Localized Fourier Neural Operator (LoFNO)

• Localized Fourier Neural Operator (LoFNO) is a neural architecture that combines global spectral aggregation with explicit localization to capture both low and high frequency features.
• It leverages parallel local-global branches, patchwise FFTs, and differential/integral kernels to enhance reconstruction accuracy and maintain resolution-independence in PDE models.
• Empirical studies demonstrate that LoFNO improves performance in fluid dynamics, urban CFD, and medical imaging by efficiently managing computational overhead and ensuring fine-scale detail recovery.

A Localized Fourier Neural Operator (LoFNO) is a neural architecture for learning operators between function spaces that combines global spectral aggregation with explicit localization in either the frequency or spatial domain. LoFNOs address the documented limitations of global FNOs, including spectral bias toward low frequencies, the inability to capture key local structures, memory inefficiency for high-resolution domains, and limited accuracy in tasks requiring fine-scale detail or geometric adaptivity. The following sections provide a detailed account of LoFNO theory, model instantiations, algorithmic structures, empirical evidence, and technical considerations across multiple research lines.

1. Theoretical Foundations

LoFNOs generalize the Fourier Neural Operator architecture by introducing localization mechanisms that break the pure globality of vanilla FNOs. The core mathematical object remains an operator mapping between functional spaces, e.g., $u\colon D\to\mathbb{R}^c$, typically solution fields of PDEs on continuous or discrete spatial domains.

The principal design paradigms are:

• Local Spectral Convolution: Partition the domain into patches or restrict spectral convolutions to windowed intervals, enabling efficient capturing of higher-frequency content.
• Localized Spatial Kernels: Employ differential (finite-difference) or integral (locally supported) kernels, often parameterized via basis expansions with shared learnable coefficients.
• Frequency-local Propagators: In frequency space, replace dense or full-spectrum multipliers with block-banded, windowed, or sparse structures tailored to the physics of the underlying operator (e.g., wave propagation locality).
• Superposition Principle Preservation: Architectures maintain linearity or approximate it closely, ensuring physical interpretability and generalization to superpositions of learned modes.

Formalizations vary from patchwise local FFT operators (Qin et al., 2024), parallel global-local spectral branches (Kalimuthu et al., 5 Apr 2025), learned localized windowed propagators (Cai et al., 15 Mar 2026), and hybrid local feature extractors plus FNO blocks (Liu et al., 22 Mar 2025).

2. Representative Architectures

2.1 Parallel Local-Global Branches (LoGlo-FNO)

The LoGlo-FNO architecture integrates three parallel streams at each Fourier layer: (i) a global Fourier convolution for dominant low-frequency and long-range dependencies, (ii) a local spectral convolution performed independently on non-overlapping spatial patches retaining all Nyquist-local modes, and (iii) a high-frequency propagation (HFP) skip connection that isolates and reinjects high-frequency residuals lost to sub-sampling or damping. The outputs are fused, typically by summation, then passed to nonlinearities and subsequent layers. This design directly improves moderate and high-frequency reconstruction, while reducing parameter count owing to the lower-dimensional local patch convolutions (Kalimuthu et al., 5 Apr 2025).

2.2 Localized Integral and Differential Kernels (LoFNO)

This class integrates localized CNN-derived stencils representing differential operators and locally supported integral kernels into the Fourier Neural Operator layer stack. Differential branches use finite-difference style kernels rescaled according to the mesh size, yielding resolution-agnostic discrete differential operators. Integral branches apply locally supported basis expansions (e.g., radial hat functions), constructing local averaging/aggregation effects. The standard global FNO is preserved as a parallel branch, and all outputs are aggregated before nonlinearity (Liu-Schiaffini et al., 2024).

2.3 Patchwise and Windowed Spectral Operators

The Local-FNO introduces overlapping physical patches wherein FFTs and spectral convolutions are performed independently, with outputs blended smoothly by partition-of-unity weights to ensure continuity and avoid artifacts. Geometry encoding (e.g., signed distance functions) is injected for local structural awareness, and the entire design is highly memory-efficient for high-resolution, multi-variable domains such as urban CFD (Qin et al., 2024).

