Skip to main content

Inside Google’s GenCast: Learn About AI in Weather Forecasting

Discover how GenCast makes use of machine learning and probabilistic modeling to deliver faster, more accurate forecasts for extreme weather events. See how it's reshaping the future of weather forecasting, offering insights that traditional models can’t match.
Jan 29, 2025  · 11 min read

Weather affects everyone in every way - what we do, how safe we feel, and how we live our lives. This makes accurate weather forecasting a necessity rather than a luxury, especially in an era marked by climate change and, alarmingly, increasingly frequent extreme weather events.

Now, weather forecasting is not something new, and has been there for decades in the form of traditional forecasting. Even though it’s effective, it often faces challenges in predicting weather accurately for long durations.

So, let me talk about GenCast, which is an extremely impressive (revolutionary, even) AI-powered model, developed by Google DeepMind, that promises to improve weather forecasting accuracy up to 15 days in advance. GenCast integrates advanced machine learning techniques with weather science to deliver more reliable, high-resolution predictions. 

Rise of AI in Weather Forecasting

Traditional weather models rely on numerical weather forecasting (NWP), which involves feeding initial weather conditions into supercomputers to create a single deterministic forecast. This is good, but the problem is that weather is a chaotic system, meaning even a small change in initial conditions - like a slight alteration in temperature, maybe, or something else - can lead to vastly different results. This phenomenon is known as the butterfly effect, a concept from chaos theory.

To address this volatility, forecasters create multiple simulations by slightly tweaking the variables. These are then combined into what’s called an ensemble forecast (ENS) by the European Centre for Medium-Range Weather Forecasting (ECMWF).

Despite its success, ENS has limitations, namely computational inefficiency, meaning there are high resource requirements for running simulations, and limited uncertainty modeling, which refers to the difficulty in representing the full spectrum of potential outcomes.

AI has emerged as a possible solution to address these challenges. While early models focused on deterministic forecasts, recent advancements like hybrid models (e.g., NeuralGCM) have improved predictive accuracy. However, these models often fall short in speed and scalability—issues that GenCast aims to resolve.

What is GenCast?

Let's first talk about the origins of GenCast and then talk about what makes it so compelling.

The origins of GenCast

GenCast represents a significant breakthrough in weather forecasting. Developed by Google DeepMind and published in the journal Nature, GenCast is the first Machine Learning-Based Weather Prediction (MLWP) model to outperform ENS in accuracy and efficiency. Initially designed to overcome the computational bottlenecks of traditional systems, GenCast leverages the latest advancements in machine learning and generative AI to revolutionize weather prediction.

The results so far have been impressive. GenCast was 97.2% more accurate than traditional models in 15-day forecasts and 99.8% more accurate for shorter 36-hour forecasts. Even more remarkable is its speed: It generates forecasts in just 8 minutes, compared to the traditional methods that may take hours on a supercomputer with tens of thousands of processors.

Understanding GenCast’s innovation

GenCast takes a fresh approach to weather forecasting. Instead of relying solely on numerical simulations, it uses probabilistic forecasting powered by its use of a diffusion model, a type of generative AI commonly used in image and video generation.  

GenCast adapts this model to the Earth's spherical geometry, enabling it to learn and generate complex probability. The diffusion models simulate realistic joint distributions for variables like wind and precipitation, enhancing probabilistic forecasts. This generative approach minimizes computational demands, making it scalable and accessible.

How GenCast Works

Gencast was trained on 40 years of historical weather data, covering over 80 variables such as temperature, humidity, and pressure. The algorithm is based on diffusion models, commonly used for generating images in tools like DALL-E or MidJourney, which were adapted for weather forecasting. The high level overview of the process is depicted below:

Source: Image by Author

Source: Image by Author

Essentially, the process begins with noisy data, which represents uncertainty or randomness in the system. Noise, in this context, refers to initial conditions or data points that lack clarity due to the chaotic nature of weather patterns. At each step, the model reduces the level of uncertainty by using learned patterns and relationships from historical weather data.

Instead of focusing on a single outcome, the algorithm produces a probabilistic ensemble of forecasts, which captures the scenario-based probabilities. This is particularly useful in predicting extreme events like cyclones, where precise timing and trajectory are super critical.

After multiple iterations, the model generates a highly refined forecast. This approach allows GenCast to create realistic joint distributions of variables, such as predicting how wind power potential correlates with temperature and pressure changes. The strength of this approach lies in its ability to work with the inherent uncertainty of weather systems.   

Applications and Real-World Implications of GenCast

GenCast is one of Google's advanced AI-driven weather models, alongside others like Google Research's NeuralGCM, SEEDS, and flood prediction models. These tools are being integrated into several of its services like Search and Maps to enhance predictions for precipitation, wildfires, flooding, and extreme heat. Some of these applications are discussed below.

