GraphCast faithfully reproduces ERA5 near-surface wind speed at Antarctic stations - but ERA5 itself deviates substantially from in-situ observations: a preliminary three-way evaluation

Standard verification approaches for AI weather prediction (AIWP) models often rely on scores derived from data-dense regions, while verification across the polar areas remains limited. Surface-level performance in data-sparse regions demands targeted evaluation using in-situ observations. We address this gap by evaluating GraphCast-predicted near-surface wind speeds against both ERA5 and in-situ observations at Antarctic stations. This study applies a three-way comparison between observations, reanalysis, and AIWP predictions to Antarctic stations, where such evaluation is currently absent; and extends the ERA5 vs in-situ observations (Obs) Antarctic wind evaluation by Rakoczy et al. [1].

We compare monthly mean 10 m wind speed from GraphCast T+24h forecasts, ERA5 reanalysis and SCAR Met READER station observations [2] at four Antarctic low-altitude stations for 2018–2020. Near-surface winds play a key role in Antarctic climate, influencing sea ice formation, precipitation, boundary layer stability and ice shelf dynamics [3]. Accurate representation in both reanalysis and forecast systems is therefore relevant to climate research as well as operational applications. Katabatic winds, channelled by ice sheet topography at scales below the model grid, may represent a particular challenge for AIWP and NWP models.

The three comparisons show a pattern. ERA5 vs Obs errors are substantial at all four stations, indicating that ERA5 shows substantial biases against observations. GraphCast vs Obs errors are systematically larger than ERA5 vs Obs at three of four stations, suggesting GraphCast may amplify ERA5's errors when evaluated against reality. GraphCast vs ERA5 errors are small at all Antarctic stations, indicating GraphCast closely replicates ERA5 patterns. The combination of these three comparisons is a noticeable preliminary finding: GraphCast faithfully reproduces ERA5 at Antarctic stations, but ERA5 deviates substantially from observations.

The station sample is too small to support strong conclusions, but these preliminary findings suggest that in-situ polar observations can expose verification blind spots not captured by reanalysis-based AIWP benchmarks, and that the fidelity of an AIWP model to ERA5 data is different from its ability to accurately predict in-situ observations. This study is ongoing; extensions to additional stations, longer records, and further AIWP models are planned prior to the conference.

—

[1] Rakoczy, B. C., Bromwich, D. H., & Wang, S. (2026). Evaluation of ERA5 Near-Surface Winds Over Antarctica: Spatial Variability, Biases, and Large-Scale Influences. Journal of Climate (published online ahead of print 2026), e250215, Article e250215. https://doi.org/10.1175/JCLI-D-25-0215.1

[2] https://legacy.bas.ac.uk/met/READER/data.html (accessed 25.03.2026) https://doi.org/10.5285/569d53fb-9b90-47a6-b3ca-26306e696706 

[3] Davrinche, C., Orsi, A., Amory, C., Kittel, C., and Agosta, C. (2025). Future changes in Antarctic near-surface winds: regional variability and key drivers under a high-emission scenario, The Cryosphere, 19, 6023–6042, https://doi.org/10.5194/tc-19-6023-2025