ml-weather-models technique sourced
Machine-learned weather forecasting
Neural networks trained on decades of atmospheric reanalysis forecast global weather 10-15 days ahead in seconds to minutes on one accelerator, matching or beating physics-based operational models on most verification targets.
mechanism
Models take the current and recent atmospheric state on a global 0.25 degree latitude-longitude grid (temperature, wind, humidity, pressure at many levels plus surface variables) and predict the state one fixed step later; rolling the prediction forward gives a 10-15 day forecast. GraphCast maps grid cells onto a multi-resolution icosahedral mesh and runs learned message passing there. GenCast is a conditional diffusion model that samples stochastic trajectories in 12-hour steps, so repeated samples form an ensemble with calibrated spread. FGN (2025) builds ensembles from learned model perturbations trained on CRPS. Aurora is a 1.3B-parameter foundation model pretrained on over a million hours of geophysical data and fine-tuned for air quality, ocean waves, cyclone tracks and 0.1 degree weather. Inference replaces hours on a supercomputer with seconds to minutes on one accelerator.
limitations
- Trained on Earth's reanalysis record, so they reproduce Earth's atmosphere only; fictional planets, altered geography or deliberate climate change lie outside the training data.
- Grids are coarse for game maps: 0.25 degree (about 28 km at the equator), and operational AIFS runs at 28 km.
- Inference needs an accelerator and gigabytes of weights (Aurora has 1.3B parameters), heavy for client-side use.
- Designers cannot author dramatic timing; the model follows the statistics of real weather, so pacing must be imposed by selecting or conditioning forecasts.
- Autoregressive rollouts are validated to 10-15 days; behavior over seasons or years is not established by these sources.
appears in
the paper, 4 sections · 1 chapter · 1 episoderelations
6 stored edges
Forecasting
Ensembles (50 GenCast members) give calibrated probabilities, so an in-game forecast can carry real uncertainty that players plan against and that occasionally misses honestly.
Planetary climate
Earth-system foundation models fine-tuned to several domains suggest climate sims coupling atmosphere, ocean waves and air chemistry, though they are trained for days-ahead weather, not decades.
Pollution spread
Aurora's fine-tuned air-quality forecasts run about 100,000x faster than CAMS, a path to pollution plumes transported by simulated wind at playable cost.
Random disasters
Storms emerge from atmospheric state with forecastable tracks (Aurora beat operational centres on 5-day cyclone tracks), turning disasters from dice rolls into threats with lead time.
Weather
Weather becomes a coherent global field with fronts, cyclones and multi-day evolution across hundreds of variables, instead of a per-region Markov chain of weather states.
backlinks
computed from other recordsclaims
5 with sourcesGraphCast produced 0.25 degree, 10-day global forecasts in under one minute on one TPU v4 machine and beat ECMWF HRES on over 90% of 1,380 verification targets.
GenCast's 50-member diffusion ensembles were more skillful than ECMWF's ENS on 97.2% of 1,320 targets; one 15-day forecast took about 8 minutes on a Cloud TPU v5.
Aurora (1.3B parameters) matched or beat operational systems on 74% of air-quality targets, 86% of ocean-wave variables and 92% of 0.1 degree weather targets, beat seven centres on 5-day cyclone tracks, and ran about 100,000x faster than CAMS for air quality.
ECMWF made its AIFS machine-learned model operational on 25 February 2025, reporting about 1,000x less energy per forecast than its physics model, at 28 km grid spacing versus 9 km.
FGN (2025) reported state-of-the-art ensemble skill and cyclone tracks while being trained only on per-location targets.