GenCast: How DeepMind Built a Weather Forecaster That Outperforms Supercomputers

For more than forty years, weather forecasting has depended on the same fundamental approach: divide the atmosphere into a three-dimensional grid, apply the laws of physics at every point, and let a supercomputer crunch the equations forward through time. The European Centre for Medium-Range Weather Forecasts (ECMWF) has refined this method to an art, operating one of the most powerful weather prediction systems on the planet. Its Integrated Forecasting System is widely considered the gold standard.

Then Google DeepMind published a paper in Nature in December 2024 showing that a machine learning model called GenCast could beat ECMWF on 97.2% of 1,320 evaluation targets. The model produces a full 15-day global forecast in about eight minutes, running on a single Google TPU v5 chip. The conventional system requires hours on a dedicated supercomputer cluster consuming megawatts of energy.

What Makes GenCast Different

GenCast is a diffusion model, similar in spirit to the AI systems that generate images from text prompts. But instead of producing photographs, it generates plausible future states of the atmosphere. The model was trained on 40 years of ERA5 reanalysis data from ECMWF itself, learning to capture the complex, chaotic patterns of global weather at a resolution of 0.25 degrees latitude and longitude, roughly 28 kilometers at the equator.

What sets GenCast apart from earlier AI weather models like Pangu-Weather and GraphCast (also from DeepMind) is its probabilistic nature. Rather than producing a single deterministic forecast, GenCast generates an ensemble of possible outcomes. This matters enormously in practice because weather is inherently uncertain. A forecast that says "there is a 70% chance of rain" is far more useful to decision makers than one that simply says "rain" or "no rain."

In head-to-head tests, GenCast produced better calibrated probability estimates for temperature, wind speed, and pressure across the globe. It was especially strong for extreme weather events and for longer lead times, where traditional models tend to lose accuracy fastest.

The Tropical Cyclone Test

One of the most striking results came from tropical cyclone track prediction. GenCast outperformed the ECMWF ensemble on predicting cyclone trajectories across all lead times from one to five days. Getting cyclone tracks right is not an academic exercise. Every additional hour of accurate warning translates directly into lives saved and evacuations completed in time.

The model also showed particular skill in predicting wind power output across Europe, generating forecasts that were more accurate than those from the ECMWF ensemble. For energy grid operators trying to balance supply and demand, better wind forecasts can mean millions of dollars saved and fewer fossil fuel plants held in reserve.

Open Source, Open Science

Perhaps the most consequential decision DeepMind made was to open source the model. The full code, model weights, and training pipeline are available on GitHub. This means national meteorological agencies, academic researchers, and even smaller countries that cannot afford supercomputer infrastructure can experiment with state-of-the-art weather prediction.

This is significant because weather forecasting capacity is deeply unequal around the world. While wealthy nations operate sophisticated numerical weather prediction systems, many developing countries rely on coarser models or delayed data. An AI model that runs on a single chip could, in principle, democratize access to high-quality forecasts. The World Meteorological Organization has estimated that early warning systems could reduce disaster deaths by up to 30% in the most vulnerable countries.

What It Does Not Replace

GenCast is not a replacement for traditional meteorology. It cannot explain why it makes a given prediction. It does not model the physics of the atmosphere explicitly, which means it may struggle with truly unprecedented weather patterns that fall outside its training data. ECMWF itself has noted that AI models and physics-based models will likely work best in combination, with each compensating for the other's weaknesses.

There are also resolution limits. At 28 kilometers, GenCast cannot capture the fine-scale phenomena that matter for local forecasts: thunderstorm cells, sea breezes, or mountain valley winds. For those, regional models with resolutions of a few kilometers or less will remain essential.

The Bigger Picture

GenCast is part of a broader transformation in weather science. Huawei's Pangu-Weather, NVIDIA's FourCastNet, and several other AI systems have all shown competitive results against traditional models in recent years. What makes GenCast notable is the comprehensiveness of its evaluation and the strength of its probabilistic forecasting.

In a world where climate change is making weather more extreme and less predictable, better forecasting is not a luxury. It is critical infrastructure. The difference between a three-day warning and a seven-day warning for a major storm can be the difference between an orderly evacuation and a catastrophe. GenCast brings that extended accuracy within reach, at a fraction of the computational cost.

Eight minutes. One chip. 97.2% of targets beaten. The numbers speak for themselves.