Scalability of power-law scheduled score-based diffusion and Lazy Diffusion to 3D turbulence, anisotropic spectra, and coupled multiphysics

Determine the scalability of score-based conditional diffusion models equipped with power-law noise schedules and the Lazy Diffusion one-step distillation to fully three-dimensional turbulent flows, to systems exhibiting anisotropic energy spectra, and to coupled multiphysics dynamical systems, beyond the moderate-resolution two-dimensional settings evaluated in this work.

Background

The paper analyzes spectral collapse in standard DDPM forward processes for turbulent power-law spectra and introduces physics-aware modifications: power-law noise schedules to preserve high-wavenumber coherence and a Lazy Diffusion distillation to enable single-step conditional generation. These approaches showed improved spectral fidelity and long-horizon autoregressive stability on two systems: 2D forced Kolmogorov turbulence and regional ocean surface reanalysis.

The authors note that the current demonstrations are at moderate resolutions and in two-dimensional domains. They explicitly state that questions remain about scaling these methods to more complex regimes, specifically fully three-dimensional turbulence, anisotropic spectra, and coupled multiphysics systems, which would test the generality and robustness of the proposed diffusion-based generative framework.

References

Moreover, the current models operate at moderate resolution and two-dimensional domains, leaving open questions regarding scal- ability to fully three-dimensional turbulence, anisotropic spectra, and coupled multiphysics systems.

— Lazy Diffusion: Mitigating spectral collapse in generative diffusion-based stable autoregressive emulation of turbulent flows  (2512.09572 - Sambamurthy et al., 10 Dec 2025) in Section 5 (Conclusion)

Scale poses a different unanswered question. The relative gain over the baselines was largest on the largest geometry evaluated, but our experiments have not yet reached million-point industrial meshes or coupled multiphysics systems. These settings are the natural next test of whether the observed advantage persists.

— Beyond Pairwise Graphs in Science: Hypergraph Adaptive Wavelet Operators for Parametric PDEs  (2608.27883 - Sarkar et al., 28 Aug 2026) in Section 5, subsection “Limitations and Outlook”

Future work should test whether the lower \ac{CFM} volume-fraction errors observed here persist in 3D settings, multi-physics design tasks, and higher-dimensional condition spaces.

— Conditional Flow Matching for ML-Based Inverse Design Problems  (2609.00863 - Felder et al., 1 Sep 2026) in Section 4, Discussion, paragraph “Limitations and future work”