The Windowed Fourier Propagator (WFP) learns a separate, compact propagator for each frequency window in the spectral domain, motivated by physical frequency locality observed in wave equations. Each block (local propagator) only mixes frequencies within a chosen window, and the large, global convolution tensor of a standard FNO is replaced by many smaller bands, enabling O(N) complexity where N is the number of excited modes (Cai et al., 15 Mar 2026).

2.4 Local Feature Extraction Hybrids

The Conv-FNO approach prepends a CNN (possibly a UNet) that extracts local spatial features from the input, which are concatenated with the original channels before feeding into global FNO blocks. This hybrid increases the function class expressivity, capturing local PDE-induced structures without significantly increasing computational overhead (Liu et al., 22 Mar 2025).

2.5 Domain-Restricted Spectral Operators with Prior

In 3D medical flow upsampling, LoFNO augments an FNO core with spatial priors derived from Laplacian eigenvectors on irregular geometries, upsampling via enhanced deep super-resolution (EDSR) blocks, and domain restriction masks that enforce operator locality within anatomical boundaries (Flouris et al., 18 Jul 2025).

3. Mathematical Formulations

The localization principle manifests via:

• Spatial Patch FFT: For each patch $\Omega_i$, compute FFT, multiply modes by learned $R_{\theta,i}(k)$, inverse FFT, blend by $\omega_i(x)$ if overlapping (Qin et al., 2024).
• Finite-Difference Stencils: Kernel convolution approximates derivatives, scaled as $1/h$ for first derivatives. For higher-order, increase stencil width or stack (Liu-Schiaffini et al., 2024).
• Integral Kernels with Local Support: Basis expansion: $k(x,y) = \sum_\ell \theta_\ell \varphi_\ell(x,y)$, with $$\mathrm{supp}\,\varphi_\ell(\cdot, y) \subset B_R(y)$$; quadrature for discretization (Liu-Schiaffini et al., 2024).
• Windowed Propagators: For Fourier mode $k$, define $\mathcal{W}(k, r)$; propagate via $P_k$ only within $\Omega_i$0, preserving superposition (Cai et al., 15 Mar 2026).
• Domain-Localized Fourier Layers: For $\Omega_i$1 mask $\Omega_i$2, interaction restricted to domain via $\Omega_i$3; spectral convolution performed within masked region (Flouris et al., 18 Jul 2025).
• Local Feature Injection: $\Omega_i$4, concatenated with input, then processed by standard FNO (Liu et al., 22 Mar 2025).

4. Algorithmic Components and Implementation

Key steps for constructing LoFNOs:

• Domain Partitioning: Divide domain into patches (non-overlapping or overlapping, possibly with geometry-aware allocations).
• Frequency Band or Patch-Local Parameterization: Learn independent or tied spectral weights for each patch/local region, with possible parameter sharing to manage memory.
• Loss Engineering: Employ frequency-sensitive loss functions, for instance, radial-binned spectral penalties that emphasize reconstruction in mid- and high-frequency bands (Kalimuthu et al., 5 Apr 2025).
• Resolution Independence: Ensure operator weights or stencils are grid-agnostic, with rescaling or quadrature schemes that guarantee transfer between training and inference resolutions (Liu-Schiaffini et al., 2024, Liu et al., 22 Mar 2025).
• Geometric Encoding: Incorporate mesh-derived spectral embeddings or signed distance functions as part of the feature stack for spatially irregular domains (Flouris et al., 18 Jul 2025, Qin et al., 2024).
• Fusion and Aggregation: Aggregate outputs of parallel branches (global/local) via elementwise or weighted sums before nonlinearity, ensuring signal flow from all scales.
• Efficiency Considerations: Favor small patch sizes or frequency window radii to reduce parameter and computational load, while leveraging parallelization—especially via optimized libraries for FFT and convolution.

