Explore the research

Weather and climate

Can AI make dependable forecasts and help people make better decisions about weather risks?

33%

field summary estimate

6 outcomes · Assessed November 7, 2026 · Low confidence

The field estimate averages its outcome scores. It is a reading aid, not a measured fraction of the field that AI manages.

How the estimate built up.

Each band is one outcome's share of the field estimate, so the top line is the estimate itself. Every field starts at 0% on December 31, 2022. Dates before the latest are retrospective estimates. Point at a band to see that outcome's score on each date.

Robustness20%
Forecast skill45%
Extreme-event performance40%
Decision benefit15%
Computation45%
Climate application validity35%

View exact historical values and reconstruction notes

What we measure

Each percentage is a judgment about progress toward independent advancement. Open the evidence to inspect the basis for it.

Climate application validity

35%estimated progress

What we track

Performance on withheld periods and regions, physical-consistency checks, and uncertainty under distribution shift; report these separately from weather scores.

Why this estimate?

Multidecadal atmospheric statistics support bounded climate simulation. Prescribed ocean conditions and uncertain extrapolation limit autonomous climate research credit.

Advancements

NeuralGCM reproduces multidecadal climate statisticsJuly 22, 2024. NeuralGCM reproduces multidecadal climate statistics with prescribed sea-surface temperatures.
ACE2 emulator reproduces 80 years of atmospheric variability at 1,500 years per day18 November 2024. The 450-million-parameter ACE2 emulator (1 degree, 6-hourly, eight vertical layers) runs about 1,500 simulated years per wall-clock day, exactly conserves global dry air mass and moisture, and reproduces atmospheric variability over the past 80 years, including the response to El Nino and global temperature trends. Its sensitivities to separately changed sea surface temperature and carbon dioxide are 'not entirely realistic'.
NVIDIA's cBottle generative model emulates global 5 km climate simulations10 May 2025. The paper 'Climate in a Bottle' describes a diffusion-based model that emulates global 5 km climate simulations and reanalysis (12.5 million pixels) using a coarse generator plus a 16x super-resolution stage. The authors report it passes tests including diurnal-to-seasonal variability, large-scale modes of variability, tropical cyclone statistics, and trends of climate change and weather extremes.
MIT study: linear pattern scaling beats deep learning for emulating temperatureAugust 26, 2025. An MIT study in the Journal of Advances in Modeling Earth Systems found that linear pattern scaling outperformed a deep-learning emulator on nearly all variables in a common climate benchmark. In a new evaluation with more data that accounts for natural variability, deep learning was slightly better for local precipitation, but linear pattern scaling remained more accurate for temperature.
NeuralGCM trained on satellite rainfall cuts multi-year precipitation error by 40%January 12, 2026. Google Research reported, with a paper in Science Advances, a NeuralGCM version at 280 km resolution trained on NASA satellite precipitation observations from 2001 to 2018. In multi-year simulations its average mean precipitation error was under 0.5 mm per day, a 40% average reduction versus leading models used in the latest IPCC report, and it outperformed the ECMWF model on most averaged precipitation measures over all 15 forecast days.

Read the dated findings and sources

Status: Climate simulation evidence with scope limits. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

July 22, 2024: NeuralGCM reproduces multidecadal climate statistics with prescribed sea-surface temperatures. Neural general circulation models, Nature.

Interpretation and tracking value

Prescribed ocean conditions are not a fully coupled climate projection. Weather skill cannot establish validity under future forcing outside the training distribution.

Next comparable measurement

Validate withheld decades/regions, conservation and forced responses; report out-of-distribution uncertainty and distinguish prescribed-boundary simulations from coupled projections. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Computation

45%estimated progress

What we track

End-to-end forecast latency, energy, and cost, including data ingestion and required upstream initialization.

Why this estimate?

Fast ensemble-member generation demonstrates strong repeatable computation. Upstream initialization, full energy costs and self-managed pipeline development remain gaps.

