22%
lower error (RMSE) than a no-change forecast on 10 m east-west wind (u10), 2020 test year · 1.17 against 1.49 m/s, latitude-weighted
results file
Four world maps of north–south surface wind for one 2020 test case: the input, the 6-hour forecast, the true field six hours later, and the forecast minus the truth. The forecast keeps the large-scale patterns; the errors are small and scattered.
The first 2020 test case, north–south wind at 10 m: the input, the 6-hour forecast, the truth and the error, in standardised units. Wind is where the model gains most over a no-change forecast.original

Evaluation

35%
lower error (RMSE) than a no-change forecast on 10 m north-south wind (v10) · 1.18 against 1.81 m/s, latitude-weighted
results file

Tested on all 1,463 six-hour steps of 2020, a year held out from training, against two baselines: persistence, which assumes nothing changes in six hours, and anomaly persistence, which carries the current departure from the seasonal average forward. Scores are latitude-weighted and in physical units. On 10 m wind the model beats both. On 850 hPa temperature and 500 hPa geopotential it does not beat anomaly persistence: 1.24 K against 1.02 K, and 213 against 191 m²/s². Over India it is level with persistence on wind and behind it on temperature and geopotential.

Error by field, 2020 test year, physical units, latitude-weighted; the pooled row is in standardised units
fieldmodel RMSEvs a no-change forecast
850 hPa temperature1.24 K3.4% worse
500 hPa geopotential213 m²/s²3.5% better
10 m wind, east–west1.17 m/s21.8% better
10 m wind, north–south1.18 m/s34.5% better
All four, pooled (standardised)0.18528.7% better
Error by field, 2020 test year, physical units, latitude-weighted; the pooled row is in standardised unitsresults file

How it works

A vision transformer with 4.8 million parameters that forecasts four ERA5 fields six hours ahead, with attention written by hand instead of nn.MultiheadAttention. Its loss adds two penalties to the error: one for jagged fields and one for a wrong global mean.

Built block by block

Every block implemented and unit-tested on its own: patch embedding, positional encoding, attention, transformer block and loss. A test checks that nn.MultiheadAttention is never used.

Stack

  • PyTorch
  • xarray
  • zarr
  • ERA5 / WeatherBench2

129 unit teststest suite · Nov 2026

Contact / Rourkela, Odisha, India

Write to me