GLEAM: спектральные потери для прогноза турбулентности на неструктурированных сетках

In July 2026, a paper was published on arXiv that transfers spectral loss functions to unstructured computational meshes: the researchers replaced Fourier frequency bands with graph Laplacian bands and introduced the GLEAM method for more accurate long-term forecasting of turbulent flows.

Why Fourier doesn't work for irregular meshes

Surrogate models replace heavy physical simulation with a neural network, but for chaotic systems, what matters is not just point-wise accuracy but preserving the structure of the field across different scales. On regular grids, this is controlled through spectral losses: Fourier modes define a standard decomposition by frequency, and the model is penalized for power discrepancies in each frequency band.

The problem is that on unstructured meshes, a canonical Fourier basis does not exist. Real geometry — a wing, a turbine, a riverbed — is almost always described by an irregular mesh, and standard spectral control cannot be applied to it directly. The spectral representation has to be built from graph operators defined by the connectivity and geometry of the mesh itself.

What the authors proposed

The authors replaced Fourier bands with frequency bands of the graph Laplacian — an operator built directly from the connectivity and geometry of the mesh. This gives a graph analogue of the same band-wise power control used on a regular grid.

Key elements of the work:

• Fourier bands are replaced with frequency bands of the graph Laplacian
• The full version uses the Laplacian's eigen-subspaces — accurate, but expensive due to spectral decomposition
• A scalable approximation — sparse polynomial Chebyshev graph filters, without explicit eigenvalue decomposition
• GLEAM (Graph Laplacian Energy Alignment for Meshes) — a multi-level variant that aligns scales across the graph hierarchy
• The task is forecasting turbulent flows on unstructured meshes

The GLEAM method applies scale-aware control at all levels of the graph hierarchy, so that coarse and fine representations are regularized simultaneously during autoregressive forecasting.

The full, "spectral" version of the method gives the most accurate match across all scales, but requires decomposing the graph Laplacian into eigenvectors — a heavy operation that scales poorly to large meshes. That is exactly why the authors propose two workarounds — Chebyshev filters and the multi-level GLEAM hierarchy.

"Our approach uses the eigen-subspaces of the graph

Laplacian to give a graph analogue of matching the power of Fourier spectral bands," the paper's abstract on arXiv states.

What the experiments showed

The proposed spectral losses improve the accuracy of long-term autoregressive forecasting and preserve the statistical invariants of turbulent flows — better than deterministic baseline models. This concerns long-horizon rollout forecasting, where ordinary models quickly accumulate error and "drift" away from physically plausible behavior.

The two approximations — Chebyshev and multi-level — solve the main practical problem of full spectral decomposition: its computational cost. Sparse Chebyshev filters avoid computing the Laplacian's eigenvectors explicitly, which makes the approach applicable to large meshes.

What this means

The work closes the gap between spectral learning and real computational fluid dynamics, where meshes are almost always irregular. If the method scales, surrogate neural networks will be able to forecast turbulence more accurately and cheaply on engineering geometries — cases where full simulation is too expensive. For engineers, this means potentially faster aerodynamics and hydrodynamics calculations without losing physical fidelity over long forecasts.

Frequently Asked Questions

What is GLEAM?

GLEAM (Graph Laplacian Energy Alignment for Meshes) is a multi-level variant of the spectral loss function that aligns signal energy across the graph hierarchy, regularizing coarse and fine mesh representations simultaneously during autoregressive forecasting.

Why replace Fourier with the graph Laplacian?

On unstructured meshes there is no canonical Fourier basis, so the standard frequency decomposition doesn't apply. The graph Laplacian is built from the connectivity and geometry of the mesh and provides an analogous frequency decomposition directly on the irregular mesh.