Real World Applications of Gencast

Real World Applications of Gencast. Source: Image by Author

Extreme weather preparedness

The perils of extreme weather events like cyclones, hurricanes, and heatwaves are well known. These events are becoming more common and also more severe due to climate change. Take the recent devastating fires in LA as an example. GenCast could offer accurate tracking of these events days in advance, enabling proactive response from the governments and emergency service agencies. For example:

  • Disaster Response: I mentioned that GenCast can predict events like the landfall of a cyclone several days ahead. This means authorities can issue timely evacuation orders, prepare shelters, or do something else that is important and useful.
  • Economic Savings: Proactive disaster management means economic losses are minimized by helping businesses secure assets and reducing downtime for critical operations.
  • Life-Saving Impact: Perhaps, the most important benefit is that by providing longer lead times for preparation, GenCast ensures that more lives can be saved during extreme events.

For instance, in a test with typhoon Hagibis, GenCast predicted the landfall seven days in advance. I think it's hard to understate the importance of this kind of predictive power in disaster management.

Energy sector

There is a clear shift in global policies towards renewable energy, particularly wind and solar. These rely heavily on accurate weather forecasts, which is where GenCast plays a role in optimizing energy production:

  • Wind Power Forecasting: Accurate prediction of wind speeds and directions helps energy companies estimate power generation from wind farms. This in-turn improves grid reliability and efficient energy distribution.
  • Solar Energy Planning: Accurate forecasting of important variables such as cloud coverage and sunlight intensity enables solar power operators to plan energy storage and distribution effectively.
  • Grid Stability: Enhanced forecasting ability with Gencast allows energy providers to create balance between the supply and demand, thereby reducing the risk of outages or energy wastage.

Agriculture and transportation

Weather conditions are a critical factor in agriculture and logistics. Some of the use cases are outlined below. 

  • Agriculture: Key farming decisions such as planting, irrigating, and harvesting are dependent on the weather, so accurate predictions matter a lot. GenCast's accuracy over a longer lead time helps farmers better plan and optimize these activities to increase yields and reduce losses.
  • Transportation and Logistics: Mathematical optimization is a well known application in transportation and logistics. However, this optimization is dependent on accurate forecasts so as to enable these logistics companies to plan routes and schedules. This is something that Gencasts state of the art forecasting model makes possible. 
  • Supply Chain Management: Companies can better predict disruptions caused by extreme weather and adjust supply chains accordingly, reducing inefficiencies and costs.

These applications underscore GenCast’s value in creating significant impact across several industries. 

Challenges and Future Directions

While GenCast is a breakthrough, it’s not perfect. Here are some of its current limitations:

  • Lower Resolution: Its resolution of 0.25 degrees is not as fine as traditional models like ECMWF’s 0.1-degree resolution. By resolution, I mean the spatial resolution of the GenCast model, which is the smallest area that the model can distinguish in its weather predictions. A 0.25-degree resolution means that each grid cell in the model represents an area approximately 28 km × 28 km at the equator, whereas ECMWF’s 0.1-degree resolution represents a finer 11 km × 11 km grid.
  • Real-Time Data Integration: Currently, GenCast cannot incorporate real-time data or predict cloud coverage effectively.
  • Rare Event Prediction: Due to its reliance on 40 years of historical data, GenCast struggles to predict rare events, such as once-in-a-hundred-year disasters.
  • High Computational Demands: While efficient, diffusion models still require significant resources. Techniques like model distillation, where a smaller model learns from a larger complex one, could help mitigate this.

How You Can Use GenCast in Your Own Work

Developers and data scientists interested in GenCast coding can access its implementation through Google DeepMind's GraphCast and GenCast GitHub repository. This repository offers a lot of resources such as sample code, pretrained model weights, and detailed guidelines for running and training the GenCast weather model. 

It also includes Jupyter notebooks like gencast_mini_demo.ipynb, which provides a step-by-step tutorial to load data, generate predictions, and compute gradients. Additionally, the repository contains essential utilities facilitating the exploration of GenCast's forecasting capabilities.

Additional Notes on Forecasting and Probability

Forecasting relies on mathematical techniques to identify patterns and predict future outcomes. Time series analysis, regression models, and machine learning algorithms all play a role in making sense of historical data. A classic example of data-driven forecasting is captured in Rob Hyndman’s famous quote: “You can forecast rain based on the number of cyclists on the road.” While it may seem humorous, it illustrates how relationships between variables—such as people avoiding biking in bad weather—can be used to make informed predictions.

Many forecasting problems involve nonlinear systems, where small changes in initial conditions can lead to vastly different outcomes. This is particularly true for complex, dynamic systems like weather or financial markets. Despite these challenges, mathematical models can extract meaningful trends and generate forecasts with varying degrees of confidence.

The role of probabilistic modeling in forecasting

Traditional forecasting methods often provide single-point estimates, but real-world systems are inherently uncertain. Probabilistic modeling improves upon deterministic forecasts by estimating a distribution of possible outcomes. Instead of saying, “It will rain tomorrow,” a probabilistic model might predict a 70% chance of rain, incorporating the uncertainty that comes with complex systems.

