Google DeepMind's 'GenCast' is a probabilistic AI weather forecasting model that outperforms existing systems in accuracy for 15-day forecasts in over 97% of test cases.
“What will the weather be like for my vacation in two weeks?” I asked the AI.
Imagine this. You are planning an important outdoor wedding or a family trip two weeks from now. You open your smartphone weather app with excitement, and while it shows a ‘15-day forecast,’ we usually tilt our heads and say, “Well, how can anyone know the weather 15 days out accurately? Let’s just take it with a grain of salt.”
In fact, it has been common sense that weather forecasts lose accuracy rapidly, like a crumbling sandcastle, as the time horizon extends. But that common sense is now starting to break in a good way, thanks to GenCast, a new weather forecasting AI announced by Google DeepMind.
GenCast goes beyond just whispering “it will rain tomorrow”; it predicts detailed weather 15 days in advance, as well as dangerous weather anomalies like heatwaves and strong winds, with incredible accuracy Source 5. Today, we’ll explain how this smart AI weather forecaster looks into the Earth’s future and how it will change our daily lives in an easy-to-understand way.
Why is this important?
Weather isn’t just trivial information for deciding ‘whether to bring an umbrella today.’ Sudden heatwaves, bitter cold snaps, and winds strong enough to swallow houses cause numerous casualties and astronomical economic losses every year. If we could accurately know about such extreme weather 15 days in advance, how would the world change?
Put simply, local governments could check levees thoroughly long before a flood occurs, and farmers could take measures to protect crops from forecasted cold snaps Source 3. According to Google DeepMind’s research, GenCast showed overwhelming performance compared to any existing system, especially in predicting these ‘dangerous weather’ events Source 4. GenCast is essentially a reliable sentinel that earns us precious ‘golden time’ to prepare for disasters.
Easy Understanding: How does GenCast predict the weather?
If traditional weather forecasting systems were ‘meticulous mathematicians’ solving giant physical formulas with supercomputers, GenCast is like a ‘veteran captain’ who has personally experienced the expressions of the sea and sky for decades and kept them all in mind.
1. An AI that learned 40 years of Earth’s memories
GenCast uses machine learning techniques (where computers learn by themselves through data). This AI has studied every detail of vast ‘reanalysis data’ that records what the Earth’s weather has been like over the past 40 years Source 2. It has mastered complex and subtle patterns, such as “when these types of clouds form in this way, a typhoon always follows 10 days later.”
2. Reviewing “50 possibilities” instead of “one right answer”
This is where GenCast’s real magic begins. Traditional weather models were often ‘deterministic models’ that gave only one result, such as “the temperature 10 days from now will be exactly 25 degrees” Source 1.
However, weather is fickle and can change with a single flap of a butterfly’s wings. So, GenCast chose the probabilistic ensemble approach. Here is an analogy:
[Analogy] When predicting which horse will win a race, instead of trusting only one expert, it’s like asking 50 experts with different perspectives for their opinions. GenCast instantly creates more than 50 different ‘future scenarios’ at once Source 8. If 45 of them say “it will rain,” we can be certain to prepare an umbrella based on that 90% probability.
3. Diffusion Model: Finding clear days in the noise
GenCast also utilizes the Diffusion Model, a state-of-the-art generative AI technology Source 3. This is the same technology used when we draw sophisticated pictures with AIs like ‘Midjourney’ or ‘DALL-E.’ It starts with fuzzy data like TV static noise and gradually carves away unnecessary information to complete a very sophisticated and clean future weather map, much like a high-resolution photograph Source 7.
Current Status: How accurate is it?
To verify GenCast’s skills, Google DeepMind pitted it against the standard model (ENS) of the European Centre for Medium-Range Weather Forecasts (ECMWF), which is renowned for being the best in the world. The results were quite shocking.
- 97.2% Overwhelming Victory: GenCast outperformed the existing model in a 15-day forecast showdown with a staggering 97.2% probability Source 6. In other words, out of 100 forecasts, the AI gave more accurate answers more than 97 times Source 15.
- A Detailed View: It divides the entire globe into a dense grid of approximately 28km (0.25°) and tracks weather changes in 12-hour intervals Source 2. Predicting 15 days into the future with this level of precision is a breakthrough that solves a long-standing challenge in meteorology Source 11.
Notably, this AI excels at identifying extreme situations such as heatwaves, deep freezes, and strong winds. Instead of just saying “it will be a bit hot,” it sends specific warnings in advance, such as “there is a certain percentage chance of a record-breaking, life-threatening heatwave in this region” Source 16.
Future Outlook: How will our lives change?
The emergence of GenCast is a major turning point showing that the protagonist of weather forecasting is shifting from ‘giant supercomputers and physical formulas’ to ‘vast data and AI.’
In the future, we will receive much more accurate long-term forecasts on our smartphones. Beyond the convenience of planning vacations, this will significantly contribute to global efficiency, such as optimizing flight paths to drastically reduce carbon emissions or accurately predicting energy demand to prevent power waste.
Of course, GenCast is not a magic ball that gets every weather prediction 100% right. However, by clearly organizing the uncertain future with the scientific tool of ‘probability,’ it will help us respond to unexpected natural disasters more resolutely and intelligently.
AI Perspective: A Word from MindTickleBytes AI Reporter
GenCast is not just a technology for predicting the weather. It is teaching us how humanity should manage the ‘uncertainty’ of the giant system called Earth. The figure of 97.2% means more than just a performance edge; it signifies that humanity has gained the world’s most powerful lens—’AI’—in understanding the whims of nature. Now, weather forecasting has completely entered the realm of “thorough preparation based on data” rather than “believe it or not” style guesswork.
References
- GenCast predicts weather and the risks of extreme conditions …
- Probabilistic weather forecasting with machine learning - Nature
- [2312.15796] GenCast: Diffusion-based ensemble forecasting …
- GenCast predicts weather and the risks of extreme conditions …
- Inside Google’s GenCast: Learn About AI in Forecasting
- Google AI boosts weather accuracy - LinkedIn
-
[Weather research WeatherNext Google for Developers](https://developers.google.com/weathernext/guides/research) - GenCastpredictsweatherandtherisksofextremeconditions…
- GenCastfrom Google DeepMind provides betterweatherforecasts
- See the latest updates, context, and perspectives about this story.
- Google Reveals New A.I. Model That Predicts Weather Better Than the Best Traditional Forecasts
- Google’s GenCast: Weather Forecasting With GenCast Mini Demo
- Approximately 50%
- Approximately 75.5%
- Approximately 97.2%
- It presents only one fixed result
- It presents over 50 different possibilities (scenarios)
- It lists past data without using a supercomputer
- 3 days
- 7 days
- 15 days