- The paper demonstrates that the SNO architecture reduces oversmoothing in shock-dominated PDEs by leveraging directional, shearlet-based spectral decompositions.
- It introduces a novel methodology that adapts scale and directionality for efficient handling of anisotropy, shocks, and multi-scale features with fewer parameters and less training data.
- Experimental evaluations reveal that SNO outperforms FNO in preserving sharp features and achieving lower errors across various anisotropic and discontinuity-rich PDE families.
Shearlet Neural Operators for Anisotropic-Shock-Dominated and Multi-Scale PDEs
Introduction and Motivation
The development of resolution- and geometry-adaptive neural surrogates for parametric PDE families is an active research area motivated by applications in uncertainty quantification, inverse problems, and physical simulation. Fourier Neural Operators (FNOs) are arguably the most widely-adopted neural operator architecture for learning the forward map of parametric PDEs due to their expressiveness, resolution-invariance, and computational efficiency. However, FNOs exhibit marked deficiencies in settings dominated by anisotropy, shocks, or multiscale directional structure—commonplace in advection, hyperbolic conservation laws, and complex diffusion, where sharp interfaces, fronts, and localized discontinuities arise.
This paper introduces the Shearlet Neural Operator (SNO) as a solution to these FNO limitations. SNO replaces the global, isotropic Fourier representation with a shearlet-based spectral decomposition, which possesses precise directional selectivity and spatial localization. This transforms the inductive bias of the neural operator, aligning it more closely with the anisotropic and discontinuous geometries typical of transport- and shock-dominated PDEs. The SNO architecture enables end-to-end differentiable learning in the shearlet domain, leveraging the theoretical sparsity optimality of shearlets for edge- and front-dominated data.
Architecture and Theoretical Properties
Neural Operator Framework
All neural operator architectures target the approximation of nonlinear solution operators between infinite-dimensional Banach spaces. They parameterize nonlocal, resolution-invariant kernel maps, learned from simulation data. FNOs use Fourier basis truncation to approximate translation-invariant kernels efficiently in spectral space, but this inherently biases learning toward smooth, isotropic phenomena and results in oversmoothing when representing localized or highly directional physics.
Shearlet-Based Spectral Parameterization
SNO embeds a scale- and orientation-dependent decomposition inspired by digital shearlet systems. The construction combines:
- Radial (scale) localization: Meyer-type smooth windows delineate dyadic bands in frequency space.
- Angular (directional) selectivity: Smooth windowing around angular orientations yields wedge-shaped bands matched to shearlet cones.
- Tight-frame normalization: Ensures energy is equitably distributed and the transform is numerically stable.
SNO thus learns complex-valued spectral mixing weights per shearlet band, supplemented by learnable scale gates (sigmoid-activated multipliers per scale) for adaptive anisotropy emphasis. The global nonlocality of spectral neural operators (shared with FNO) is preserved, but with enhanced geometric adaptivity.
The SNO block replaces the FNO spectral convolution with this shearlet-based operation, followed by pointwise feedforward mixing and nonlinearities. Multiple SNO layers are stacked to build a deep surrogate mapping from parameter field to PDE solution field. Critically, SNO's capacity is distributed anisotropically, adaptively targeting the fine-scale, directionally-organized content in the data.
Key Properties
- Directional multiscale locality: SNO can capture sharp, oriented features such as ridges, edges, and discontinuities efficiently.
- Spectral sparsity and efficiency: Reflecting the optimality of shearlets for cartoon- and singularity-rich functions, SNO requires fewer parameters and less training data than FNO for anisotropic regimes.
- Resolution invariance and global receptive field: Retains the strengths of spectral operator learning.
- Scalability: The digital implementation is FFT-based and differentiable, compatible with large-scale operator learning routines.
Experimental Evaluation
Benchmark Design
Seven distinct PDE families are posed to evaluate SNO against FNO:
- Diffusion-Dominated: Anisotropic, multi-orientation wave textures.
- Convection-Dominated: Bent ridge advection, anisotropic ridge advection, sheared Kelvin-Helmholtz stripes, and polygonal shocks—problems with strong directional advection and sharp, front-like features.
- Combined-Effect: 2D Burgers equation with multiscale, multi-angle shocks and spiral shock patterns, testing shock-capturing capabilities and nonlinear interaction.
Both models are parameter-matched (SNO with 12k parameters; FNO with 17k), trained for next-step field prediction using a supervised, autoregressive setup (MSE loss, AdamW optimizer).
