homehome Home chatchat Notifications


Google's DeepMind claims they have the best weather prediction model

After chess, Go, and proteins, has DeepMind taken over weather forecasting as well?

Mihai Andrei
December 5, 2024 @ 6:13 pm

share Share

DeepMind, the AI powerhouse behind groundbreaking innovations like AlphaGo or AlphaFold, has unveiled a bold claim: its new machine learning model, GenCast, surpasses the world’s best operational weather forecasting system.

According to a new study, the new algorithm combines decades of meteorological data with cutting-edge AI techniques to achieve faster, more accurate predictions than existing systems. This breakthrough promises not only more accurate weather forecasts but also better anticipation of extreme weather events — potentially saving lives and optimizing resource planning globally.

A conceptual representation of AI in weather forecasting. AI-generated image.

Predicting the weather

Weather forecasting is an important part of modern life, influencing decisions from daily commutes to disaster preparedness. It’s not just about whether you take your umbrella today or not. Accurate weather forecasts help protect lives and livelihoods. They support economic activities like agriculture and energy management, and enable preparedness for extreme weather events that can disrupt communities.

Yet weather forecasting is inherently challenging. Traditional methods, based on numerical weather prediction (NWP), involve solving complex equations that simulate atmospheric dynamics. These mathematical equations describe how air, water, and energy interact in the atmosphere over time. By inputting current weather data into these equations, NWP predicts future states of the weather, such as temperature, wind patterns, and precipitation.

Researchers have made great progress with this method. However, NPW is computationally intensive and can be slow, often requiring large supercomputers and hours of processing time to generate forecasts, particularly for long-term or high-resolution predictions.

DeepMind has a different idea: using machine learning.

This isn’t exactly a new concept. Last year, DeepMind unveiled GraphCast, which produces a single best-guess forecast at a time. But a single, perfect weather forecast is not possible. Instead, weather agencies typically work with probabilistic ensembles — likelihoods of particular weather patterns. What you’re seeing as a weather forecast is only the most likely outcome; behind it, there’s a range of likelihoods for other types of weather.

“Such ensemble forecasts are more useful than relying on a single forecast, as they provide decision makers with a fuller picture of possible weather conditions in the coming days and weeks and how likely each scenario is,” writes DeepMind in an explanatory post.

Now, the company has unveiled GenCast — which essentially builds on GraphCast by generating an ensemble of 50 or more forecasts, each with their own probability.

Image of possible weather predictions via GenCast
Image credits: DeepMind.

As good as the best models, and much faster

At its core, GenCast generates an ensemble of 15-day global forecasts with unprecedented resolution (0.25° latitude-longitude grid, or around 27 km) and speed. In only eight minutes, the model produces forecasts for over 80 weather variables at 12-hour intervals.

The secret lies in its ability to model probabilistic outcomes. Unlike traditional deterministic models, GenCast provides a range of possible weather scenarios and their likelihoods. Powered by DeepMind’s machine learning, this is much more efficient than deterministic approaches.

The system was tested against the state-of-the-art European Centre for Medium-Range Weather Forecasts (ECMWF), widely regarded as the world leader in the field. DeepMind’s GenCast performed as good or better in nearly every evaluation metric. It excelled in three critical areas:

  1. Resolution and Detail
    With its fine-grained resolution, GenCast provides a more detailed view of atmospheric changes, capturing localized weather phenomena with greater clarity than its counterparts.
  2. Accuracy in Extreme Weather Prediction
    GenCast consistently outperformed ENS in forecasting rare and extreme events, such as tropical cyclones and high-wind scenarios. This improvement has profound implications for disaster preparedness, enabling earlier and more reliable warnings.
  3. Speed and Efficiency
    Generating a full set of forecasts in mere minutes, GenCast offers a substantial advantage over traditional systems, which require significantly longer processing times. This speed ensures timely updates and better responsiveness during rapidly evolving weather situations.

“We comprehensively tested both systems, looking at forecasts of different variables at different lead times — 1320 combinations in total. GenCast was more accurate than ENS on 97.2% of these targets, and on 99.8% at lead times greater than 36 hours,” DeepMind says.

Exciting, but not ready to take over

Researchers praised the emergence of this new algorithm. Steven Ramsdale, a Met Office chief forecaster with responsibility for AI, told The Guardian the work was “exciting”, while a spokesperson for the ECMWF called it “a significant advance”, adding that some components of GenCast were being used in some of its forecasts. Meanwhile, Sarah Dance, a professor of data assimilation at the University of Reading, said weather forecasting could be on the brink of a “fundamental shift in methodology”.

Yet, despite its groundbreaking achievements, GenCast is not without limitations. Its reliance on historical reanalysis data, for instance, constrains its performance in scenarios requiring real-time updates. Enhancements such as fine-tuning with operational data could further improve its accuracy and adaptability. Experts also cautioned that it’s not clear just how well the AI can handle “butterfly effects,” a cascade of growing uncertainties that becomes prevalent when you try to zoom in on microclimates.

Although GenCast already operates at a high resolution, expanding its scope to even finer scales could unlock new applications, such as urban-level weather forecasting. However, this requires addressing the computational demands of diffusion models.

Finally, integrating GenCast with existing weather infrastructure remains a critical step. Collaboration with meteorological agencies and stakeholders will ensure that the model’s insights are effectively translated into actionable strategies.

Even with all these limitations, the advent of GenCast is a testament to the transformative potential of generative AI. Beyond weather forecasting, the model’s underlying technology has applications in climate modeling, disaster risk assessment, and even planetary science. By capturing the complexity of atmospheric dynamics with unprecedented accuracy, GenCast paves the way for more resilient and informed societies.

With further refinements, GenCast and its successors could redefine our understanding of weather and the role of AI in shaping it.

The article was published in Nature.

share Share

Mexico Will Give U.S. More Water to Avert More Tariffs

Droughts due to climate change are making Mexico increasingly water indebted to the USA.

Chinese Student Got Rescued from Mount Fuji—Then Went Back for His Phone and Needed Saving Again

A student was saved two times in four days after ignoring warnings to stay off Mount Fuji.

The perfect pub crawl: mathematicians solve most efficient way to visit all 81,998 bars in South Korea

This is the longest pub crawl ever solved by scientists.

This Film Shaped Like Shark Skin Makes Planes More Aerodynamic and Saves Billions in Fuel

Mimicking shark skin may help aviation shed fuel—and carbon

China Just Made the World's Fastest Transistor and It Is Not Made of Silicon

The new transistor runs 40% faster and uses less power.

Ice Age Humans in Ukraine Were Masterful Fire Benders, New Study Shows

Ice Age humans mastered fire with astonishing precision.

The "Bone Collector" Caterpillar Disguises Itself With the Bodies of Its Victims and Lives in Spider Webs

This insect doesn't play with its food. It just wears it.

University of Zurich Researchers Secretly Deployed AI Bots on Reddit in Unauthorized Study

The revelation has sparked outrage across the internet.

Giant Brain Study Took Seven Years to Test the Two Biggest Theories of Consciousness. Here's What Scientists Found

Both came up short but the search for human consciousness continues.

The Cybertruck is all tricks and no truck, a musky Tesla fail

Tesla’s baking sheet on wheels rides fast in the recall lane toward a dead end where dysfunctional men gather.