Forecasting
4 labs · 11 models and products Theme: Companies · Generated from this page's data. Click a name to open its panel on the Labs page. Records resting on one source are tagged "one source".AI forecasting uses machine learning to predict what happens next, most often the weather, and more recently world events. Machine-learning weather models are being tested alongside the physics-based models forecasters have long used; the Met Office says they can give useful forecasts faster and with less computing power. Google DeepMind's GraphCast and GenCast make global forecasts up to 10 and 15 days ahead. The Met Office, with the Alan Turing Institute, is testing its own model, FastNet, which is not yet in operational use. Mantic, a London start-up, builds an AI system that forecasts world events (per the site).
Context
- Google DeepMind's GraphCast beat the most accurate operational forecasting systems on 90% of 1,380 test measures, its authors report. 24 December 2022 (arXiv v1) arxiv.org/abs/2212.12794
GraphCast significantly outperforms the most accurate operational deterministic systems on 90% of 1380 verification targets
- Google DeepMind's GenCast, published in Nature, was more accurate than Europe's main ensemble forecast (ENS) on 97.2% of targets, up to 15 days ahead. 4 December 2024 deepmind.google/blog/gencast-predicts-weather-an
GenCast was more accurate than ENS on 97.2% of these targets
- The Met Office and the Alan Turing Institute launched a partnership on AI for weather prediction in 2023. 2023 (accessed 31 October 2026) metoffice.gov.uk/research/approach/collaboration
launched the Artificial Intelligence for Numerical Weather Prediction (AI4NWP) Partnership
- Met Office research found FastNet's accuracy comparable to the Met Office Global Model on several measures; the Met Office says it is not yet operational. 8 July 2026 metoffice.gov.uk/about-us/news-and-media/media-c
comparable to the Met Office's Global Model
- The Met Office runs FastNet experimentally in near real time, for research and comparison only. accessed 31 October 2026 metoffice.gov.uk/research/approach/collaboration
The Met Office is running the experimental Artificial Intelligence-based FastNet weather forecast system in near real-time.
- Mantic, a London start-up co-founded by former Google DeepMind researcher Toby Shevlane and Ben Day, emerged from stealth to forecast world events (trade). 27 August 2025 tech.eu/2025/08/27/mantic-which-claims-to-foreca
building a cutting-edge AI system that can predict the complex world of human affairs
Key papers
- GraphCast: Learning skillful medium-range global weather forecasting arXiv preprint · 24 December 2022
- GenCast: Diffusion-based ensemble forecasting for medium-range weather arXiv preprint · 25 December 2023
Labs Round 8 (31 Oct 2026): introduction and facts as the researchers wrote them; "(per the site)" marks a sentence resting on this site's own description of that lab.
| Record | What it is | Place | Links | Confidence |
|---|---|---|---|---|
| Aardvark Weather (end-to-end AI weather prediction)Chips, models and products | AI system that goes straight from raw weather observations to forecasts, replacing the traditional forecasting pipeline with one model. research (Nature paper)Key facts Published in Nature vol. 641, issue 8065, pp. 1172–1179 (2025); built by the Alan Turing Institute with the University of Cambridge and Microsoft Research; local station forecasts skilful up to 10 days ahead; Turing states it is "tens of times faster" and uses "thousands of times less computing power", running on a regular desktop computer; access: paperMade by The Alan Turing Institute (AI assurance and security work) · 2025 (Nature vol. 641; Turing blog 20 Mar 2025) | London | turing.ac.uk/science-innovation/environment-and- · turing.ac.uk/news/publications/end-end-data-driv | lab's own pages |
| AIFS ENS (ECMWF machine-learned ensemble forecast)Chips, models and products | ECMWF's machine-learned ensemble forecast: 51 stochastic forecasts per run from a 229-million-parameter model trained on reanalysis and operational analyses. operationalKey facts 229m parameters; about 30 km; 50 perturbed members plus 1 control; trained on 38 years of ERA5 (1979–2017) and 8 years of IFS operational analyses (2016–2023); up to 25% better than IFS ENS, mainly upper-air variables; authors Simon Lang and Linus Magnusson (ECMWF Newsletter 185)Made by European Centre for Medium-Range Weather Forecasts (ECMWF): machine-learned forecasting (AIFS) · 2025-10-28 (newsletter date; start date not given) | Reading | ecmwf.int/en/newsletter/185/earth-system-science | one source |
