14 papers in the archive · tier-1 watchlist.
Aurora is a large Earth-system foundation model trained on over a million hours of data that beats operational systems for air quality, waves, cyclone tracks and high-resolution weather.
Foundation modelsThe Alan Turing InstituteMicrosoft Research (AI for Science)Silurian AIUniversity of Cambridge (Aardvark)
End-to-end model ingests raw observations to produce global and station forecasts, beating operational NWP on several variables using 8% of input data.
Data assimilationThe Alan Turing InstituteECMWFGoogle DeepMindMicrosoft Research (AI for Science)University of Cambridge (Aardvark)
Sim2Real pre-trains a ConvCNP on reanalysis and fine-tunes on German weather stations, outperforming reanalysis-only or station-only training for surface air temperature interpolation.
Observations & remote sensingMicrosoft Research (AI for Science)University of Cambridge (Aardvark)
Vine copulas model dependence among weather variables at three African sites, proposed as a dependency layer for subseasonal forecast calibration.
Post-processing & calibrationMicrosoft Research (AI for Science)
Probabilistic structural forecasts simulate geostationary IR radial profiles with PixelSNAIL and feed a CNN, giving intensity errors near NHC official forecasts.
ClimaX is a transformer foundation model pretrained on CMIP6 that fine-tunes to many weather and climate tasks, outperforming data-driven baselines.
Foundation modelsMicrosoft Research (AI for Science)University of California, Los Angeles (UCLA)code
Machine-learning adaptive bias correction of ECMWF subseasonal forecasts raises contiguous-US temperature skill 60-90% and precipitation skill 40-69%.
Post-processing & calibrationMicrosoft Research (AI for Science)Massachusetts Institute of Technology (MIT)
ANFIS models forecast typhoon rainfall at a Tamsui station up to three hours ahead and are packaged in a mobile app, with best RMSE of 7.9 mm.
Local & station forecastingMicrosoft Research (AI for Science)
Microsoft's operational ConvLSTM radar nowcasting model, conditioned on HRRR, beats optical flow and HRRR by 20–25% for lead times up to 6 hours.
A CNN-based cubed-sphere ensemble of 32 deep-learning weather models gives 2-6 week forecasts that are only modestly worse than the ECMWF S2S ensemble over land.
Sub-seasonal to seasonalMicrosoft Research (AI for Science)University of Washington
CNN on a cubed-sphere grid with multi-step loss gives stable global forecasts for weeks, beating persistence, climatology and coarse NWP at short-to-medium range.
Global medium-range NWPMicrosoft Research (AI for Science)University of Washington
CNNs forecast 500-hPa geopotential height up to 14 days, beating persistence and barotropic vorticity model but not the operational NWP model.
Global medium-range NWPMicrosoft Research (AI for Science)University of Washington
Opinion article arguing machine learning should be adopted for S2S forecasting, showing statistical methods beat dynamical models on a polar vortex case.
Sub-seasonal to seasonalMicrosoft Research (AI for Science)Massachusetts Institute of Technology (MIT)Potsdam Institute for Climate Impact Research (PIK)Stanford University
Ensemble of regression models with the SubseasonalRodeo dataset beats the top Rodeo entry and improves debiased CFSv2 precipitation skill by 129-169%.
Sub-seasonal to seasonalMicrosoft Research (AI for Science)Stanford University
