Reproduce “On the Effectiveness of Neural Operators at Zero-Shot Weather Downscaling” on OpenResearch

About this paper

Saumya Sinha, Brandon Benton, Patrick Emami

An evaluation from NREL investigates the effectiveness of neural operators for zero-shot weather downscaling, demonstrating that non-operator deep learning models like SwinIR achieve better average error metrics, while ESRGAN is more adept at preserving kinetic energy spectra, suggesting current models struggle to infer fine-scale physics at unobserved resolutions.

Key insights

• Neural operators, despite their resolution invariance, did not consistently outperform non-operator deep learning models like SwinIR and ESRGAN in zero-shot weather downscaling, challenging previous assumptions about their broad applicability.
• Effective spatial feature extraction (e.g., via RRDB blocks) is critical for neural operators in downscaling, indicating that overall architectural design is as important as the operator layers themselves.
• All tested data-driven models faced a fundamental limitation in zero-shot scenarios, consistently underestimating energy content at the highest wavenumbers, which suggests difficulty in inferring unobserved fine-scale physical phenomena (like turbulence).

Summary by alphaXiv.