textup{projection}\circ\mathcal{B}_n\circ\cdots\circ\mathcal{B}_1\circ\mathcal{P}_\textup{lifting},$$0 on the 2D test set, compared with $$\mathcal{G}_\theta = \mathcal{P}_\textup{projection}\circ\mathcal{B}_n\circ\cdots\circ\mathcal{B}_1\circ\mathcal{P}_\textup{lifting},$$1 for 2D FNO, $$\mathcal{G}_\theta = \mathcal{P}_\textup{projection}\circ\mathcal{B}_n\circ\cdots\circ\mathcal{B}_1\circ\mathcal{P}_\textup{lifting},$$2 for 2D FNO (heavy), and $$\mathcal{G}_\theta = \mathcal{P}_\textup{projection}\circ\mathcal{B}_n\circ\cdots\circ\mathcal{B}_1\circ\mathcal{P}_\textup{lifting},$$3 for 3D FNO. The reported inference time for $$\mathcal{G}_\theta = \mathcal{P}_\textup{projection}\circ\mathcal{B}_n\circ\cdots\circ\mathcal{B}_1\circ\mathcal{P}_\textup{lifting},$$4 was about $$\mathcal{G}_\theta = \mathcal{P}_\textup{projection}\circ\mathcal{B}_n\circ\cdots\circ\mathcal{B}_1\circ\mathcal{P}_\textup{lifting},$$5 s, compared with $$\mathcal{G}_\theta = \mathcal{P}_\textup{projection}\circ\mathcal{B}_n\circ\cdots\circ\mathcal{B}_1\circ\mathcal{P}_\textup{lifting},$$6 s for WENO-Z, or roughly $$\mathcal{G}_\theta = \mathcal{P}_\textup{projection}\circ\mathcal{B}_n\circ\cdots\circ\mathcal{B}_1\circ\mathcal{P}_\textup{lifting},$$7 faster on the reported setup (Kim et al., 2024).

In low-data regimes, graph-based local surrogates also show a marked advantage over neural-operator baselines. For 2D Burgers, coupled Burgers, and Allen–Cahn, GNS achieved under $$\mathcal{G}_\theta = \mathcal{P}_\textup{projection}\circ\mathcal{B}_n\circ\cdots\circ\mathcal{B}_1\circ\mathcal{P}_\textup{lifting},$$8 relative $$\mathcal{G}_\theta = \mathcal{P}_\textup{projection}\circ\mathcal{B}_n\circ\cdots\circ\mathcal{B}_1\circ\mathcal{P}_\textup{lifting},$$9 errors with only $h\to0$0 training samples out of $h\to0$1, and averaged across all cases reduced autoregressive error by $h\to0$2 relative to FNO AR and $h\to0$3 relative to DON AR (Nayak et al., 7 Sep 2025). Hybrid FNO designs report comparable improvements from explicitly encoded local structure: UNet-FNO improves over vanilla FNO on Darcy flow, Allen–Cahn, incompressible Navier–Stokes, and compressible Navier–Stokes (Liu et al., 22 Mar 2025), while LoGlo-FNO reports parameter reduction by up to $h\to0$4 while matching baseline FNO accuracy and reports substantial gains on Kolmogorov flow, diffusion-reaction, and 3D turbulent radiative mixing (Kalimuthu et al., 5 Apr 2025).

6. Beyond forward prediction: system-level analysis, control, and current directions

A notable development is the use of local neural operators as short-time coarse timesteppers for equation-free computation. In this setting, the learned object is

$h\to0$5

and fixed points are zeros of

$h\to0$6

The learned timestepper is embedded inside Jacobian-free Newton-Krylov iterations,

