DeepMind WeatherNext предсказывает ураганы раньше существующих моделей прогноза

Google DeepMind announced WeatherNext in August 2026 — an AI model that predicts hurricanes earlier and more accurately than any existing weather forecasting system, while using lower-resolution data.

What WeatherNext Can Do

WeatherNext addresses two tasks simultaneously: it determines a hurricane's track (where the storm is heading) and its intensity (how powerful it is). Traditional numerical models often handle one parameter better than the other — WeatherNext forecasts both in tandem and outperforms existing systems in lead time for warnings.

• Developer — Google DeepMind
• The model will be released as open source
• Forecasts hurricane track and intensity simultaneously
• Operates on lower-resolution data
• Issues warnings earlier than traditional methods

Every additional hour of forecast lead time has direct implications for safety: it allows for the evacuation of populations, the deployment of emergency services, and the preparation of critical infrastructure before the storm arrives. The more accurately and earlier the path and force of a hurricane are known, the more effectively meteorological services and authorities can allocate resources.

Why

Even the Researchers Themselves Don't Understand How It Works

The defining feature of WeatherNext — and its central mystery — is that the mechanism by which the model operates is not fully understood even by its developers. As Wired reports, the DeepMind team cannot fully explain how the model extracts reliable forecasts from lower-resolution data.

"Researchers don't yet fully understand how exactly the model does this," states the

Wired article, citing DeepMind.

This is the classic problem of neural networks — the opacity of the architecture (the "black box"). For meteorological services and emergency agencies, this has practical implications: decisions to evacuate entire regions are made based on forecasts. If the model makes an error and the cause of the error is opaque, correcting its behavior is extremely difficult. DeepMind describes uncovering this mechanism as the subject of ongoing research.

Why Low-Resolution Data Is a Breakthrough

In classical meteorology, forecast accuracy depends directly on the density and detail of observations: more data and higher quality means a better forecast. WeatherNext inverts this logic: the model achieves high accuracy with input data that traditional systems would consider insufficient.

The practical significance of this is especially great for tropical regions and small island states: these are precisely the places where hurricanes are most destructive, and precisely where meteorological infrastructure is often outdated or incomplete — sparse ground stations, gaps in data, a lack of modern equipment. A model capable of delivering early and accurate forecasts under conditions of information scarcity fundamentally changes the capabilities of warning systems for millions of people.

DeepMind plans to publish WeatherNext as open source, which will allow national meteorological services and independent researchers worldwide to adapt the model, verify its results, and improve it using their own regional data.

What This Means

If WeatherNext passes independent verification, AI-based hurricane forecasting could become part of the standard toolkit for meteorological services around the world. Open source will accelerate this transition: regions with the greatest vulnerability to hurricanes will gain access to a cutting-edge model without having to develop one themselves.

Frequently Asked Questions

When Will WeatherNext Be Released as Open Source?

DeepMind has announced its intention to publish WeatherNext as open source, but no specific timeline is given in the Wired publication.

What Does Operating "on Lower-Resolution Data" Mean?

Traditional forecasting models require a dense network of high-detail meteorological observations. WeatherNext demonstrates high accuracy with more "sparse" input data — this is critically important for countries and regions with underdeveloped weather monitoring infrastructure.