Advancements

FourCastNet generates a week-long global forecast in under 2 seconds22 February 2022. NVIDIA and collaborators reported that FourCastNet produces a week-long global forecast at 0.25-degree resolution in less than 2 seconds, orders of magnitude faster than ECMWF's IFS, while matching IFS accuracy at short lead times for large-scale variables. Predates the window; included as the landmark basis for fast AI forecasting.
GraphCast makes a 10-day forecast in under a minute on one TPU machineNovember 14, 2023. Google DeepMind stated that making 10-day forecasts with GraphCast takes less than a minute on a single Google TPU v4 machine.
GenCast produces a 15-day forecast in eight minutesDecember 4, 2024. One 15-day GenCast ensemble member takes about eight minutes on one TPU v5; members can run in parallel.
Aardvark Weather replaces the whole forecast pipeline with one AI model20 March 2025. Cambridge-led researchers reported in Nature that Aardvark Weather takes raw observations and outputs global and local forecasts with a single machine learning model, outperforming the US GFS on many variables using 10% of the input data, and running in minutes on a desktop computer.
ECMWF says AIFS ENS runs over 10 times faster with about 1,000 times less energy1 July 2025. On making its AI ensemble operational, ECMWF stated that AIFS ENS produces forecasts more than 10 times faster than the physics-based system and cuts energy use by approximately 1,000 times.
WeatherNext 2 generates forecasts 8 times faster, under a minute per predictionNovember 17, 2025. Google stated that WeatherNext 2 generates forecasts 8x faster than before, with each prediction taking under a minute on a single TPU, and produces hundreds of possible outcomes from one starting point.
NOAA's operational AIGFS uses 0.3% of the computing of the GFSDecember 2025 (article dated 18 December 2025). NOAA put three AI-driven global models into operation: AIGFS, the 31-member AIGEFS and the 62-member hybrid HGEFS. A 16-day AIGFS forecast uses 0.3% of the operational GFS computing resources and takes about 40 minutes; AIGEFS needs 9% of GEFS computing. AIGFS v1.0 has degraded tropical cyclone intensity forecasts.

Read the dated findings and sources

Status: Measured inference latency proxy. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

December 4, 2024: One 15-day GenCast ensemble member takes about eight minutes on one TPU v5; members can run in parallel. DeepMind: GenCast runtime and evaluation.

Interpretation and tracking value

Parallelism uses additional hardware. This excludes the full observation/initialization pipeline, training amortization and measured electricity cost.

Next comparable measurement

Time and meter the complete pipeline; specify member count, hardware, initialization dependency and cost/energy per delivered ensemble. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Decision benefit

15%estimated progress

What we track

Avoided losses or improved resource allocation from actual forecast-informed decisions against a prespecified operational baseline.

Why this estimate?

Wind-power prediction is a useful downstream proxy, but realized savings from actual decisions and autonomous policy improvement are not demonstrated.

Advancements

GenCast is evaluated for wind-power predictionDecember 4, 2024. GenCast evaluates wind-power prediction as a downstream application.
GiveDirectly pays Nigerian households ahead of floods using Google AI forecastsJune 27, 2025. GiveDirectly reported that in its 2024 Kogi State pilot, triggered by Google Flood Hub forecasts, 3,250 people received a $105 transfer between August 29 and October 6, before the October 21 flood peak, and over 4,600 were paid in total. All 53 targeted communities flooded. Reported food shortages fell from 97% to 7% of recipients; there was no comparison group.
AI monsoon-onset forecasts reach 38 million Indian farmers by SMSSeptember 15, 2025. India's agriculture ministry sent AI-based monsoon-onset forecasts to 38 million farmers, with nearly 1 million more reached by voice message in Odisha. The forecast blended Google's NeuralGCM and ECMWF's AIFS, had up to a 30-day lead time, and predicted a roughly three-week stall in the 2025 monsoon's progress. Effects on farmer decisions had not yet been measured; surveys were planned.

Read the dated findings and sources

Status: Application proxy; realized operational benefit unknown. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

December 4, 2024: GenCast evaluates wind-power prediction as a downstream application. GenCast, Nature.

Interpretation and tracking value

Prediction improvements do not establish realized avoided losses or better resource allocation from actual decisions; operational policies and constraints intervene.

Next comparable measurement

Run a prospective decision trial with frozen policies, logging actions and costs; report net benefit after forecast and response costs. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Extreme-event performance

40%estimated progress

What we track

Event recall, false-alarm rate, probability calibration, cyclone-track error, and useful warning lead time on prospectively observed events.

Why this estimate?

Historical extreme-event and cyclone evaluation supports useful forecasting breadth. Prospectively calibrated warnings and autonomous adaptation to new extremes remain incomplete.

