Geophysical Research Letters· 2026Q1
Probing Atmospheric Gravity‐Wave‐Relevant Representations in a Weather Foundation Model via Multi‐Method Fine‐Tuning
- 1citations
- Q1SCImago
- 2026year
Short summary
A weather foundation model (Aurora) shows pretraining benefits for predicting gravity-wave momentum fluxes, particularly zonal flux (38.9% vs. 41.6% normalized error with FISH + TADP vs. random control), with a 1% parameter fine-tuning recovering significant skill.
AI-generated from the title and abstract; the full text is not read.
Key points
- Pretraining in the Aurora weather foundation model shows a benefit for predicting gravity-wave momentum fluxes.
- The zonal flux prediction improved with parameter-efficient fine-tuning methods (e.g., FISH + TADP) compared to random initialization (38.9% vs. 41.6% normalized error).
- Meridional flux prediction remained difficult across all methods (60%-80% NRMSE).
- The pretraining benefit was level-dependent, with significant gains at 200-300 hPa but minimal gains at 50 hPa.
- Fine-tuning approximately 1% of parameters recovered a significant portion of the skill seen with full fine-tuning, indicating regularization from the frozen backbone.
AI-generated from the title and abstract; the full text is not read.
Abstract
Abstract Weather foundation models pretrained on reanalysis data may encode subgrid‐scale physics beyond their forecasting objectives. Using Aurora, we probe this with a multi‐method fine‐tuning framework: agreement across structurally different adaptation strategies signals shared backbone information, while a matched randomly initialized control isolates pretraining's contribution. We evaluate six methods (0.5%–100% of trainable parameters) for predicting ERA5‐resolved, T21‐filtered gravity‐wave momentum fluxes (313–1,820 km). Evidence for a pretraining benefit is strongest for the zonal flux : several parameter‐efficient methods outperform random initialization, most clearly FISH + TADP (38.9% vs. 41.6% normalized error, = 0.007). The meridional flux remains hard for all methods (NRMSE 60%–80%). The benefit is level‐dependent: 11–14 percentage points at 200–300 hPa but only 0.2 at 50 hPa. Approximately 1% of parameters recover part of full‐fine‐tuning skill (38.9% vs. 31.8%), indicating regularization by the frozen backbone.
The authors' abstract, as published at the source. Geophysical Research Letters, 2026 · DOI ↗
Continue with a free account
Ask the paper: 3 free questions a day about this paper; save it, get its citation, new summaries every day for your field. Takeaways are Premium.
Continue free on the webSign in with Google or Apple; no card needed. You come back to this paper.
On your phone:
Field: Atmospheric Science
Atmospheric ScienceEarth and Planetary Sciences