| AIFS Single (ECMWF Artificial Intelligence Forecasting System)Chips, models and products | ECMWF's first operational machine-learned weather model, run alongside its physics-based forecasts and trained on past weather data. operationalKey facts Operational 25 Feb 2025; up to 20% gains on many measures incl. tropical cyclone tracks; about 1,000 times less energy per forecast; 28 km grid (ECMWF, 25 Feb 2025); AIFS v1.1 operational since Feb 2025; AIFS v2 fine-tuned with prototype 50r1 data, operational date not stated (ECMWF blog, 11 May 2026)Made by European Centre for Medium-Range Weather Forecasts (ECMWF): machine-learned forecasting (AIFS) · 2025-02-25 | Reading | ecmwf.int/en/about/media-centre/news/2025/ecmwfs · ecmwf.int/en/about/media-centre/aifs-blog/2026/f | lab's own pages |
| Anemoi (open-source machine-learning weather framework)Chips, models and products | Open-source framework for building machine-learned weather models, first developed at ECMWF and now co-developed with European national weather services. in use (open source)Key facts Covers dataset curation, model training, graph selection and inference; used in ECMWF's Machine Learning Project with 17 national weather services, led by MET Norway and MeteoSwiss, launched 2024 under EUMETNET E-AI (ECMWF blog, 14 Jul 2026)Made by European Centre for Medium-Range Weather Forecasts (ECMWF): machine-learned forecasting (AIFS) · 2026-07-14 | Reading | ecmwf.int/en/about/media-centre/aifs-blog/2026/b | one source |
| DGMR precipitation nowcasting model (with the Met Office)Chips, models and products | Rain-forecasting model built with the Met Office that predicts the next 90 minutes of rainfall from radar. research (paper; Nature)Key facts Announced 29 Sep 2021; uses past 20 minutes of radar to predict next 90 minutes; more than 50 Met Office meteorologists rated it first choice in 89% of cases; Nature paper s41586-021-03854-z; access: paper onlyMade by Google DeepMind · 2021-09-29 | London | deepmind.google/blog/nowcasting-the-next-hour-of | one source |
| European Centre for Medium-Range Weather Forecasts (ECMWF): machine-learned forecasting (AIFS)Networks, bodies, events and media | European weather centre hosted in Reading, UK, that trains and runs its own machine-learned forecasting models, AIFS Single and AIFS ENS.Key facts AIFS Single operational 25 Feb 2025: gains of up to 20% on many measures, about 1,000 times less energy per forecast, 28 km grid (ECMWF); AIFS ENS: 229m parameters, up to 25% better than the physics-based IFS ENS mainly for upper-air variables (ECMWF Newsletter 185, Oct 2025); 11 May 2026: stopped real-time runs of four external ML models (Pangu-Weather, GraphCast, Aurora, FourCastNet) with IFS Cycle 50r1; AIFS v2 fine-tuned on prototype 50r1 data; ML project with 17 national weather services (ECMWF blog, 14 Jul 2026); 2025–2034 Strategy approved Dec 2024Models and products (3)
| Reading | ecmwf.int/en/about/media-centre/news/2025/ecmwfs · ecmwf.int/en/about/media-centre/news/2025/new-10 | verified |
| FastNet (Met Office and Alan Turing Institute)Chips, models and products | A machine-learning weather model built by the Met Office and the Alan Turing Institute to forecast global weather up to two weeks ahead. in development (not operational)Key facts Versions: v1 (global model built and trained), v1.1 (more physical realism), v2 in development for global and UK regional forecasts, built in the Anemoi framework. Research published 8 Jul 2026 (American Meteorological Society journal, article ID AIES-D-25-0090.1): accuracy comparable to the Met Office Global Model and better on some metrics; uses a modified spherical harmonic loss; tested on Hurricane Ian (2022) and Storm Ciarán (2023). Trained on decades of historical atmospheric data. Being evaluated beside existing systems; not yet operational.Made by The Alan Turing Institute (AI assurance and security work) · 2026-07-08 | metoffice.gov.uk/research/approach/modelling-sys · metoffice.gov.uk/about-us/news-and-media/media-c | lab's own pages | |
| GenCastChips, models and products | An AI weather model that gives a range of possible forecasts up to 15 days ahead, including extreme-weather risk. released (open code and weights)Key facts Theme: Forecasting. Nature paper 4 Dec 2024; code and weights released. Figures (4 Dec 2024): more accurate than ECMWF's ENS on 97.2% of targets; one 15-day forecast in 8 minutes on a TPU v5. Predecessor: GraphCast.Made by Google DeepMind · 2024-12-04Read the paper | London (EC2M 4AG) (via Google DeepMind) | deepmind.google/blog/gencast-predicts-weather-an · deepmind.google/blog/page/8/ · nature.com/articles/s41586-024-08252-9 | one source |
| GraphCastChips, models and products | An AI weather model that makes 10-day global forecasts in under a minute. released (open code)Key facts Theme: Forecasting. Science paper 14 Nov 2023; model code open-sourced. Figures (14 Nov 2023): more accurate than the leading conventional system (HRES) on more than 90% of 1,380 test variables and lead times; a 10-day forecast takes under a minute on one TPU v4. Succeeded by GenCast (Dec 2024) and WeatherNext 2/3.Made by Google DeepMind · 2023-11-14Read the paperKey facts and news (1)News
