Scale-Aware Learning for Chaotic Dynamics on Unstructured Meshes

Kanad Sen, Romit Maulik· July 23, 2026 View original

Summary

This research extends binned spectral losses to unstructured meshes for surrogate modeling of high-dimensional chaotic systems, using graph-Laplacian frequency bands. The approach, including scalable approximations like Graph Laplacian Energy Alignment for Meshes (GLEAM), improves long-horizon forecasting fidelity and preserves statistical invariants for turbulent flows.

Modeling high-dimensional, chaotic dynamical systems, especially with surrogate models, requires not just point-wise accuracy but also the preservation of scale-dependent physical field structures. While binned spectral power losses effectively provide this supervision on structured grids using Fourier modes, applying this to irregular meshes, which lack a canonical Fourier basis, presents a significant challenge. This study addresses this by extending the binned spectral power loss concept to unstructured-mesh surrogate modeling. It achieves this by replacing traditional Fourier bands with graph-Laplacian frequency bands, derived from graph operators that reflect mesh connectivity and geometry. The researchers also introduce scalable approximations, such as Chebyshev polynomial graph filters to avoid costly eigendecomposition, and Graph Laplacian Energy Alignment for Meshes (GLEAM) for multilevel graph architectures. These innovations significantly improve long-horizon rollout fidelity and maintain statistical invariants when forecasting turbulent flows on unstructured meshes, outperforming deterministic baselines.

Why it matters

Professionals in fields relying on complex simulations can achieve more accurate and stable long-term predictions for chaotic systems, especially those involving irregular geometries, leading to better design and operational decisions.

How to implement this in your domain

  1. 1Explore integrating graph-Laplacian based spectral losses into existing simulation and surrogate modeling frameworks for complex physical systems.
  2. 2Investigate the use of Chebyshev polynomial graph filters as a scalable alternative for spectral decomposition in large-scale simulations.
  3. 3Apply the GLEAM approach in multilevel graph architectures to improve regularization and accuracy across different scales in simulations.
  4. 4Collaborate with research teams to adapt these advanced spectral loss techniques for specific industry applications involving unstructured meshes.

Who benefits

AerospaceAutomotiveClimate ScienceEnergyManufacturing

Key takeaways

  • New spectral loss methods enable scale-aware learning for chaotic systems on unstructured meshes.
  • Graph-Laplacian frequency bands extend Fourier-like analysis to irregular geometries.
  • Scalable approximations like GLEAM improve long-horizon forecasting fidelity.
  • This research enhances the accuracy and stability of surrogate models for turbulent flows.

Original post by Kanad Sen, Romit Maulik

"arXiv:2607.19387v1 Announce Type: cross Abstract: Surrogate modeling for high-dimensional nonlinear dynamical systems that exhibit chaos requires mechanisms that preserve not only pointwise accuracy but also the scale-dependent structure of physical fields. Bandwise spectral powe…"

View on X

Originally posted by Kanad Sen, Romit Maulik on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses