ECMWF Reports AI Forecast Gains of Up to 20 Percent on Surface Temperature Over Physics, While IMD's 2026 Season Report Does Not Mention the AI Monsoon Forecast It Launched on 12 May

The European Centre for Medium-Range Weather Forecasts (ECMWF) put an AI ensemble forecast into operation on 1 July 2025, 51 forecasts at a time, and says it beats the centre’s physics-based models by up to 20 percent on surface temperature. The release does not name the score or the model version behind that figure. Fifteen months later the published evidence on AI weather models comes in three kinds: numbers from the builders, numbers from outside groups, and, for India’s first AI monsoon forecast, a launch release with no accuracy figure in it.

Europe’s AI ensemble runs next to the physics model

An ensemble is a set of forecasts started from slightly different conditions, so the spread between them shows how sure the forecast is. ECMWF’s AI Forecasting System (AIFS) went operational as a single forecast at the end of February 2025 and as the 51-member ensemble, AIFS ENS, on 1 July.

ECMWF’s release gives AIFS ENS a grid of 31 km against 9 km for its physics-based ensemble. It still relies on the physics system’s data assimilation, the step that blends observations into a starting picture of the atmosphere. ECMWF says it runs more than 10 times faster on roughly 1,000 times less energy. Florian Pappenberger, the centre’s Director of Forecasts and Services, said: “We see the AIFS and IFS as complementary.”

Where ECMWF’s own scorecard shows losses

In the October 2025 issue of its newsletter, ECMWF’s Simon Lang and Linus Magnusson compare AIFS ENS with the physics ensemble, IFS ENS, against the centre’s analyses and against weather-station and weather-balloon observations. Improvements reach up to 25 percent for upper-air variables. Two-metre temperature scores better at every lead time, and 24-hour rainfall scores better in the first half of the forecast. IFS ENS is the more skilful on 10-metre wind speed against station data, and AIFS ENS is worse for temperature at 100 hPa, a level high in the atmosphere.

Two April 2025 cases are in the same article. During Storm Hans on 16 and 17 April, up to 446 mm of rain fell in 24 hours in northern Italy and southern Switzerland. Both ensembles reproduced the pattern and both underestimated the peak, with the AI ensemble further off. The authors name grid size, about 30 km against about 9 km, as a likely factor. On 30 April a heatwave hit France, and the physics ensemble underestimated it even at short range. The AI ensemble came closer, and ECMWF says it is investigating why the physics system fell short.

On 12 May 2026 ECMWF switched on AIFS v2 together with a new physics cycle, 50r1. The May 2026 newsletter reports that significant wave height errors fell by around 10 percent against IFS 50r1. It also says more realistic vertical air motion “did not translate directly into better verification scores.” ECMWF built and scored both systems, so these tables come from one organisation.

Google, Microsoft and NVIDIA publish their own numbers

Google DeepMind’s GenCast, described in a December 2024 post alongside a Nature paper, was more accurate than ECMWF’s physics ensemble on 97.2 percent of 1,320 combinations of variable and lead time, and on 99.8 percent beyond 36 hours. DeepMind trained it on data up to 2018 and tested it on 2019. On 17 November 2025 Google announced WeatherNext 2, which it says beats its earlier WeatherNext model on 99.9 percent of variables and lead times from 0 to 15 days. That comparison is against Google’s own previous model, not an outside system. DeepMind’s separate cyclone model and its Nature paper of 6 August 2026 are covered in our article on the WeatherNext cyclone results.

Microsoft’s Aurora, published in Nature on 21 May 2025, beat ECMWF’s high-resolution physics forecast, IFS HRES, on 92 percent of targets over 10 days at 0.1 degree grid spacing. It also beat seven operational centres on 100 percent of targets for five-day cyclone tracks, on 2022 and 2023 storms. The authors say HRES beats Aurora on many targets at the shortest lead times, that Aurora still needs a traditional data-assimilation system for its starting conditions, and that an ensemble version is not yet built. Several authors work for Microsoft or hold its stock, and the paper says so.

NVIDIA’s Earth-2 family of open models, announced on 26 January 2026, says its Medium Range model outperforms leading open models on the most common variables on “standard benchmarks.” The post does not say which benchmarks. Its one quantified user result is from Amir Givati, director of the Israel Meteorological Service, who reports a 90 percent cut in compute time at 2.5 km against a classic physics model run on a CPU cluster.

