What is the AI weather big model? Full analysis of principles, mainstream models, and new energy applications

While supercomputers are still struggling to solve atmospheric equations, another "new gameplay" can already calculate global weather forecast fields for multiple days in the future within minutes, with accuracy comparable to traditional top-level models. This technological wave is reshaping the way we 'see the sky'.

If you ask what is the hottest keyword in the field of meteorology in the past two years, "AI weather big models" will definitely rank first. From short-term severe convective warning to medium - to long-term climate prediction, this new technological paradigm is quietly changing the entire chain capability of weather forecasting.

But what exactly is the AI weather model? What is the difference between it and the familiar 'traditional weather forecast'? Why can it be fast and accurate? This article takes you through it all at once - all key data comes from publicly available academic papers and industry publications.

To understand the AI weather model, we need to first look at the object it wants to "revolutionize" - traditional numerical weather forecasting (NWP).

The traditional approach is very "hardcore": scientists write atmospheric motion as a set of fluid dynamics equations, and then use supercomputers to cut the global atmosphere into countless small grids, step by step deducing future evolution. It has extremely rigorous physical logic and has been the cornerstone of weather forecasting for the past few decades.

But the cost is also very obvious: the computational cost is extremely high, and it consumes a huge amount of computing power. Taking the high-resolution model (IFS) of the European Centre for Medium Range Weather Forecasts (ECMWF) as an example, generating a 10 day global forecast requires running on a supercomputer for several hours. Moreover, in some small-scale and extreme weather scenarios, such as sudden thunderstorms and local rainstorm, the resolution limit of the equation will bring some limitations.

The AI meteorological big model has taken a different path: it does not start from physical equations, but "learns" statistical laws of atmospheric evolution from massive historical meteorological data.

Its working method is roughly to use the global meteorological status of the past few decades, including temperature, humidity, wind, pressure, etc., as training samples; Build deep learning networks to enable models to learn to 'see today's atmosphere and infer tomorrow's and the day after tomorrow's atmosphere'; When reasoning, directly take the current meteorological state as input and output the future meteorological field.

Compared to traditional modes, its most intuitive advantage is its extremely fast inference speed and low energy consumption. Taking Google GraphCast as an example, traditional numerical systems require "several hours and tens of thousands of processors" to generate a 10 day global forecast, while GraphCast can be completed in less than 1 minute on a single TPU, with energy efficiency about 1000 times that of traditional methods. Compressing the original hours of computing power expenditure into one minute is precisely the most fascinating aspect of the AI paradigm.

This technology is not just talk on paper. A number of mature models have already emerged globally, and many achievements have been published in top journals such as Nature and Science.

Google DeepMind, published in Science's GraphCast in 2023, is a 0.25 ° global medium-term forecasting model that can generate 10 day forecasts in less than 1 minute with a single TPU. It outperforms ECMWF HRES in 90.3% of testing metrics and has open-source code. Google DeepMind, published in Nature's GenCast in 2024, uses a diffusive probability set to generate over 50 possible 15 day paths, outperforming ECMWF ensemble in predicting ENS in 97% of indicators. The FourCastNet, launched by Nvidia and others in 2022, uses Fourier neural operators and has a global resolution of 0.25 °, which is the basic model of Nvidia's Earth-2 Earth platform. Huawei's Pangu Weather, published in Nature in 2023, is based on a 3D Earth Specific Transformer architecture customized for Earth meteorology. It completes a 24-hour global forecast in 1.4 seconds on NVIDIA V100 and is the first AI model invited by ECMWF to be featured on its experimental dashboard. Fudan University and Shanghai AI Lab's Fuxi, using the U-Transformer architecture with approximately 4.5 billion parameters, supports 15 day global forecasting. In the experimental forecast, it was predicted that Typhoon "Bebijia" would make landfall in Pudong, Shanghai in 2024 five days in advance. The Fengwu, launched by Shanghai AI Lab in 2023, achieved effective forecasting for over 10 days for the first time, generating 10 day global forecasts in 30 seconds, and the GHR version resolution was increased to about 9 kilometers. The "Wind" series of the China Meteorological Administration includes Wind Thunder, Wind Clear, and Wind Smooth, which can generate global forecast products with a resolution of 25 kilometers for the next 15 days in 3 minutes.

It is worth mentioning that in 2025, the China Meteorological Administration organized the "Artificial Intelligence Weather Forecasting Large Model Demonstration Plan", with 14 models participating in real-time forecasting demonstrations. In the comprehensive evaluation, Fuxi, Puyun, Fengwu, Fengqing, Pangu and other models performed well. Even ECMWF put its self-developed AIFS machine learning system into trial operation in February 2025, which was a trial operation and did not completely replace the original IFS formal business system. AI has become a standard direction for global meteorological agencies.

