Observation Assimilation Transformer: a network corrects weather forecasts from thousands of stations, up to a resolution of 1 km

This probabilistic transformation refines a numerical prediction through scattered observations and, downhill from ERA5, does better than the COSMO model on temperature and rain.

Combining numerical weather prediction with actual station measurements is one of the pillars of meteorology, but conventional assimilation methods are cumbersome. Researchers present the Observation Assimilation Transformer (OAT), a neural process that takes "a priori" forecasts and scattered observations as input, and then produces a high-resolution corrected field with a probability distribution rather than a single value.

No grid, no absolute coordinates

The model relies on sparse attention that depends only on relative positions between points, allowing it to jump to thousands of stations, tolerate missing observations, and produce values at any location, without projecting onto a grid. It is trained by randomly drawing the stations used, with a loss function based on the CRPS score.

Tested on a regional set of a few hundred stations and a continental set of several thousand, OAT exceeds several reference methods and provides well-calibrated predictions. In a case study, it downscales the ERA5 reanalysis (about 31 km) to the scale of the COSMO model (1 km) and obtains a lower mean absolute error and CRPS than COSMO itself on temperature and precipitation at the test stations.