When we train a standard convolutional neural network or vision transformer to simulate physical systems—such as fluid dynamics, wave propagation, or weather patterns—we inadvertently bind the model to the specific grid resolution of the training data. If a model is trained on a $64 \times 64$ mesh, evaluating it on a $256 \times 256$ mesh typically fails or requires destructive interpolation. This is because traditional architectures learn mappings between discrete, finite-dimensional vector spaces ($\mathbb{R}^N \to \mathbb{R}^M$).
In contrast, the underlying physical laws governed by partial differential equations (PDEs) are continuous and independent of any discretization. Resolution-invariant operator learning shifts the paradigm by designing neural networks that operate directly on infinite-dimensional function spaces. By learning the underlying differential or integral operators rather than a mapping between fixed grids, these models achieve a remarkable property: zero-shot super-resolution. A model trained entirely on coarse, computationally cheap data can be deployed directly onto highly refined meshes at inference time without any retraining.
The Core Mechanism: Mapping Function Spaces
To understand why resolution-invariant models work, we must look at how they process data. Instead of treating an input as a matrix of pixel values, they treat it as discrete samples from a continuous function $u(x)$ defined over a domain. The goal is to learn an operator $\mathcal{G}$ that maps this input function to an output function $v(x) = \mathcal{G}(u)(x)$.
Traditional NN: [Coarse Grid Tokens] ---------> [Coarse Grid Predictions]
(Bound to Discretization)
Neural Operator: [Coarse Grid Samples]
│
▼ (Lifting to Function Space)
[Continuous Representation] ───> [Continuous Operator Kernel]
│ │
▼ (Zero-Shot Evaluation) ▼ (Zero-Shot Evaluation)
[Fine Grid Samples] ─────────> [Fine Grid Predictions]
Because the internal layers of these networks parameterize continuous mathematical operations—such as integration or coordinate-based mappings—the network's weights do not depend on the grid size. When evaluated on a finer mesh, the model simply samples the continuous learned operator at a higher density. As shown by Guanhang Lei and Lei Shi, learning operators between infinite-dimensional function spaces allows the architecture to maintain physical consistency and accuracy across entirely different spatial scales.
Architectural Paradigms for Discretization Invariance
How do we build a neural network layer that can process continuous functions? Researchers have converged on three primary mechanisms to decouple network parameters from spatial resolution.
1. Global Spectral and Wavelet Transforms
The most prominent approach leverages the convolution theorem. Instead of performing localized convolutions with fixed-size pixel kernels, the network transforms the input into the frequency domain, applies a linear transformation to the continuous spectral coefficients, and transforms back.
- Fourier Neural Operators (FNOs): By operating in the frequency domain, FNOs treat data as continuous fields. As Boris Bonev et al. demonstrated with Spherical Fourier Neural Operators (SFNOs) for global atmospheric modeling, spectral-based operator learning allows a model trained on a low-resolution grid to generalize seamlessly to higher-resolution data. Similarly, Sebastian Mizera applied FNOs to the Schrödinger equation, achieving zero-shot super-resolution from coarse training grids to fine simulation scales.
- Multiwavelets and Localized Spectra: While Fourier transforms excel at smooth, global periodic structures, they struggle with sharp discontinuities or shocks. To address this, Gaurav Gupta et al. introduced Multiwavelet-based Neural Operators (MWT), which compress integral kernels using the mathematical properties of multiwavelets. This allows the model to capture localized, high-frequency details while maintaining a $10\times$ reduction in error during zero-shot resolution transfer. Chenhong Zhou and Jie Chen further hybridized this via dual-path spectral attention, combining the global reach of Fourier transforms with the local precision of wavelets.
2. Attention as Numerical Quadrature
Another way to achieve resolution independence is to reframe the attention mechanism found in Transformers. If we view the continuous operator as an integral kernel:
$$v(x) = \int_{\mathcal{D}} \kappa(x, y) u(y) dy$$
we can interpret the attention matrix as a numerical approximation (a quadrature) of this continuous integral.
Kazem Meidani et al. exploited this in the Operator Transformer (OFormer), using cross-attention to decouple the input and output domains. By querying the model at arbitrary continuous coordinates rather than fixed grid points, the model maintains consistent error rates across varying spatial densities. To scale this concept to massive datasets, Seungjun Lee et al. proposed using a small set of learnable "inducing points" as a latent bottleneck, ensuring that the computational complexity scales linearly regardless of how dense the input or output meshes are.
3. Implicit Neural Representations (INRs)
Instead of passing grids through layers, some frameworks use coordinate-based networks (like MLPs) to parameterize the continuous field directly. Oussama Boussif et al. developed MAgNet, an "Encode-Interpolate-Forecast" framework that pairs Graph Neural Networks with INRs. Because the INR represents the solution as a continuous function mapping a coordinate $(x, y)$ to a physical state, the model can evaluate meshes up to four times denser than those seen during training while preserving physical consistency.
Why It Works: Learning Physics, Not Pixels
The fundamental reason resolution-invariant operators succeed where traditional networks fail comes down to what they are optimizing for. A standard CNN minimizes pixel-wise loss, leading it to memorize high-frequency grid patterns or aliasing artifacts unique to that specific discretization.
An operator learning model, however, isolates the scale-invariant properties of the physical system. Whether a fluid vector field is sampled at 50 points or 5000 points, the underlying Navier-Stokes equations governing its behavior remain identical. By forcing the network to learn mappings in the continuous spectral domain or via coordinate-free integral kernels, the model is incentivized to capture the true differential relationships.
This property allows for creative cross-disciplinary applications. For example:
- Medical Imaging: Ken C. L. Wong et al. developed ultra-lightweight 3D medical image segmentation models (HNOSeg-XS) using frequency-domain operators. Trained on low-resolution scans, these models generalize to high-resolution clinical data without a drop in accuracy, bypassing the "memory wall" that plagues standard 3D CNNs.
- Data Assimilation: Kun Chen et al. introduced Fourier Neural Processes (FNP) for weather data assimilation, allowing models to ingest sparse, arbitrary-resolution observations directly without requiring pre-processing or interpolation.
Challenges and the Role of Priors
Despite the theoretical elegance of resolution invariance, zero-shot scaling is not entirely free. As Alasdair Tran and Alexander Mathews observed, neural operators can still experience performance degradation if the testing resolution deviates drastically from the training resolution. If a model is trained only on extremely coarse data, it may never observe the high-frequency micro-dynamics (such as turbulent eddies or sharp shock waves) that naturally emerge at finer scales.
To bridge this gap, researchers inject physical or multi-scale priors into the learning process:
- Physics-Informed Constraints: Zongyi Li et al. introduced the Physics-Informed Neural Operator (PINO). By penalizing violations of the underlying PDE constraints on the fine grid during training, PINO uses explicit physical laws as a guide to accurately construct high-frequency details that are completely absent from the coarse training data.
- Multi-Resolution Wavelet Training: Rather than relying on a single grid, Peiyan Hu et al. utilized a generative approach in the wavelet domain (WDNO). By explicitly training a model to map coefficients across a multi-resolution wavelet hierarchy, the network learns how energy cascades from large scales to small scales, drastically improving its zero-shot super-resolution fidelity on highly dense grids.
By decoupling the architecture from the discretization mesh, resolution-invariant operator learning provides a mathematically rigorous foundation for building scientific AI models that are flexible, scalable, and fundamentally aligned with the continuous nature of the physical world.
Key Foundational Work
Multiwavelet-based Operator Learning for Differential Equations (2021) — contributed 2 concept elements to this direction.