AI and High‑Performance Computing in Weather Forecasting: From Radar Images to Neural Networks

On August 12, Beijing experienced the strongest rain of the season, prompting public interest in how weather forecasts are made and how artificial intelligence is now being integrated into meteorology.

Traditional forecasting relies on a network of ground stations, Doppler radars, and weather satellites that collect temperature, humidity, pressure, and other atmospheric data, which are then visualized on multi‑layered maps.

Modern meteorology is highly interdisciplinary, requiring physics to explain atmospheric and oceanic motions, chemistry to understand material changes, and mathematics for statistical analysis, reflecting a nation’s scientific and computational capabilities.

Google’s recent paper on precipitation nowcasting demonstrates a data‑driven approach that abandons explicit atmospheric physics, using a U‑Net convolutional neural network to learn a direct image‑to‑image mapping from radar snapshots to short‑term rain forecasts.

Image‑recognition techniques extract hue, saturation, and brightness features from radar images to differentiate weather phenomena such as rain, snow, hail, fog, and mist.

Optical‑flow (OF) algorithms, originally developed in the 1940s for computer vision, are applied to sequential radar panels to model motion vectors and predict short‑term weather evolution, though they may miss intensity decay.

IBM operates the Global High‑Resolution Atmospheric Forecast System (GRAF), the world’s highest‑resolution global weather model, updating hourly and running on 84 AC922 nodes each equipped with four Nvidia V100 GPUs and 3.5 PB of storage, processing up to 10 TB of weather data per day.

Despite these advances, AI faces limitations because weather is influenced by thousands of variables—solar radiation, ocean currents, and more—requiring ever‑greater data volumes and computational power, and predictions for extreme events like Beijing’s heavy rain still contain uncertainties.

The article concludes that AI accelerates meteorological research and public understanding, yet substantial challenges remain before AI can fully match the complexity of atmospheric systems.

References: - iScientist: "Why Weather Forecasts Can Be Inaccurate" - Google: "Using Machine Learning to Nowcast Precipitation in High Resolution" - Machine Heart: "AI‑Empowered Precise Weather Forecasting" - IBM GRAF website