$h\to0$7

with Jacobian-vector products approximated by finite differences, and Arnoldi iteration is used for leading eigenvalues. On Allen–Cahn, Liouville–Bratu–Gelfand, and FitzHugh–Nagumo, local-in-time, local-in-space, and patch operators were used to compute steady states, stability spectra, and bifurcation diagrams. Reported results include RandONet MSE around $h\to0$8 and median $h\to0$9 error $$\frac{\partial u_{t+\Delta t}(x_a)}{\partial u_t(x_b)}=0,\qquad \forall \|x_a-x_b\|>r,$$00 for a single-parameter Allen–Cahn case, homotopy-based recovery of the Bratu saddle-node at $$\frac{\partial u_{t+\Delta t}(x_a)}{\partial u_t(x_b)}=0,\qquad \forall \|x_a-x_b\|>r,$$01, and a Gap-Tooth RandONet training time of $$\frac{\partial u_{t+\Delta t}(x_a)}{\partial u_t(x_b)}=0,\qquad \forall \|x_a-x_b\|>r,$$02 s with MSE $$\frac{\partial u_{t+\Delta t}(x_a)}{\partial u_t(x_b)}=0,\qquad \forall \|x_a-x_b\|>r,$$03 (Fabiani et al., 5 May 2025).

The same short-time local-operator idea has been extended to control of distributed parameter systems. A learned solution operator

$$\frac{\partial u_{t+\Delta t}(x_a)}{\partial u_t(x_b)}=0,\qquad \forall \|x_a-x_b\|>r,$$04

is used to compute steady states, dominant eigenspectra, and reduced linear models by Newton-Krylov and Arnoldi, after which dLQR and pole-placement controllers are designed on the reduced system and lifted back to the full nonlinear dynamics. In the parabolic Liouville–Bratu–Gelfand example with three Gaussian actuators, the NO-based controller stabilized the unstable steady state with final error around $$\frac{\partial u_{t+\Delta t}(x_a)}{\partial u_t(x_b)}=0,\qquad \forall \|x_a-x_b\|>r,$$05, compared with near machine precision for the controller based on the exact finite-difference model (Fabiani et al., 28 Sep 2025). This recasts local neural operators as computational kernels for coarse analysis rather than only as rollout surrogates.

Current work increasingly treats locality as something to be allocated across representation, kernel, and coordinates. ACT reports average relative error reductions of $$\frac{\partial u_{t+\Delta t}(x_a)}{\partial u_t(x_b)}=0,\qquad \forall \|x_a-x_b\|>r,$$06 for Transolver + ACT, $$\frac{\partial u_{t+\Delta t}(x_a)}{\partial u_t(x_b)}=0,\qquad \forall \|x_a-x_b\|>r,$$07 for FNO + ACT, and $$\frac{\partial u_{t+\Delta t}(x_a)}{\partial u_t(x_b)}=0,\qquad \forall \|x_a-x_b\|>r,$$08 for ConvNeXt-UNet + ACT across six PDE datasets, with layer-wise coordinate adaptation outperforming simple parameter scaling (Liu et al., 7 May 2026). Patch-based latent reductions pursue a different route: localized PCA-Net decomposes solution fields into patches, applies PCA within each patch, and trains a neural operator in the reduced latent space, reducing end-to-end pipeline time by a factor of $$\frac{\partial u_{t+\Delta t}(x_a)}{\partial u_t(x_b)}=0,\qquad \forall \|x_a-x_b\|>r,$$09 to $$\frac{\partial u_{t+\Delta t}(x_a)}{\partial u_t(x_b)}=0,\qquad \forall \|x_a-x_b\|>r,$$10 relative to global PCA while requiring refinements such as overlap smoothing or CNN postprocessing to suppress block artifacts (Dhingra et al., 9 Sep 2025).

Several tensions remain stable across this literature. Locality that is too weak leaves models data-hungry, over-smooth, or unstable under long rollout; locality that is too strong can truncate the true dependence range, lose global coherence, or require expensive multiscale patch coupling. Some papers therefore advocate explicitly combining local and global mechanisms rather than choosing only one. A plausible implication is that the mature form of local neural-operator design is not a purely local architecture, but a structured decomposition in which local spatial interactions, local temporal evolution, and selected nonlocal channels are assigned distinct computational roles.