Google DeepMind has opened a new path for increasing the accuracy of weather forecasting by releasing 'PARDIG' and 'SatWeave,' the vast meteorological datasets that form the foundation of their weather prediction AI, 'WeatherNext 3.'
Imagine this: You’ve planned an outdoor camping trip for the weekend and are excitedly packing, expecting a clear, cloudless day. But just before you leave, your AI assistant tells you very accurately, “At 2:00 PM, there is a high probability of a 15-minute shower near the campsite.” Thanks to this, you remembered to pack an umbrella and were able to avoid the sudden rain. How is such ‘precise weather forecasting’ possible?
Recently, Google DeepMind made a very important decision in the field of weather forecasting. They have released to the world the massive meteorological data that acts as the core heart powering their cutting-edge weather AI, ‘WeatherNext 3’ [Source 1].
Why is the release of weather forecasting data important?
Weather affects everything in our daily lives, from agriculture, aviation, and travel plans to deciding what clothes we wear each morning. Until now, predicting the weather was a very difficult domain that required mobilizing the world’s fastest supercomputers to calculate complex physical laws [Source 14]. However, with the emergence of AI, the landscape of forecasting is changing completely.
The fact that Google DeepMind has released this valuable data means that countless researchers around the world can now use it to jointly study more accurate and faster weather forecasting models. This signifies that AI meteorological technology is evolving as a ‘public asset’ that protects everyone’s safety and responds to climate change, rather than being a technology exclusive to a specific company.
Simply put: How does AI learn the weather?
The process by which AI predicts the weather can be compared to the principle of an experienced ‘seasoned fisherman’ looking at the sky to guess tomorrow’s weather.
An fisherman observes cloud shapes and wind directions tens of thousands of times in his lifetime, accumulating in his mind a pattern that says, “When clouds look like this and the wind blows like that, it rains.” AI is the same. In meteorology, this is called a ‘statistical model,’ and AI analyzes data by learning from weather records accumulated over decades and physical phenomena that actually occurred [Source 15].
The dataset Google released this time is an excellent ‘textbook’ for AI to learn from.
- PARDIG: This is data that meticulously records global precipitation in 15-minute increments [Source 2]. It shows the flow of rain starting and stopping in very fine detail.
- SatWeave: This is a mosaic map made by stitching together satellite photos taken every hour [Source 2]. It is essentially a video where you can scan the movement of clouds across the entire Earth by time period.
AI studies these textbooks repeatedly and realizes for itself, “When clouds look like this right now, the probability of rain in 15 minutes is X%.”
Where are we now: The amazing capabilities of ‘WeatherNext 3’
Currently, Google DeepMind’s latest model, ‘WeatherNext 3,’ is demonstrating technological prowess far ahead of existing weather forecasting systems [Source 8].
Its biggest feature is that it generates new forecasts ‘every hour’ [Source 8]. While existing systems updated forecasts only a few times a day at specific times, WeatherNext 3 tracks the Earth’s weather, which changes in real-time 24 hours a day, without rest. This is a huge strength in detecting and responding to sudden, unexpected weather changes.
What can we expect in the future?
Weather forecasting is no longer just a matter of ‘luck.’ Looking at past data, AI is going beyond the level of simply guessing tomorrow to capturing complex and subtle atmospheric flows that humans find difficult to identify.
In the future, researchers around the world will use the released PARDIG and SatWeave data to create more diverse and smarter models in their respective fields. Perhaps one day, a time will come when we can check the weather at our exact location down to the minute on our smartphone apps. As data becomes more transparent and collaborative research becomes more active, our skies will be forecasted more clearly and safely.
MindTickleBytes’ AI Reporter’s Perspective The way AI predicts the weather is very similar to how humans build experience. However, unlike humans, AI can scan decades of global records in just a few seconds. Now that Google has opened this vast meteorological textbook to everyone, we have entered a new era where all of humanity can piece together the massive puzzle of climate much faster.
References
- Opening the datasets behind our AI weather models
- PARDIG and SatWeave: DeepMindweatherdatasets
-
[AIWeatherForecasting Can’t Replace Humans… Scientific American](https://www.scientificamerican.com/article/ai-weather-forecasting-cant-replace-humans-yet/) - WeatherNext 3 isourmost advanced globalweatherAImodel.
- AIpredicted Cyclone Alfred’s path and could be the future of…
- WeatherNext 3
- PARDIG and SatWeave
- GenCast and Gemma
- Direct calculation by supercomputers
- Pattern analysis of weather records accumulated over decades
- Intuitive judgment of human meteorologists
- Generates forecasts only once a day
- A global weather AI that generates forecasts every hour
- A model that only uses satellite data