GLEAM: Spectral Losses for Turbulence Forecasting on Unstructured Meshes

A paper published on arXiv in July 2026 carries spectral loss functions over to unstructured computational meshes. The researchers replaced Fourier frequency bands with graph Laplacian bands and introduced GLEAM, a method for more accurate long-term forecasting of turbulent flows.

Why Fourier doesn't work on irregular meshes

Surrogate models replace a heavy physics simulation with a neural network. For chaotic systems, though, accuracy at each individual point is not enough: the model also has to preserve the structure of the field across different scales. On regular grids, this is controlled through spectral losses. Fourier modes provide the standard decomposition by frequency, and the model is penalized for any mismatch in power within each frequency band.

The problem is that unstructured meshes have no canonical Fourier basis. Real-world geometry (a wing, a turbine, a riverbed) is almost always described by an irregular mesh, so standard spectral control cannot be applied to it directly. The spectral representation has to be built from graph operators, which are 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 mesh's connectivity and geometry. This gives a graph analogue of the same band-wise power control used on regular grids.

Key elements of the work:

• Fourier bands are replaced with frequency bands of the graph Laplacian
• The full version uses the Laplacian's eigensubspaces: it is exact, but expensive because it requires a spectral decomposition
• The scalable approximation uses sparse polynomial Chebyshev graph filters, with no explicit eigenvalue decomposition
• GLEAM (Graph Laplacian Energy Alignment for Meshes) is a multilevel variant that aligns scales across the graph hierarchy
• The target task is forecasting turbulent flows on unstructured meshes

GLEAM applies scale-aware control at every level of the graph hierarchy, so coarse and fine representations are regularized simultaneously during autoregressive forecasting.

The full "spectral" version of the method gives the closest match across all scales, but it requires an eigendecomposition of the graph Laplacian. That is a heavy operation that scales poorly to large meshes. This is why the authors offer two workarounds: Chebyshev filters and the multilevel GLEAM hierarchy.

"Our approach uses the eigensubspaces of the graph

Laplacian to provide a graph analogue of Fourier spectral band power matching," the paper's arXiv abstract says.

What the experiments showed

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

The two approximations, Chebyshev and multilevel, address the main practical problem of the full spectral decomposition: its computational cost. Sparse Chebyshev filters make it possible to avoid computing the Laplacian's eigenvectors explicitly, which makes the approach applicable to large meshes.

What this means

The work closes a gap between spectral learning and real computational fluid dynamics, where meshes are almost always irregular. If the method scales, surrogate neural networks could forecast turbulence on engineering geometry more accurately and more cheaply, in cases where full simulation is too expensive. For engineers, this could mean faster aerodynamics and hydrodynamics calculations without losing physical fidelity on long forecasts.

Frequently asked questions

What is GLEAM?

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

Why replace Fourier with the graph Laplacian?

Unstructured meshes have no canonical Fourier basis, so the standard frequency decomposition does not apply. The graph Laplacian is built from the mesh's connectivity and geometry, and it provides an analogous frequency decomposition directly on the irregular mesh.