- WeatherNext 2 uses a Functional Generative Network to create coherent and realistic weather scenarios.
- Generate hundreds of forecasts in less than a minute with a single TPU, with hourly resolution.
- It improves accuracy to 99,9% of variables between 0 and 15 days and is already integrated into Search. Gemini and Pixel.
- Data available in Earth Engine and BigQuery, and early access in Vertex AI for custom models.
If you're one of those people who check the weather forecast before leaving home, you'll know that the apps Weather apps are now almost as essential as your mobile phone. Among the most popular are BBC Weather, Met Office, Apple Weather, The Weather Channel, and AccuWeather, all very reliable for planning your day. However, what comes with the new model of Google promises to raise the bar: WeatherNext 2 It comes with the idea of accelerating, refining and making much more useful the forecasts we consult at all times.
The result of a joint effort between Google DeepMind and Google Research, this system IA aims to generate forecasts eight times faster than traditional approaches, with a temporal resolution of up to one hour and the ability to simulate hundreds of scenarios from a single starting point. This leap is already noticeable in products you use daily, because the forecasts in the Search, Google Gemini, Pixel Weather and the Google Maps Platform API They have been updated to take advantage of the new technology, with more new features on the way for Google Maps.
What is WeatherNext 2 and why is everyone talking about it?
WeatherNext 2 is the second iteration of Google's AI-powered weather forecasting system. Unlike a classic deterministic model, this approach focuses on generate multiple plausible scenarios that respect the physics of the atmospheric system. With a single initial weather state, it is capable of producing in less than a minute a range of results that helps to anticipate both the most probable behavior and low-probability extreme cases, key to risk management.
The change is also significant operationally: instead of relying on calculations that run for hours on supercomputers with numerical models, WeatherNext 2 runs its inferences on a TPU (Tensor Processing Unit) And it obtains hundreds of forecasts in record time. This efficiency isn't just cosmetic: it means faster decisions when urgent action is needed, for example, in the face of severe storms or winds that affect critical infrastructure.
How it works: the Functional Generative Network behind the model
The technical heart of the system is called Functional Generative Network (FGN)This architecture injects controlled "noise" directly into the model's functional space, helping simulations maintain physical consistency and interdependence between variables. In practice, instead of fixing a single trajectory, the model generates several possible atmospheric state paths that fit together and are compatible with the system's own laws.
An intuitive way to understand this is to think of an orchestra rehearsing the same piece in sections: each group works on its part with small variations, but They all fit together harmoniously when combined. Similarly, FGN allows different networks to learn aspects of time separately, introduce variability, and still build consistent joint forecasts on a physical level.
This approach overcomes the limitations of "marginal" forecasts, those that describe an isolated variable (such as the exact temperature at a specific point and time). The FGN facilitates the transition to joint predictions where temperature, wind, humidity, or other magnitudes evolve in an interrelated way, something crucial for capturing complex phenomena and for incorporating low probability extreme events within the range of scenarios.
Another key aspect is that the model is trained with what the researchers describe as “loose data” about the climateMeasurements of variables such as wind, temperature, or humidity at specific locations. From this data, the FGN is able to infer and anticipate the dynamics of larger-scale weather systems, identify areas at risk of heat waves or estimate how much energy a wind farm could generate under future conditions.
Speed, resolution, and prediction horizons
In terms of metrics, the leap compared to the previous version is significant: according to Google, WeatherNext 2 It outperforms 99,9% of the variables analyzed. (temperature, wind, humidity and more) across all forecast horizons 0 to 15 daysFurthermore, it offers a time resolution of up to one hour, allowing for finer forecasts where extra detail makes a difference, such as in power grid management or logistics operations.
Computational performance also changes the game. Where conventional simulations require several hours on supercomputersWeatherNext 2 calculates hundreds of scenarios in under a minute using a TPU. And there's an interesting operational nuance: the system processes the most recent atmospheric conditions and, from there, generates four six-hour forecasts each one per day, incorporating variations thanks to the FGN to cover the range of possibilities in more detail.
- Covered horizons: from nowcasting and short term up to 15 days, with clear improvements in accuracy.
- Temporary resolution: up to one hour, useful for immediate decisions in sensitive sectors.
- Scenario capacity: hundreds in less than a minute from a single initial state.
- Efficient computing: execution on a TPU, without the need for massive supercomputing.
Integration into Google products and availability to the community
WeatherNext 2 technology is already driving improvements in Search, Gemini, Pixel Weather and the Google Maps Platform API. Google has also announced that in the coming weeks we will see visible improvements in Google Maps related to meteorological information, making it easy to check at a glance the conditions and the expected evolution according to different scenarios.
Regarding data access, Google states that WeatherNext 2 forecasts are available in Earth Engine and BigQueryFurthermore, they have launched a early access program at Vertex AI (Google Cloud) which allows companies and technical teams to run customized inferences and adjust the model to their needs, a direct path to adoption in real-world environments.
