Error in ERA5 2m Temperature identified using GraphCast
Hannah M. Christensen, Jack Barker, Bobby Antonio, Massimo Bonavita, Mohamed Dahoui, Patricia de Rosnay · 2026
Read Error in ERA5 2m Temperature identified using GraphCast by Hannah M. Christensen, Jack Barker, Bobby Antonio, Massimo Bonavita, Mohamed Dahoui, Patricia de Rosnay (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-04701.txt.
Description
Hannah M. Christensen and five co-authors caught a flagship weather dataset making the same mistake for decades, by racing it against a machine learning model, in a 2026 arXiv paper. ERA5, ECMWF's great reanalysis, now doubles as training and verification data for learned forecast models. Comparing six-hour GraphCast forecasts against ERA5 surfaced a recurrent error in two-metre temperature over the Ethiopian Highlands, appearing mostly at one time of day.
The forecast, they show, was right and the reanalysis wrong. The culprit is the interpolation scheme that assimilates surface reports. When reports arrive displaced in time from the background forecast, it writes spuriously warm corrections over Ethiopia on roughly seven percent of dates at that hour. the same flaw sits in ECMWF's operational analysis. GraphCast mostly learned to ignore the phantom warmth. The authors recommend fixes and warn against treating reanalysis as unquestioned truth.
Catalog record
| CREATOR | Hannah M. Christensen, Jack Barker, Bobby Antonio, Massimo Bonavita, Mohamed Dahoui, Patricia de Rosnay |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 38 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-04701