Accurate weather forecasts are crucial for our daily lives, from helping us plan our activities to protecting us from hazardous weather conditions. As the climate changes, the demand for reliable and timely weather information will only increase. That’s why Google has been investing in weather and climate research, and we are excited to announce our latest innovation: Scalable Ensemble Envelope Diffusion Sampler (SEEDS).
SEEDS is an AI that makes weather forecasts fast and cheaply. It costs less than traditional weather prediction methods. This technology is key for researching weather and climate. It is among the first uses of probabilistic diffusion models in this field.
The Need for Probabilistic Forecasts
In 1972, Ed Lorenz, a meteorology professor at MIT, presented a talk. It was titled “Does the Flap of a Butterfly’s Wings in Brazil Set Off a Tornado in Texas?” He gave this talk at the American Association for the Advancement of Science meeting. This introduced the term “butterfly effect.” It means small initial changes can greatly affect future outcomes. Lorenz had already explored this idea in a 1963 paper. There, he explained how small errors grow exponentially over time in weather models.
This growth of errors is called chaos. It makes predicting the weather, especially extreme events, very difficult.
Weather agencies globally offer forecasts that show different potential outcomes. They create these forecasts by mixing regular predictions with changes. These changes reflect the real-world uncertainties in weather and physical processes. By merging these predictions, they reduce mistakes and estimate the weather’s uncertainty.
SEEDS: A Game-Changing Technology
Traditional methods for generating ensembles of weather forecasts rely on physics-based simulations, which can be computationally costly. This makes it challenging to generate a large enough ensemble to accurately predict rare and extreme weather events. SEEDS overcomes this limitation by using generative AI models to efficiently produce ensembles at scale.
SEEDS is a big advancement in weather forecasting technology. It is now part of Google’s advanced forecast model, MetNet-3. This model forecasts the weather up to 24 hours ahead. SEEDS has also helped create GraphCast. This is a weather model that predicts up to 10 days into the future.
The Future of Weather Forecasting
Google is working to improve weather forecasts. They use SEEDS and other advanced technologies. Their goal is to offer reliable information to help people make decisions. They want to create a society better prepared for severe weather conditions. Google plans to achieve this by continuing to invest in these efforts.
Posted by Lizao (Larry) Li, Software Engineer, and Rob Carver, Research Scientist at Google Research
