AI and the future of weather forecasting
Emerging technologies and trends
Over the past five years, artificial intelligence has increasingly been applied to weather forecasting, offering faster and often more scalable alternatives to traditional numerical methods.
While conventional forecasting methods remain essential to global meteorological operations, recent developments suggest that AI could play a complementary—and in some cases, transformative—role in the future of weather prediction.
This post explores how AI-based models work, how they compare to traditional systems, and what real-world benefits and challenges they present.
Traditional forecasting methods
For decades, weather forecasting has relied on numerical weather prediction (NWP) models. These physics-based systems simulate atmospheric processes using mathematical equations and are supported by high-performance computing infrastructure. Models such as the European Centre for Medium-Range Weather Forecasts (ECMWF) IFS and NOAA’s Global Forecast System (GFS) are widely used and provide high-resolution, skillful forecasts.
However, these systems come with considerable computational costs. High-resolution global forecasts can take several hours to compute and require substantial supercomputing resources. This has prompted researchers and agencies to explore alternative approaches, including machine learning.
How AI-based forecasting works
AI weather forecasting models use machine learning techniques, particularly deep learning, to identify patterns in historical weather data. Once trained, these models can generate forecasts directly, often in a fraction of the time required by traditional methods.
Most models use historical reanalysis datasets like ERA5, which combine past observations from satellites, radars, and weather stations into a consistent global archive. These datasets provide the input needed to train deep neural networks that approximate the dynamics of the atmosphere.
AI models do not solve the governing equations of motion explicitly; instead, they predict future states based on the statistical relationships learned during training. In practice, this allows for faster generation of forecasts and, in many cases, competitive levels of accuracy.
Key AI weather models developed since 2020
Several organizations have introduced AI-based weather forecasting models with promising results:
GraphCast (Google DeepMind, 2023)
GraphCast uses a graph neural network architecture trained on over 40 years of ERA5 data. It generates 10-day global forecasts at 0.25° resolution and has demonstrated comparable or higher accuracy than ECMWF’s high-resolution model across a broad range of metrics. The model is capable of running forecasts in under one minute on a single TPU.
FourCastNet (NVIDIA, UC Berkeley, and collaborators, 2022–2024)
This model applies a Fourier neural operator to learn atmospheric dynamics. FourCastNet also uses 0.25° resolution and can produce a seven-day forecast in approximately two seconds on a high-end GPU. It has shown strong performance on both large-scale and small-scale variables, including precipitation.
Pangu-Weather (Huawei, 2023)
Pangu-Weather utilizes a three-dimensional transformer architecture. It forecasts global weather at hourly intervals for up to 10 days, using a method called hierarchical temporal aggregation to reduce error. In benchmark comparisons, it has outperformed both ECMWF’s IFS and FourCastNet in several key areas.
GenCast (Google DeepMind, 2024)
Building on GraphCast, GenCast introduces probabilistic forecasting using diffusion models. Rather than producing a single forecast, it generates ensembles of possible outcomes. Tests have shown it can outperform ECMWF’s operational ensemble system on the majority of target metrics.
Aardvark Weather (Independent researchers, 2025)
This experimental model represents a different approach by functioning as a fully end-to-end system. It uses raw observations—including radar and satellite imagery—and outputs forecasts without relying on numerical data assimilation or physics-based modules. Initial results suggest it performs at or above the level of traditional operational models, but more evaluation is needed.
Comparing AI and traditional forecasting
While AI models currently operate at coarser resolutions compared to high-end NWP models, they deliver considerable speed and efficiency advantages. These features make them especially useful for applications where frequent, low-cost forecasting is needed, or where ensemble predictions are valuable.
However, many AI models still rely on initial conditions derived from traditional systems. Data assimilation—the process of integrating real-time observations into forecasts—is one area where NWP maintains a clear advantage, though research is ongoing to close this gap.
Use cases and applications
AI-based forecasting systems are being explored for a range of practical uses:
Disaster preparedness: Improved forecasts for tropical cyclones, heatwaves, and atmospheric rivers may help provide earlier warnings.
Aviation and logistics: Rapid updates and ensemble forecasts could support more efficient routing and scheduling.
Agriculture: Better short- and medium-range weather information can inform planting, harvesting, and irrigation decisions.
Energy management: Forecasts of wind and solar potential can support grid operations and renewable energy integration.
Climate risk modeling: Low-cost ensemble forecasting allows for broader simulation of weather-related risks in insurance and infrastructure planning.
Adoption around the world
Interest in AI forecasting is growing globally. Key developments include:
Europe: ECMWF is evaluating GraphCast, Pangu-Weather, and FourCastNet in operational settings. The UK Met Office is also investing in AI-enhanced systems.
Asia: China’s meteorological agencies have developed several AI-based models for short- and medium-range forecasting. Japan and India are actively exploring AI applications in typhoon tracking and monsoon prediction.
North America: NOAA and NASA have launched AI-focused research initiatives, though operational models remain in early stages.
These models are also becoming more accessible. Open-source implementations and public data releases are allowing national weather services, researchers, and NGOs in lower-resource settings to run AI forecasts without access to supercomputers.