Many people are curious about what data AI weather models rely on to "learn" weather. Models with different positioning have significantly different training datasets, which can easily lead to misunderstandings. "All AI weather models use the same set of data" is a simplified statement and not accurate.

Large models for global medium-term forecasting ranging from 3 to 15 days are mostly trained using reanalysis datasets, with ERA5 being the most widely used. ERA5 is the fifth generation global atmospheric reanalysis data released by ECMWF through the Copernicus Climate Change Service (C3S). It has a spatiotemporal resolution of approximately 31 kilometers, or 0.25 °, with 137 vertical layers and hourly updates, covering the period from 1940 to 1950. It integrates multiple sources of observations such as satellites, sounding, ground observations, and aircraft reports to produce a set of spatiotemporal continuous and physically consistent global grid data, which is open source and free of charge through the C3S Climate Data Store. With its long time series, global coverage, and rich variables, it has become the training base that mainstream global models such as GraphCast, FourCastNet, Pangu Weather, and Fuxi rely on.

But ERA5 is not the only option. The training of AI weather models is a "multi-source combination", where reanalysis data is generally used as the backbone, and observation data is added for fine-tuning. Short term models do not even rely on reanalysis as the main force. Common data sources other than ERA5 and CMA-RA can be roughly classified into several categories based on their purposes.

Short term and imminent forecast, that is, 0-3 hour forecast, is to capture local strong convection such as rainstorm, thunderstorm and gale, relying on radar echo, meteorological satellite cloud image and minute level ground reality. Specifically, it includes the 6-minute composite reflectance puzzle of national Doppler radar, which is the core material for severe convective short-term AI and is not used for 10 day global forecasting; High density ground automatic stations, sounding and aircraft reports, mainly used for deviation correction and true value verification; Lightning location data, used for thunderstorm tagging. In terms of satellites, domestic models use high-frequency AGRI geostationary satellite data from Fengyun-4 FY-4A and FY-4B, including infrared and visible channels, 15 minute frequency, to enhance convective clouds and radiation, and make East Asian fine-tuning. Overseas, they rely on GOES, Himawari Sunflower Meteosat。 The GPM-IMERG global satellite precipitation inversion product often provides truth labels for precipitation models and new energy irradiation corrections.

In terms of domestic regional optimization, China has established its own global atmospheric reanalysis CMA-RA V1.5, with a horizontal resolution of 10 kilometers and hourly updates, covering 204 variables. According to public reports, this is the first set of 10 kilometer level global atmospheric reanalysis products internationally, and has been used as the training benchmark data for the "Artificial Intelligence Weather Forecasting Model Demonstration Program", supporting the research and evaluation of 13 large models and achieving domestic substitution of AI training data. Domestic models generally adopt a combination of global pre training and local fine-tuning, which adds CMA-RA, Fengyun satellite, and high-density station observations on top of ERA5 to correct systematic biases in the East Asian region.

Long term climate models will introduce climate datasets such as CMIP multimodal sets to expand their learning ability for decadal evolution. The typhoon specific model will also supplement the global tropical cyclone dataset, including path and intensity labels. Other global reanalysis is often used for assistance or comparison, such as NASA's MERRA-2, which is about 50 kilometers away and has rich aerosol variables; JRA-55 from the Japan Meteorological Agency, long time series, commonly used for scientific research verification in the Asian region; The previous generation domestically produced reanalysis CRA-40 by the China Meteorological Administration, covering approximately 30 kilometers from 1979 to 2018, has been replaced by CMA-RA V1.5. These three sets are mostly used for climate research, model comparison, and effectiveness verification, and are rarely used as the main training library for business level global models. In terms of benchmark and evaluation datasets, Google WeatherBench and WeatherBench2 are used to uniformly score and test AI weather models, and are not the original training sources; In the field of scientific research, there is also a publicly available SEVIR storm short-term reference dataset, which includes satellite and radar storm samples and is commonly used for short-term model training and testing.

If viewed by task, the main data is completely different. The global mid-term forecast for 3 to 15 days is usually based on ERA5 or CMA-RA V1.5 as the main training data, with auxiliary and fine-tuning data including Fengyun satellites, ground stations, and live observations. Short term forecast of strong convection from 0 to 3 hours, with radar mosaic and geostationary satellites as the main components, and reanalysis for background field reference only. Regional fine-tuning usually involves ERA5 pre training, followed by CMA-RA, Fengyun, and high-density stations. For long-term climate, the CMIP multimodal set is introduced, with ERA5 as the basic reference.