5. Empirical Performance and Benchmarks

Multiple experimental studies establish the empirical superiority of LoFNO variants across challenging benchmarks:

Task / Domain	Metric	FNO	LoFNO / Variant	Relative Improvement
2D Kolmogorov Flow (Kalimuthu et al., 5 Apr 2025)	1-step nRMSE	0.107	0.072 (LoGlo-FNO)	–27%
Turbulent Navier-Stokes (Liu-Schiaffini et al., 2024)	rel $\Omega_i$5	0.138	0.0902	–35%
Urban Microclimate (3D) (Qin et al., 2024)	RMSE (velocity)	–	0.35	–23.9% (vs FNO)
Hemodynamic Upsampling (Flouris et al., 18 Jul 2025)	Rel. $\Omega_i$6 error (u, WSS)	0.0647/0.9832	0.0452/0.7625	≈ –30–40% (u), ≈ –20–30% (WSS)

Ablation studies consistently demonstrate that pure spectral global operators are insufficient for recovering high-frequency or localized physical features, and that removing local/differential/integral/K-local frequency or spatial branches reverts accuracy to global FNO baselines. Adding explicit local feature preprocessing or geometric prior channels provides further systematic gains (Liu et al., 22 Mar 2025, Flouris et al., 18 Jul 2025).

Resolution-invariance and robustness to distribution shift are achieved via discretization-aware kernel scaling, spectral blending, and strong inductive biases toward physical localization (Liu-Schiaffini et al., 2024, Qin et al., 2024, Cai et al., 15 Mar 2026).

6. Applications and Limitations

LoFNOs have broadened the applicability of neural operator models to:

• Turbulence and Fluid Dynamics: Accurate prediction of vorticity, energy, and dissipation at intermediate and high Reynolds numbers (Kalimuthu et al., 5 Apr 2025, Liu-Schiaffini et al., 2024).
• Urban-Scale CFD: Fast, memory-efficient surrogate modeling for 3D wind/heat fields at meter-scale resolution over kilometer-scale domains (Qin et al., 2024).
• Medical Imaging: Denoising and upsampling of hemodynamic flow data, including the accurate interpolation of wall shear stress in irregular vascular geometries (Flouris et al., 18 Jul 2025).
• Wave Propagation: Efficient and physically faithful learning of wave equations in variable media, with explainable block-banded Fourier-domain operators (Cai et al., 15 Mar 2026).
• Generic PDE Surrogates: Robustness on reaction-diffusion, Allen–Cahn, and Darcy flow equations, especially in low-data regimes (Liu et al., 22 Mar 2025, Liu-Schiaffini et al., 2024).

Limitations include additional implementation complexity from patching, geometric blending, and feature fusion; increased overhead for patch overlap; and potential failure to capture long-scale phenomena if only local kernels are used. Purely local operators may underperform on problems where distant/global couplings are physically essential (Qin et al., 2024). Hybrid and two-level operator designs are a proposed future direction to mitigate this (Qin et al., 2024).

7. Future Directions

Ongoing research aims to further integrate global and local spectral modeling, exploit more expressive geometric encodings, and adapt localization dynamically based on learned problem structure. Data-driven kernel learning informed by underlying PDE structure remains a central theme.

Generalizing LoFNOs to non-Euclidean or non-uniform domains, incorporating attention-based fusion mechanisms, and developing scalable training for very high-dimensional or multi-physics settings represent current frontiers. A plausible implication is that as these innovations mature, the LoFNO framework will increasingly serve as the default architecture for operator learning tasks where local structure, fine-resolution fidelity, and computational efficiency are simultaneously required (Kalimuthu et al., 5 Apr 2025, Qin et al., 2024, Liu-Schiaffini et al., 2024, Cai et al., 15 Mar 2026, Liu et al., 22 Mar 2025).