Looking ahead
AI in weather forecasting is still a developing field, but its progress over the past five years has been substantial. As resolution improves, data assimilation techniques mature, and probabilistic forecasting becomes more widely adopted, AI is likely to become an increasingly important part of the forecasting toolkit.
Rather than replacing traditional systems, AI models are currently best viewed as complementary tools—offering faster, more scalable alternatives for certain applications while relying on the strengths of physics-based models for others.
The landscape is evolving quickly, and continued collaboration between meteorologists, data scientists, and policymakers will be key to ensuring AI’s safe and effective integration into global forecasting systems.
📚 Bibliography
📘 Academic research
Lam, Remi, et al. “Learning Skillful Medium-Range Global Weather Forecasting.” Science, vol. 382, no. 6675, 2023, pp. 749–755. https://www.science.org/doi/10.1126/science.adi2336
Lam, Remi, et al. “GraphCast: Learning Skillful Medium-Range Global Weather Forecasting.” arXiv, 2022. https://arxiv.org/abs/2212.12794
Bi, Kaifeng, et al. “Pangu-Weather: A 3D High-Resolution Model for Fast and Accurate Global Weather Forecast.” arXiv, 2022. https://arxiv.org/abs/2211.02556
Cheng, Wenze, et al. “The Compatibility between the Pangu-Weather Model and Meteorological Operational Data.” arXiv, 2023. https://arxiv.org/abs/2308.04460
Price, Isaac, and Matthew Willson. “GenCast Predicts Weather and the Risks of Extreme Conditions with State-of-the-Art Accuracy.” Google Deep Mind Blog, 2024. https://deepmind.google/discover/blog/gencast-predicts-weather-and-the-risks-of-extreme-conditions-with-sota-accuracy/
Vonich, Paul T., and Gregory J. Hakim. “Testing the Limit of Atmospheric Predictability with a Machine Learning Weather Model.” arXiv, 2025. https://arxiv.org/abs/2504.20238
📗 Technical documentation and model releases
Huawei Cloud. “Pangu-Weather AI Model Published in Nature.” Huawei Newsroom, 6 July 2023. https://www.huawei.com/en/news/2023/7/pangu-ai-model-nature-publish
Huawei Cloud Blog. “Pangu-Weather from Huawei Cloud Outperforms NWP Methods in Terms of Accuracy for Medium-Range Forecast.” Huawei Cloud Blog, 7 July 2023. https://www.huaweicloud.com/intl/en-us/about/blogs/20230707.html
Huawei Cloud. “Huawei Cloud and Shenzhen Meteorological Bureau Announce Regional AI Model.” Huawei Newsroom, 23 Mar. 2024. https://www.huawei.com/en/news/2024/3/pangu-weather
📙 Media coverage
Cookson, Clive. “AI Outperforms Conventional Weather Forecasting Methods for First Time.” Financial Times, 14 Nov. 2023. https://www.ft.com/content/ca5d655f-d684-4dec-8daa-1c58b0674be1
Simonite, Tom. “Google DeepMind’s AI Weather Forecaster Handily Beats a Global Standard.” Wired, 14 Nov. 2023. https://www.wired.com/story/google-deepmind-ai-weather-forecast
Sample, Ian. “Google DeepMind Predicts Weather More Accurately than Leading System.” The Guardian, 4 Dec. 2024. https://www.theguardian.com/science/2024/dec/04/google-deepmind-predicts-weather-more-accurately-than-leading-system
Morton, Adam. “Cyclone Forecasting Boosted by Artificial Intelligence Offers Earlier Path Tracking.” The Guardian, 18 Jan. 2024. https://www.theguardian.com/australia-news/2024/jan/18/cyclone-forecasting-boosted-by-artificial-intelligence-offers-earlier-forecasts
Ducharme, Jamie. “How Meteorologists Are Using AI to Forecast Hurricane Milton and Other Storms.” TIME, 9 Oct. 2024. https://time.com/7081372/ai-hurricane-forecasting
de Miguel, Mercedes. “El Sistema que ‘Reinventa’ la Predicción Meteorológica.” El País, 10 Apr. 2025. https://elpais.com/tecnologia/2025-04-10/el-sistema-que-reinventa-la-prediccion-meteorologica-un-modelo-reivindica-ser-miles-de-veces-mas-preciso-y-rapido-que-el-actual.html
More news!
Lovable aims for $150M raise at nearly $2B valuation
Lovable, a fast-growing Swedish AI startup known for its text-to-app web development platform, is reportedly raising over $150 million at a nearly $2 billion valuation, according to the Financial Times. This follows a $15 million pre-Series A round in February led by Creandum. Accel is said to be leading the new round, with participation from Creandum and 20VC. Founded in 2023, Lovable reached $50 million in ARR within six months of launching its product in late 2023. The platform allows users to generate full web apps from prompts and recently launched a beta AI agent to automate coding tasks, using a usage-based credit pricing model.
Keep reading with a 7-day free trial
Subscribe to The PhilaVerse to keep reading this post and get 7 days of free access to the full post archives.