There are two key understandings here. One is that the backbone training base is not equal to all training data. Global models use ERA5 or CMA-RA to learn atmospheric evolution laws, and then add satellites and stations for fine-tuning and correction. Another issue is that the vast majority of mainstream models for 3 to 15 day global medium-term forecasts still use ERA5 or CMA-RA V1.5 as their backbone, while other data sources are mostly used for regional fine-tuning, short-term training, or model performance verification.

Another key distinction is that training data is not equal to business input. Further analysis of data, such as ERA5, is mainly used during the training phase of the model's "learning historical patterns"; When the real business is running, the AI model receives real-time numerical forecast initial fields, such as ECMWF HRES, CMA-GFS and other products, provided by observation driven data assimilation systems. It does not directly use reanalysis data for real-time forecasting. Confusing these two is one of the most common misconceptions in science popularization.

After discussing the advantages, it is necessary to clarify the limitations, which is also a part of responsible science popularization. It needs to be clarified that the AI weather model still relies on traditional NWP to provide initial atmospheric conditions, and it has not completely separated from the infrastructure of numerical forecasting. In terms of performance, it excels in global large-scale circulation and medium-term weather trends, but for local strong convection and small-scale extreme weather, radar and satellite observations are still needed as supplements. Due to the bias of the training data itself, the results directly output by the model may have systematic errors in some areas. When implementing business, it is often necessary to combine local observations for bias correction. To summarize its position in one sentence, it is a powerful accelerator and supplement to traditional modes, rather than a substitute.

The value of AI weather models ultimately lies in the industry. Among many scenarios, the new energy industry is one of its most imaginative application scenarios, and there are already real implementation cases. The power generation output of photovoltaics and wind power is highly dependent on meteorological factors such as radiation, wind speed, and temperature. With the help of AI meteorological models, it is possible to carry out wind and solar resource assessments, based on historical and predicted meteorological fields, to evaluate the endowment of regional solar and wind energy resources; It can conduct power forecasting, predict future power generation in advance, and help the power grid absorb new energy more smoothly; It can also support scheduling and trading, providing meteorological decision-making basis for power station scheduling and electricity market trading.

These abilities have been practically demonstrated in China. The "Yuheng" energy meteorological AI model in Guangxi is the first "physical and data dual engine driven" model in China, which updates real-time monitoring of new energy stations, wind and solar output forecasts, wind turbine icing risks, and other information in the entire region. After its implementation, the accuracy of wind and solar power prediction has significantly improved, promoting an annual increase of about 8 billion kilowatt hours of clean energy, reducing carbon emissions by about 6.8 million tons, and improving spot trading efficiency by 40%. The "Water, Wind, and Solar" prediction model for the Wujiang River Basin was developed by Huadian Wujiang, extending the forecasting time to 15 days. The accuracy of wind prediction reached 91.95%, and the illumination reached 94.54%, both higher than ECMWF's reference for the same period. It was selected as one of the first central enterprise AI strategic high-value scenarios by the State owned Assets Supervision and Administration Commission of the State Council. Its high-precision runoff prediction model was included in the International Telecommunication Union's "Artificial Intelligence for Good" case collection in 2026. The "meteorological big model combined with multi-source data fusion" system of Yunnan Power Grid has achieved an ultra short term accuracy rate of 92.46% and a short-term accuracy rate of 90.94% for new energy power prediction by 2025, supporting daily and real-time scheduling and water solar complementary optimization operation.

The forefront is also accelerating. Google's WeatherNext 3, released in September 2026, is specifically designed to optimize predictions for renewable energy production. It can predict wind speeds at a height of 100 meters, which corresponds to the height of the wind turbine hub, to accurately estimate wind power production. It also provides high-resolution cloud cover and solar radiation data to help estimate solar field illumination, with a maximum resolution of 5 kilometers and updates every hour. It is worth noting that directly applying the original model output is often insufficient, and localization correction needs to be combined with local measured data to meet the accuracy requirements of engineering business. This also means that obtaining high-resolution, locally correctable meteorological data, such as the services provided by the meteorological data open platform, is a key link in truly implementing AI model outputs into power plant scheduling and trading operations.

Looking ahead to the future, AI weather will not completely replace traditional numerical models, and the two are more complementary and integrated. The traditional model adheres to physical mechanisms to ensure that forecasts are "reasonable and reliable"; AI models are responsible for speeding up, correcting errors, reducing costs, and extending forecasting capabilities to more scenarios. With the continuous enrichment of observation data and the continuous iteration of algorithms, AI meteorological models will take root in more fields such as disaster prevention and reduction, new energy, agriculture, ecology, etc., and unleash greater value.

In a sense, we are already at the starting point of a transformation in the way we look at the sky, and this wave has just begun.