Editing Weather forecasting (section)

=== Artificial intelligence ===
Initial attempts to use [[artificial intelligence]] began in the 2010s. [[Huawei]]'s Pangu-Weather model, [[Google]]'s GraphCast, WindBorne's WeatherMesh model, [[Nvidia]]'s FourCastNet, and the [[European Centre for Medium-Range Weather Forecasts]]' Artificial Intelligence/Integrated Forecasting System, or AIFS all appeared in 2022–2023. In 2024, AIFS started to publish real-time forecasts, showing specific skill at predicting hurricane tracks, but lower-performing on the intensity changes of such storms relative to physics-based models.<ref name=":02">{{Cite web |last=Berger |first=Eric |date=June 3, 2024 |title=No physics? No problem. AI weather forecasting is already making huge strides. |url=https://arstechnica.com/ai/2024/06/as-a-potentially-historic-hurricane-season-looms-can-ai-forecast-models-help/ |access-date=June 6, 2024 |website=Ars Technica |language=en-us}}</ref>

Such models use no physics-based atmosphere modeling or [[large language model]]s. Instead, they learn purely from data such as the [[ECMWF re-analysis]] ERA5.<ref>{{Cite web |last=Setchell |first=Helen |date=February 19, 2020 |title=ECMWF Reanalysis v5 |url=https://www.ecmwf.int/en/forecasts/dataset/ecmwf-reanalysis-v5 |access-date=June 11, 2024 |website=ECMWF |language=en}}</ref> These models typically require far less compute than physics-based models.<ref name=":02" />

[[Microsoft]]'s Aurora system offers global 10-day weather and 5-day air pollution ([[Carbon dioxide|{{chem|CO|2}}]], [[NOx|NO]], [[Nitrogen dioxide|{{chem|NO|2}}]], [[Silicon dioxide|{{chem|SO|2}}]], [[Ozone|{{chem|O|3}}]], and particulates) forecasts with claimed accuracy similar to physics-based models, but at orders-of-magnitude lower cost. Aurora was trained on more than a million hours of data from six weather/climate models.<ref>{{Cite journal |last=Wong |first=Carissa |date=June 4, 2024 |title=Superfast Microsoft AI is first to predict air pollution for the whole world |url=https://www.nature.com/articles/d41586-024-01677-2 |journal=Nature |language=en |doi=10.1038/d41586-024-01677-2|pmid=38834696 |url-access=subscription }}</ref><ref>{{cite arXiv |last1=Bodnar |first1=Cristian |title=Aurora: A Foundation Model of the Atmosphere |date=May 28, 2024 |eprint=2405.13063 |last2=Bruinsma |first2=Wessel P. |last3=Lucic |first3=Ana |last4=Stanley |first4=Megan |last5=Brandstetter |first5=Johannes |last6=Garvan |first6=Patrick |last7=Riechert |first7=Maik |last8=Weyn |first8=Jonathan |last9=Dong |first9=Haiyu|class=physics.ao-ph }}</ref>

In 2024, a group of researchers at Google's DeepMind AI research laboratories published a paper in Nature to describe their machine-learning model, called GenCast, that is expected to produce more accurate forecasts than the best traditional weather forecasting systems.<ref>{{Cite journal |last=Price |first=Ilan | display-authors=etal | year=2025 |title=Probabilistic weather forecasting with machine learning  |journal=Nature |volume=637  |issue=8044 |pages=84–90 |language=en |doi=10.1038/s41586-024-08252-9 |pmid=39633054 |pmc=11666454 |bibcode=2025Natur.637...84P }}</ref>

In a study conducted using the AIFS, Lang et al. (2024) presented 30-day ensemble simulations of the Madden-Julia Oscillation.''<ref>{{Citation |last1=Lang |first1=Simon |title=AIFS-CRPS: Ensemble forecasting using a model trained with a loss function based on the Continuous Ranked Probability Score |date=2024 |arxiv=2412.15832 |last2=Alexe |first2=Mihai |last3=Clare |first3=Mariana C. A. |last4=Roberts |first4=Christopher |last5=Adewoyin |first5=Rilwan |last6=Bouallègue |first6=Zied Ben |last7=Chantry |first7=Matthew |last8=Dramsch |first8=Jesper |last9=Dueben |first9=Peter D.}}</ref>