NAGO: Noise-Aware Generative Operator via Flow Matching

Guanyu Chen, Shengze Xu, Pengwei Liu, Xingyu Ren, Pengkai Wang, Antonios Armaou, Tieyong Zeng, Dong Ni

TL;DR

NAGO (Noise-Aware Generative Operator) significantly accelerates the generation of complex physical fields by replacing generic Gaussian noise with custom-designed Matérn noise. By matching the source noise's regularity to the target data, NAGO reduces the computational cost of sampling by 6–10× while maintaining high physical accuracy and resolution consistency.

What Problem Does This Paper Address?

Stochastic Partial Differential Equations (SPDEs) are the mathematical workhorses for modeling everything from fluid turbulence to thin-film growth. However, sampling from their "invariant measures"—the long-term statistical equilibrium of these systems—is notoriously difficult.

Modern generative models like Flow Matching offer a path forward by learning to "transport" a simple source distribution (usually white noise) into the complex target distribution. The problem is that white noise is mathematically "rough" (it has excessive high-frequency energy), while physical fields are often smooth. This mismatch creates spectral stiffness: the learned model has to work incredibly hard to suppress those high frequencies, leading to unstable training and requiring dozens of expensive numerical steps to generate a single sample.

Methods

The authors introduce NAGO, which treats the source noise not as a fixed default, but as a design variable.

  1. Matérn Source Design: Instead of standard Gaussian noise, NAGO uses Matérn fields. These are Gaussian random fields where you can explicitly tune the smoothness ($\nu$) and correlation length ($\kappa$).
  2. Admissibility-Constrained Spectral Matching: The authors prove that for a flow to be well-defined in infinite-dimensional function space, the source noise must be "rougher" than the target data (the Feldman-Hájek theorem). NAGO uses a new rule: choose a source that is just rough enough for mathematical validity, then tune its parameters to minimize the "covariance mismatch" across all frequencies.
  3. Neural Operator Backbone: NAGO uses a Fourier Neural Operator (FNO) to learn the velocity field. Because FNOs operate in the frequency domain, they are naturally suited to handling the multiscale nature of SPDE data.
Figure 1: The structure of the NAGO model.

Datasets

The model was evaluated on three representative benchmarks:

  • Linear Matérn Fields: A controlled synthetic setting to validate the core theory.
  • Stochastic Allen-Cahn (SAC): A 1D nonlinear equation used to model phase separation, featuring a bimodal invariant measure.
  • Stochastic Navier-Stokes (SNS): A 2D fluid dynamics benchmark representing high-dimensional, multiscale complexity.

Key Insights and Ideas

  • The Stiffness Mechanism: The paper identifies that the ratio of source-to-target covariance at high frequencies is the primary driver of numerical stiffness. If this ratio is poorly tuned, the velocity field becomes "spiky" and difficult to integrate.
  • Source Regularity Matters: In function-space modeling, the choice of noise isn't just an implementation detail—it's a fundamental requirement for the model to be well-posed.
  • Resolution Consistency: Because NAGO aligns the noise regularity with the physics, the model remains stable even when you change the grid resolution (e.g., training at $64^2$ and sampling at $128^2$), a common failure point for standard generative models.
Comparison of marginal velocity fields: Standard white noise (bottom right) creates extremely sharp, stiff gradients compared to the smoother, well-conditioned NAGO flow (top right).

Key Results

  • Efficiency Gains: In controlled tests, NAGO reduced the required number of ODE solver steps from ~30 (for standard Gaussian flow matching) to just 3–5 steps. This represents a 6–10× reduction in computational effort.
  • Throughput: When compared to traditional numerical SPDE solvers (like the SNS pseudo-spectral solver), NAGO achieved a 89.1× raw throughput ratio post-training.
  • Spectral Accuracy: NAGO significantly outperformed baselines in "log-MSE," a metric that measures how well the model reproduces the energy spectrum across different scales.

Limitations and Future Work

  • Geometry and Stationarity: The current implementation assumes stationary measures on periodic, regular grids. Extending this to irregular geometries (like complex engine parts) or non-stationary systems is a key next step.
  • Spectral Tail Recovery: While NAGO is excellent at preserving the frequencies it was trained on, it does not "invent" or reconstruct unseen fine-scale physics (the spectral tail) beyond the training resolution.
  • Anisotropy: The current source selection rule focuses on isotropic noise. The authors suggest that learning anisotropic (direction-dependent) sources could further improve performance for complex physical systems.

Summary updated

Abstract

We introduce NAGO, a noise-aware generative operator for efficient sampling from invariant measures of stochastic partial differential equations (SPDEs) in function space. Our central idea is to treat the source measure as a design variable rather than a fixed Gaussian default. We establish admissibility conditions for well-defined infinite-dimensional flows and show that the finite-band source--target covariance ratio controls the spectral stiffness of the marginal velocity field. This analysis yields an admissibility-constrained spectral-matching rule: choose a source that is sufficiently rough for well-posedness, then minimize covariance mismatch over the resolved Fourier modes. On three representative stochastic-field benchmarks, this source design improves few-step spectral accuracy and several complementary distributional and physical statistics. In a controlled Matern-field ablation, it reduces the accurate RK4 budget from about 30 steps for Gaussian functional flow matching to 3--5 steps, a $6$--$10\times$ reduction in neural ODE evaluations. These gains are strongest for stationary multiscale fields and depend on the metric, resolution, and guidance protocol.
BibTeX
@misc{chen2026nago,
  author = {Guanyu Chen and Shengze Xu and Pengwei Liu and Xingyu Ren and Pengkai Wang and Antonios Armaou and Tieyong Zeng and Dong Ni},
  title = {{NAGO: Noise-Aware Generative Operator via Flow Matching}},
  year = {2026},
  eprint = {2610.38273},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.38273},
}
Figure