DeepMind weather model outperforms government forecasters in accuracy benchmarks

Developing story first seen 2 hours ago

· 2 hours ago

Google has released WeatherNext 3, a new AI weather forecasting model built by DeepMind and Google Research, which has already topped independent accuracy benchmarks and will begin powering weather data across Google Search, Maps and Gemini. The model marks a further step in the shift from traditional supercomputer-run physics simulations towards deep learning approaches, which are faster and cheaper while now matching or beating government forecasters on core metrics such as temperature, wind speed and humidity.

On Operational WeatherBench, an independent testing platform built by startup Brightband, WeatherNext 3 outperformed rival AI models from Microsoft, Nvidia and the ECMWF, as well as official forecasts from the US National Weather Service and ECMWF. With 2.4 times more parameters than its predecessor, it can forecast key variables at a resolution of 5km rather than the previous 15–25km, produces hourly rather than six-hourly updates, and has improved rainfall predictions by 60%. It has also been trained to target forecasts at specific weather stations, allowing more localised predictions that can be checked directly against ground measurements.

Google's WeatherNext 3 AI model now tops independent weather forecasting benchmarks.
It offers hourly, 5km-resolution forecasts and 60% better rain predictions.
It will feed weather data into Google Search, Maps and Gemini.

New here? Start with this

Google has developed an artificial intelligence system, called WeatherNext 3, designed to predict the weather. It was built by DeepMind and Google Research, the same teams behind other well-known AI projects, and it works very differently from the way forecasts have traditionally been made. Rather than running enormous physics-based simulations on supercomputers, it uses machine learning trained on weather data to generate predictions, which is generally quicker and cheaper to run.

For decades, the most trusted forecasts have come from national weather agencies, such as the US National Weather Service, and international bodies like the European Centre for Medium-Range Weather Forecasts (ECMWF), all using physics-based modelling. In recent years, tech companies including Google, Microsoft and Nvidia have been building rival AI systems that learn patterns from historical weather data instead, and these are increasingly being tested against the official forecasters to see which approach is more accurate.

This matters because weather forecasts affect everything from daily travel plans to farming, aviation and disaster preparedness, so improvements in accuracy or detail have real practical value. It also reflects a broader trend of AI being used to tackle large-scale scientific and public-service tasks once dominated by government institutions, raising questions about the future role of state-run forecasting agencies versus commercial technology firms.

Both sides, in good faith

The strongest fair case each way — we don't pick a winner.

The case for

Advocates of this shift welcome AI weather models as a genuine leap forward in a field where accuracy saves lives and money: faster, higher-resolution and more frequently updated forecasts mean better warnings for storms, floods and heatwaves, achieved at a fraction of the computing cost of traditional supercomputer simulations. They argue that independent, transparent benchmarking such as Operational WeatherBench keeps the competition honest, and that if a well-resourced private lab can outperform government agencies, the sensible response is to adopt or license the better tool rather than cling to institutional pride, ultimately benefiting the public through Search, Maps and Gemini at no extra cost to users.

The case against

Others, while not disputing the benchmark results, urge caution about ceding such critical public infrastructure to a private company's proprietary model. They point out that national weather services carry statutory duties, public accountability and continuity obligations that a commercial product does not, and that dependence on a single corporation's black-box system for life-safety forecasting raises legitimate questions about transparency, long-term access, pricing power and what happens if incentives or ownership change. For them, government forecasters remain essential as an independently verifiable, publicly accountable backbone, even if AI tools are used to complement rather than replace that capability.

Full account

Google has unveiled WeatherNext 3, a new artificial intelligence weather forecasting system developed jointly by Google DeepMind and Google Research, which the company says produces sharper, more frequent and more accurate forecasts than its predecessor and, in independent testing, than established government meteorological services. The system has reportedly topped Operational WeatherBench, a benchmarking tool built by the forecasting startup Brightband that compares AI and traditional models on measures such as temperature, wind speed and humidity, outperforming rival AI systems from Microsoft, Nvidia and the European Centre for Medium-Range Weather Forecasting (ECMWF), as well as conventional forecasts issued by the ECMWF and the US National Weather Service.

Conventional forecasting has long depended on government-run supercomputers that simulate atmospheric physics by solving vast systems of equations, an approach that is highly accurate but computationally expensive and relatively slow to update. Machine-learning alternatives, which have proliferated since the ECMWF released decades of historical weather data in 2018, instead learn statistical patterns from past observations, allowing forecasts to be generated far more quickly. Google says WeatherNext 3 pushes this further by drawing on live satellite observations rather than relying solely on the archived, standardised datasets that most AI weather models are trained on, refreshing its forecasts every hour instead of the six-hour cycle typical of earlier systems.

In practical terms, Google says the new model can resolve certain variables, such as temperature and atmospheric moisture, at a grid resolution of around five kilometres, compared with 25 kilometres for its previous model, WeatherNext 2 — a roughly fivefold sharpening that the company says is especially valuable for tracking fast-moving rain and snow systems. Reported gains in rainfall prediction vary depending on the comparison being drawn: one account cites precipitation forecasts up to 50 percent more accurate than prior benchmarks a day or more ahead, while another cites a 60 percent improvement in rainfall evaluation specifically against WeatherNext 2. Google says the use of satellite data should also benefit regions with sparse ground-based rain-gauge networks, particularly outside the US and Europe, where forecasting has historically been weaker.

Beyond day-to-day weather prediction, Google says WeatherNext 3 has been built to forecast conditions relevant to renewable energy generation, including wind speeds at the roughly 100-metre height of a turbine, a capability researchers linked to the company's own rising electricity demands from AI data centres. The model is also set to be woven into consumer-facing Google products, including Search, Maps and its Gemini assistant, as well as being made available to researchers and developers via Google's cloud platform. Coverage of the launch diverged somewhat in focus: one outlet emphasised the technical leap in resolution and its implications for renewable-energy planning, while another placed greater weight on the model's competitive benchmark results and its imminent rollout across Google's own products; the two accounts also gave differing job titles for the Google researchers quoted, Samier Merchant and Ferran Alet.

Where outlets differ

Precipitation accuracy improvement figures differ: one source cites up to 50% more accurate forecasts a day or more out, the other cites a 60% improvement specifically versus WeatherNext 2 — likely different baselines rather than a contradiction.

The Verge emphasises technical detail (resolution, satellite data use, renewable-energy/wind forecasting tied to Google's own energy needs) while TechCrunch emphasises competitive benchmarking against government forecasters and rollout into Google consumer products (Search, Maps, Gemini).

Job titles given for the Google researchers differ between outlets: Samier Merchant is described as a 'research engineer' by The Verge and a 'senior staff engineer' by TechCrunch; Ferran Alet is described as a 'research scientist' by The Verge and a 'staff research scientist manager' by TechCrunch.

Coverage

AI Art Business Celebrity Companies Culture Entertainment Research Science Software Technology

Read the full article at the source →