Labs Round 8 (31 Oct 2026), from the researchers' rows; one-source items are marked. | London (EC2M 4AG) (via Google DeepMind) | deepmind.google/blog/graphcast-ai-model-for-fast · deepmind.google/blog/page/9/ | one source |
| ManticCompanies | London start-up building an AI system that forecasts world events.Profile
Labs Round 7 deep profile (30 Oct 2026), as the researchers wrote it; confidence: single_source (no primary source opened; mantic.com blocked). Sources Models and products (1)
Key facts and news (4)
Labs Round 8 (31 Oct 2026), from the researchers' rows; one-source items are marked. | Kensal Road, London (W10 5BN) | Companies House · tech.eu/2025/08/27/mantic-which-claims-to-foreca · find-and-update.company-information.service.gov. · find-and-update.company-information.service.gov. · uktech.news/ai/ai-startup-that-predicts-world-ev | one source |
| Mantic forecasting system (Scan, Forecast, Monitor)Chips, models and products | Mantic's AI that gives probabilities, with reasoning and sources, for future world events and updates them as news changes. live (product, by demo)Key facts Product modules Scan, Forecast and Monitor; site states it beat all 676 human forecasters in the Metaculus Cup (Summer 2025 label); Polymarket backtest calibration error 0.022 vs 0.058 for Polymarket on markets with $1k+ traded; blog (5 Oct 2026) reports a 45.12% forecast for Lula against a 45.16% result; access: product by demoMade by Mantic · — (launch date not stated) | London | mantic.com/ · mantic.com/news | lab's own pages |
| Met Office AI research (National Capability AI Programme)Schemes, programmes and AI lines | The UK's national weather service runs an AI research programme to test machine-learning weather and climate forecasting beside its physics models.Key facts Programmes named: National Capability AI Programme (NCAIP) and AI4Climate, funded by DSIT (Met Office blog, 4 Feb 2026). AI for Numerical Weather Prediction (AI4NWP) partnership with the Alan Turing Institute launched 2023; current focus is co-developing FastNet (BIZ-LAB-6008). AI research themes: observations, simulation, analysis, products and warnings. Chief AI Officer: Prof Kirstine Dale (8 Jul 2026). Operational forecasting model (Momentum) remains physics-based. Headcount not published.Profile
Labs Round 8 deep profile (31 Oct 2026), as the researchers wrote it; confidence: one source (Met Office pages; the Turing role is confirmed through its named mission director). Sources Models and products (1)
Key facts and news (1)
Labs Round 8 (31 Oct 2026), from the researchers' rows; one-source items are marked. | metoffice.gov.uk/blog/2026/how-does-the-met-offi · metoffice.gov.uk/research/approach/collaboration · deepmind.google/blog/nowcasting-the-next-hour-of · metoffice.gov.uk/about-us/news-and-media/media-c | one source | |
| Open Climate FixCompanies · also on Start-ups | A not-for-profit that builds open machine-learning forecasts of solar power, used by Britain's grid operator to balance the system.Models and products (1)
Key facts and news (3)
Labs Round 8 (31 Oct 2026), from the researchers' rows; one-source items are marked. | County Hall, London (SE1 7PB) | Companies House · openclimatefix.org/insights/neso-adopts-ai-solar · pv-magazine.com/2025/11/07/ai-powered-solar-fore · find-and-update.company-information.service.gov. · digicatapult.org.uk/about/press-releases/post/la | verified |
| WeatherNext 2Chips, models and products | Google DeepMind's November 2025 weather forecasting model. releasedKey facts Theme: Forecasting. Announced November 2025 (index). Used to support the US National Hurricane Center on Hurricane Melissa (post 16 May 2026); cyclone forecasting result 6 Aug 2026.Made by Google DeepMind · 2025-11 | London (EC2M 4AG) (via Google DeepMind) | deepmind.google/blog/page/5/ · deepmind.google/blog/rss.xml | lab's own pages |
| WeatherNext 3Chips, models and products | Google DeepMind's AI global weather model, which produces a new forecast every hour and feeds Google Search, Maps and Gemini. releasedKey facts New forecast every hour; temperature and humidity at 5 km, wind at 10 km; ensemble model taking satellite imagery directly; used by the US National Hurricane Center for Hurricane Melissa's Jamaica landfall; offered via BigQuery, Earth Engine, Google Maps Platform and Cloud Storage. UK scientists offered priority access to WeatherNext (Dec 2025)Made by Google DeepMind · — (no date on page) | London | deepmind.google/science/weathernext/ · deepmind.google/blog/strengthening-our-partnersh | one source |