Quantitative Results
SNO consistently exceeds FNO accuracy in six of seven PDE families:
| Dataset |
SNO L2 (N=128) |
FNO L2 (N=128) |
Relative L2 Ratio |
SNO SSIM |
FNO SSIM |
| Anisotropic Ridge Advect |
1.6e-5 |
2.0e-4 |
0.08 |
0.9998 |
0.9862 |
| Multi Angle Shocks |
2.3e-5 |
7.3e-5 |
0.32 |
0.9973 |
0.9859 |
| Bent Ridge Advect |
1.4e-5 |
1.4e-4 |
0.10 |
0.9999 |
0.9909 |
| Multi Orientation Texture |
1.3e-5 |
2.2e-5 |
0.59 |
0.9994 |
0.9960 |
| Spiral Shock |
4.2e-5 |
3.4e-5 |
1.23 |
0.9921 |
0.9950 |
| Polygonal Shock |
1.3e-5 |
3.5e-5 |
0.37 |
0.9987 |
0.9900 |
| Sheared Kelvin-Helmholtz |
2.7e-5 |
1.1e-4 |
0.26 |
0.9986 |
0.9729 |
On advection and combined shock test cases, SNO achieves L2 errors that are 3–10x (or more) lower than FNO. SNO near-perfectly preserves sharp structures, whereas FNO suffers severe oversmoothing and phase errors, with absolute L2 error up to several orders of magnitude larger, especially at reduced grid resolution or with less data.
SSIM metrics confirm SNO's superior structure and orientation preservation—crucial for scientific surrogates. The sole exception is the spiral shock configuration, where SNO and FNO are statistically tied, likely due to the dominance of smooth, globally periodic features in that specific problem.
Loss curve analysis reveals that SNO achieves lower loss plateaus faster and with smaller datasets, whereas FNO's convergence saturates early. This directly supports the claim that shearlet-based operators are both more data-efficient and offer higher representational capacity for anisotropic, multiscale, and discontinuity-rich PDE systems.
Qualitative Analysis
Case-by-case inspections show:
- Diffusion-dominated (multi-orientation textures): SNO captures all scales and directions with high fidelity, requiring less data and grid resolution for accurate generalization.
- Shock and advection-dominated (bent/anisotropic ridges, polygonal/multi-angle/spiral shocks, Kelvin-Helmholtz): SNO produces visually indistinguishable predictions from ground truth, with errors winnowing to negligible values even at sharp gradients and fronts. FNO’s artifacts—smearing, ringing, and loss of orientation—are persistent and pronounced.
- Mesh refinement: SNO benefits more from increased resolution than FNO, but even at moderate resolutions, SNO exhibits lower errors and better feature preservation.
Implications and Future Directions
Practically, SNO enables neural operator learning for challenging physical regimes where FNO or standard convolutional surrogates fail: high-Péclet, shock-rich, turbulent, or multi-phase transport, where resolving moving, interacting, and curved fronts and interfaces is essential. Its improved fidelity in representing discontinuities, anisotropy, and multi-scale textures positions it as a superior surrogate for scientific computing tasks such as:
- Turbulence modeling, multiphase and reactive flows, atmospheric/oceanic simulation, seismic imaging
- UQ and inverse problem setups where parameter variability creates dynamically shifting and oriented localized structures
- Scenarios requiring sample efficiency (reduced training data) and less grid refinement due to optimal spectral sparsity
Theoretically, SNO bridges state-of-the-art neural operator learning with harmonic analysis advances. Its design is extensible: advanced polar or parabolic tilings, more sophisticated scale/shear gating, or task-informed hybridization with wavelets or graph-based representations. Future development could extend SNO to irregular meshes, higher dimensions, and physical constraints via PINO hybridization. Its architecture can serve as a model for operator learning in computer vision and scientific domains where geometric complexity dominates.
Conclusion
The Shearlet Neural Operator systematically outperforms the Fourier Neural Operator for PDE classes dominated by anisotropy, discontinuities, and multi-scale structure. The adoption of shearlet-based, directional spectral parameterizations amends the inherent oversmoothing and isotropy bias of Fourier-based operators. SNO realizes sharper, more structurally faithful, and data-efficient surrogates in shock- and front-rich regimes, without sacrificing the global coupling and FFT-based scalability needed for large-scale operator learning. This work points toward a new synthesis of machine learning and harmonic analysis, opening directions for neural operators that are both theoretically grounded and practically robust in physical simulation and scientific machine learning.