Outside tests find the gap at the extremes

Zhongwei Zhang, Erich Fischer, Jakob Zscheischler and Sebastian Engelke, writing in Science Advances in 2026, scored GraphCast, Pangu-Weather and Fuxi, three deterministic AI models, against ECMWF’s HRES on record-breaking heat, cold and wind. HRES had lower error across nearly all lead times. The AI models underestimated both how often records occur and how large they are, and they under-forecast hot records while over-forecasting cold ones. The authors write that these models need more verification before they are relied on alone for early warning. Their August 2025 preprint says most of its analysis rests on one test year, 2020, and that probabilistic models were not tested.

A PNAS paper from May 2025 by Y. Qiang Sun, Pedram Hassanzadeh and colleagues retrained the FourCastNet model with every Category 3 to 5 tropical cyclone of 1979 to 2015 removed, then tested it on 20 Category 5 storms of 2018 to 2023. It did not forecast their intensification. For Hurricane Lee the lowest predicted pressure never fell below 980 hPa, against 960 hPa in the reference data. The authors say the study covers one model and one kind of extreme event.

On physical consistency, the same group reports that FourCastNet’s wind and pressure fields violate gradient-wind balance, a basic link between the two. ECMWF’s newsletter lists related flaws in AIFS ENS version 1, including odd pressure values over mountains and faint spurious rain in dry regions.

A 2025 trial in India that did publish a score

In 2025 India’s Ministry of Agriculture and Farmers’ Welfare sent weekly probabilistic monsoon-onset forecasts to 38 million farmers in 13 states. A preprint of 9 March 2026 by Colin Aitken and colleagues, with Michael Kremer, Pedram Hassanzadeh and William Boos among the authors, describes the method. It blends Google’s NeuralGCM and ECMWF’s AIFS with a statistical model of what farmers expect from past rain-gauge records of the India Meteorological Department (IMD). A companion preprint of 3 February 2026 says AI models skillfully predict an onset index weeks ahead when tested out of sample.

The numbers come from leave-one-year-out tests over 2000 to 2024. At one week, the blend’s Brier score, an error measure for probability forecasts, improved by roughly 15 percent on the farmer-expectations model alone and 25 percent on plain climatology. Skill fell with lead time but stayed positive to four weeks. In 2025, when the monsoon stalled for more than two weeks from 29 May, the blend scored about 20 percent better than climatology.

The authors list caveats. The test years 2000 to 2018 overlap the AI models’ training years, which could overstate accuracy. Picking the best-performing models afterwards can bias scores upward. Only 28 grid cells, each 2 degrees wide, received forecasts in 2025.

India’s 2026 launch and what the season report covers

On 12 May 2026 Jitendra Singh, Minister of State for Earth Sciences, launched two products from the Ministry of Earth Sciences, built jointly by IMD, the Indian Institute of Tropical Meteorology (IITM) in Pune and the National Centre for Medium Range Weather Forecasting (NCMRWF). The Press Information Bureau (PIB) describes the first as IMD’s first AI-driven system. It gives probabilistic forecasts of monsoon progression every Wednesday, up to four weeks ahead, for farmers in 16 states and more than 3,000 sub-districts. It combines AI models, extended-range prediction systems and statistical techniques.

The second is a pilot for Uttar Pradesh that produces rainfall forecasts at 1 km up to 10 days ahead. It uses AI downscaling of rain-gauge, weather-station, Doppler radar and satellite data. The PIB release gives no accuracy figure for either product. The only accuracy figures in it describe conventional forecasting, such as the minister’s claim of nearly 40 percent better accuracy on severe weather over the past decade. How IMD’s cyclone landfall errors have moved is laid out in our piece on India’s cyclone warnings.

On 23 July 2026 the minister told the Rajya Sabha in a written reply that the global models running in real time are IMD’s GFS and NCMRWF’s Mithuna, both at 12 km, and that the Bharat Forecast System runs at 6 km. The reply calls these numerical weather prediction systems.

IMD’s season report of 30 September 2026 scores the seasonal forecast of 90 to 92 percent of the long period average against the outcome of 87 percent. It records that the Kerala onset forecast of 26 May, with a margin of four days, missed the actual date of 4 June, only the second miss since the operational onset forecast began in 2005. How the season unfolded is in our article on the 2026 monsoon. The report does not mention the AI monsoon-advance product.