Determine whether the Aurora prior causes the noise floor

Determine whether the persistent high-wavenumber noise floor in Xaurora forecasts originates from the pretrained Aurora deterministic foundation-model prior.

Background

The paper identifies a persistent noise floor in Xaurora’s forecast spectra: beyond a variable-dependent cutoff wavenumber, the power spectral density becomes flat and predictable fine-scale structure is not recovered. Experiments indicate that the phenomenon is not explained by insufficient denoiser training, the SDE solver, the generative parameterisation, or the finetuning method.

Additional comparisons show that both a denoiser trained from scratch and a model initialized from Aurora-L exhibit similar noise-floor behavior. However, the available experiments do not establish whether the phenomenon is inherited from the pretrained Aurora prior, leaving its causal origin unresolved.

References

Whether this comes from the Aurora prior itself cannot be settled by the current analysis.

— Xaurora: Generative Weather Forecasting with Denoising Stochastic Interpolants from a Foundation Model Prior  (2610.06509 - Walt et al., 5 Oct 2026) in Appendix, Section Supplementary results, subsection “Noise floor investigation,” subsubsection “Sufficient denoiser training”