This method involves conditional probability distributions, Bayesian inference, and stochastic processes to quantify uncertainty. Similar techniques are used in AI models like the ones uses in ChatGPT, which predict the next word in a sentence by weighing probabilities of various options. Instead of predicting words, forecasting models use probabilities to anticipate events, trends, and conditions.

How GenCast applies these concepts

GenCast leverages both forecasting techniques and probabilistic modeling to improve prediction accuracy. Rather than providing a single deterministic outcome, it models the likelihood of different future states based on historical and real-time data. By doing so, it accounts for uncertainty and variability, making it more reliable than traditional models.

Conclusion

GenCast is more than just a weather forecasting model; it’s a glimpse into the future of AI-powered predictions. Truly, it has the potential to revolutionize industries, improve disaster preparedness, and even save lives. It will be exciting to witness the transformative leap Gencast will take and the impact it will make in the years to come.

I talked about a lot of different ideas in this article, and thankfully, we here at DataCamp have a full range of resources to help. I'm going to link options below, so you can keep learning about forecasting, probabilistic modeling, and generative AI concepts.

  • Generative AI Concepts: Learn how generative AI models are developed and how they will impact society moving forward.
  • Forecasting in R: Learn how to make predictions about the future using time series forecasting in R including ARIMA models and exponential smoothing methods.
  • Foundations of Probability in Python: Learn fundamental probability concepts like random variables, mean and variance, probability distributions, and conditional probabilities.
  • AI Courses: Build your skills in topics such as generative AI, prompt engineering, deep learning and more.

Vikash Singh's photo
Author
Vikash Singh
LinkedIn

Seasoned professional in data science, artificial intelligence, analytics, and data strategy.

Gencast FAQs

What is GenCast?

GenCast is an advanced AI-based weather forecasting model developed by Google DeepMind. It predicts weather conditions up to 15 days in advance, by using a probabilistic approach instead of traditional methods, and has achieved high accuracy.

How does GenCast differ from traditional weather models?

Traditional models rely on numerical weather prediction (NWP), which uses physics-based equations and can take days to generate forecasts. GenCast, on the other hand, uses machine learning and historical data to predict a range of possible weather scenarios in a much shorter time.

What is the key innovation behind GenCast?

The key innovation is that GenCast uses a diffusion model, which refines noisy data in an iterative manner  to generate realistic weather predictions.

Can GenCast predict extreme weather events?

Yes, GenCast is effective at predicting extreme weather events like cyclones, hurricanes, and heatwaves. It can track cyclones and predict their landfall days in advance, allowing for better emergency preparedness.

How does GenCast handle uncertainty in weather forecasting?

GenCast uses a probabilistic framework, meaning it generates a range of possible outcomes and their likelihoods. This approach accounts for the volatility in the weather systems and makes its forecasts more reliable.

How is GenCast more accurate despite having lower resolution?

GenCast is more accurate because it uses probabilistic modeling and diffusion models to analyze historical weather data and generate a range of likely outcomes. However, its resolution is currently 0.25 degrees, which means it provides predictions for larger grid areas which means it may miss very localized weather variations. However, for broader areas, this approach allows it to predict with a higher accuracy than traditional models.
Topics

Learn with DataCamp

course

Forecasting in R

5 hr
49.4K
Learn how to make predictions about the future using time series forecasting in R including ARIMA models and exponential smoothing methods.
See DetailsRight Arrow
Start Course
See MoreRight Arrow
Related

blog

AI Time Series Forecasting: A Beginners' Guide

Learn how AI is transforming time series forecasting through adaptability, uncovering hidden patterns, and handling multiple variables.
Stanislav Karzhev's photo

Stanislav Karzhev

8 min

blog

GPT-3 and the Next Generation of AI-Powered Services

How GPT-3 expands the world of possibilities for language tasks—and why it will pave the way for designers to prototype more easily, streamline work for data analysts, enable more robust research, and automate content generation.
Adel Nehme's photo

Adel Nehme

7 min

blog

What is a Generative Model?

Generative models use machine learning to discover patterns in data & generate new data. Learn about their significance & applications in AI.
Abid Ali Awan's photo

Abid Ali Awan

11 min

OpenAI Google AI Data Science

blog

The Latest On OpenAI, Google AI, and What it Means For Data Science

Learn about the disruptive language, vision, and multimodal technologies and how it is making us more productive and effective.
Abid Ali Awan's photo

Abid Ali Awan

13 min

podcast

The Data to AI Journey with Gerrit Kazmaier, VP & GM of Data Analytics at Google Cloud

Richie and Gerrit explore AI in data tools, the evolution of dashboards, the integration of AI with existing workflows, the challenges and opportunities in SQL code generation, the importance of a unified data platform, and much more.
Richie Cotton's photo

Richie Cotton

55 min

tutorial

Time Series Forecasting With TimeGPT

Learn how to fine-tune TimeGPT, the first foundational model for time series datasets, for forecasting and anomaly detection with just a few lines of code.
Abid Ali Awan's photo

Abid Ali Awan

14 min

See MoreSee More