Looking ahead, the Google DeepMind and Google Research team assures that they are investigating new capabilities to integrate more data sources, expand access, and continue converting cutting-edge research in high-impact applicationsThe stated commitment is to make recent tools available to the global community to accelerate scientific progress and improve decision-making.
From the labs to everyday life: what changes for the user
The practical difference is that, starting from a single input, WeatherNext 2 can simulate hundreds of possible evolutions It will show you the most likely and worst-case scenarios to watch out for. This is especially useful for planning. travel, flights, commutes, supply chains or the operations of delivery companies, where small weather variations imply costs or delays.
If you already use a Pixel phone, you know its weather app is known for its clarity. With WeatherNext 2, predictions in Pixel Weather and in Search They gain accuracy and speed, so you can check key information "at a glance," including the range of possibilities instead of a single optimistic figure that then changes at the last minute.
Google suggests that this advance reduces dependence on radars and satellites as they had been used until now, since the model learns patterns from large volumes of historical and most recent data. In any case, the beauty of the new system is not to replace everything, but create a more useful forecast integrating different signals and quickly covering uncertainty.
Use cases: from energy to emergency management
- Emergency services: Evaluate in advance scenarios of severe storms, floods or strong winds to activate protocols in time.
- Renewable energy: to estimate future production in wind or solar farms, adjusting supply to demand with greater precision.
- Transport and logistics: Optimize routes and operating windows in the face of rain, snow or gusts of wind, minimizing delays and costs.
- Public sector and insurance: analyze risks in worst case scenarios low probability and high impact to size response and coverage.
- Research and education: explore the dynamics of extreme events and compare joint scenarios to better understand the atmosphere.
What does it offer compared to traditional methods?
Numerical weather models solve physical equations of the atmosphere with high-resolution grids and, by their very nature, They require a lot of computing powerWeatherNext 2's AI doesn't ignore physics: it learns from historical data and the current state to generate consistent plausible scenarios with those laws, but it does so with a lower computational cost and in very short times.
Its competitive advantage is the ability to cover probabilistically the range of possibilities, including extreme cases that are relevant for planning. Where previously a large set of costly simulations was needed to approximate that uncertainty, FGN introduces "intelligent" variability that is physically consistent, allowing exploration hundreds of trajectories almost in real time.
Limits, responsibilities and good practice
It's worth remembering something basic: the atmosphere is a chaotic systemEven with AI, 100% infallible predictions will not be achieved. The danger of a very accessible and seemingly accurate tool is that the average user will assume certainties where there are only probabilities. Here, it is essential that the interfaces communicate the uncertainty and that minimum meteorological literacy be maintained.
There are voices raising doubts about whether it is a good idea.putting WeatherNext 2 in anyone's hands"Without expert filtering, just as an algorithm that detects cancer needs a radiologist to validate the result. The tool is incredibly powerful, but the nuanced interpretation, especially in the face of..." safety and self-protection situations, still requires professional judgment.
Recent experience with severe events demonstrates this. In episodes such as a DANA in ValenciaMaking decisions late can worsen the impact. A system that quickly provides you with worst-case scenarios and clearly warns you, gain valuable minutes to activate plans and reduce risks. That's where improvements in speed and resolution are most transformative.
The key will be to combine the best of both worlds: the AI power to produce rich and rapid scenarios, and the validation of specialists to refine messages and action guidelines. If, in addition, honest communication is maintained regarding ranges and probabilities, the result will be a better-prepared citizenry without a false sense of certainty.
Access for companies, scientists, and developers
For those who need to work with data at scale, Google has opened the WeatherNext 2 forecasts in Earth Engine and BigQueryfacilitating consultations, analysis, and cross-referencing with other sources. And for those seeking to adapt the model to their specific case, there is a early access program at Vertex AI with support for custom model inference and deployments in production environments.
The roadmap involves integrating new data sourcesto expand access and continue translating research advances into high-impact, real-world applications. With this impetus, agencies and companies can accelerate their transition to a data-driven decision making where There atmospheric conditions, far from being an unpredictable factor, become a measurable and manageable risk.
WeatherNext 2 is presented as a qualitative leap in weather forecasting: eight times faster than traditional approaches, capable of exceed 99,9% of the variables Compared to its predecessor in horizons of 0 to 15 days, and ready to produce hundreds of scenarios in less than a minute from a single initial state. Integration with Search, Gemini, Pixel Weather, and the Maps API is already underway, while data in Earth Engine and BigQuery, along with early access to Vertex AI, opens the door to professional and scientific uses that, properly guided, can make a real difference in how we prepare for what might happen.
Passionate writer about the world of bytes and technology in general. I love sharing my knowledge through writing, and that's what I'll do on this blog, show you all the most interesting things about gadgets, software, hardware, tech trends, and more. My goal is to help you navigate the digital world in a simple and entertaining way.