Advancements

GraphCast predicted Hurricane Lee's Nova Scotia landfall about nine days aheadNovember 14, 2023. Google DeepMind reported that in September 2023 GraphCast predicted Hurricane Lee would make landfall in Nova Scotia about nine days in advance, while traditional forecasts settled on Nova Scotia only about six days ahead.
Google's AI flood model forecasts river floods up to 7 days ahead in 80 countriesMarch 20, 2024. With a paper in Nature, Google reported that its AI river-forecasting model provides flood information up to 7 days in advance, extended the average reliability of global nowcasts from zero to five days, and covers 80 countries where 460 million people live, including regions without streamflow gauges.
GenCast improves extreme-weather and cyclone-track forecastsDecember 4, 2024. The GenCast paper evaluates extreme weather and tropical-cyclone tracks and reports improvements over its ENS comparator.
Aurora beats seven forecasting centres on 2022-2023 tropical cyclone tracksMay 21, 2025. Microsoft reported that Aurora outperformed seven major forecasting centres on all cyclone track forecasts globally for the 2022-2023 season and beat the US National Hurricane Center on 5-day track forecasts, a first for a machine-learning model. It predicted Typhoon Doksuri's Philippine landfall four days in advance.
Google DeepMind cyclone model cuts 5-day track error by 140 km versus ECMWF ENSJune 12, 2025. Launching Weather Lab, Google DeepMind reported that on 2023-2024 North Atlantic and East Pacific storms its experimental cyclone model's 5-day track prediction was on average 140 km closer to the true location than ECMWF's ENS, and beat NOAA's HAFS on average intensity error. It generates 50 scenarios up to 15 days ahead, and the US National Hurricane Center is seeing live predictions.
Preliminary analysis ranks Google DeepMind's model best for 2025 Atlantic hurricanesNovember 9, 2025. A preliminary analysis by University of Miami researcher Brian McNoldy covering all 13 named Atlantic storms of 2025 found Google DeepMind's Weather Lab model was the best for both track and intensity, while the US GFS performed worst. The article notes official National Hurricane Center verification was still to come.
Nature paper: WeatherNext cyclone model gains a day or more of forecast lead time2026. The Nature paper 'Operational tropical cyclone forecasting with AI' (Alet et al.) reports that on 2023-2025 tropical cyclones the model's five-day ensemble-mean track error was 230 km, against 370 km for ECMWF ENS and 335 km for GenCast, an average lead-time advantage of a day or more, with lower intensity errors than NOAA's HAFS from 0.5 to 3.25 days.

Read the dated findings and sources

Status: Historical event-skill evidence; prospective baseline incomplete. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

December 4, 2024: The GenCast paper evaluates extreme weather and tropical-cyclone tracks and reports improvements over its ENS comparator. GenCast, Nature.

Interpretation and tracking value

Retrospective performance is not a prospective event recall/false-alarm baseline for current models. Rare-event sample counts and uncertainty matter.

Next comparable measurement

Register event thresholds before the season; measure recall, false alarms, calibration, track error and warning lead time on every eligible observed event. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Forecast skill

45%estimated progress

What we track

Continuous ranked probability score for forecast distributions, Brier score for specified events, and appropriate deterministic error metrics, by variable, region, and lead time.

Why this estimate?

Broad historical skill evaluation and an operational model lineage support substantial bounded progress. Human-designed verification, initialization and research agendas remain necessary.

Advancements

GraphCast beats ECMWF's HRES on more than 90% of 1,380 test targetsNovember 14, 2023. Google DeepMind's GraphCast, published in Science, was more accurate than ECMWF's deterministic HRES forecast on more than 90% of 1,380 test variables and lead times, and on 99.7% of them in the troposphere.
GenCast beats ECMWF's ensemble on 97.2% of targetsDecember 4, 2024. GenCast's published evaluation outperformed ECMWF ENS on 97.2% of 1,320 evaluated targets.
ECMWF makes AIFS Single its first operational AI forecast model25 February 2025. ECMWF took its Artificial Intelligence Forecasting System (AIFS Single, a deterministic model) operational alongside the physics-based IFS. ECMWF says it outperforms state-of-the-art physics-based models for many measures, including tropical cyclone tracks, with gains of up to 20%.
Microsoft's Aurora outperforms existing models on 91% of forecasting targetsMay 21, 2025. Microsoft reported, with a paper in Nature, that the Aurora foundation model beat existing numerical and AI models on 91% of forecasting targets when fine-tuned for 0.25-degree medium-range forecasts, and matched or exceeded existing ocean-wave forecasts in 86% of comparisons over a test year.
ECMWF's 51-member AI ensemble AIFS ENS becomes operational1 July 2025. ECMWF took the AIFS ENS ensemble model operational, producing 51 forecasts per run at 31 km resolution alongside the 9 km physics-based ensemble. ECMWF says it outperforms state-of-the-art physics-based models on many measures, including surface temperature, with improvements of up to 20%.
WeatherNext 2 beats the previous WeatherNext on 99.9% of variables and lead timesNovember 17, 2025. Google DeepMind and Google Research released WeatherNext 2, which outperforms the previous WeatherNext model on 99.9% of variables and lead times from 0 to 15 days, with predictions down to 1-hour resolution.
ECMWF puts its AI forecast models into operationMay 12, 2026 deployment. ECMWF records operational AIFS Single v2 and Ensemble v2 from May 12, 2026. The historical GenCast score is not their score.

Read the dated findings and sources

Status: Measured historical benchmark; current operational version identified. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

December 4, 2024: GenCast's published evaluation outperformed ECMWF ENS on 97.2% of 1,320 evaluated targets. GenCast, Nature.

May 12, 2026 deployment: ECMWF records operational AIFS Single v2 and Ensemble v2 from May 12, 2026. The historical GenCast score is not their score. ECMWF AIFS version history.

Interpretation and tracking value

This is the fraction of target metrics improved, not forecast accuracy. Historical ENS and evaluation periods must not be confused with November 2026 operational comparisons.

Next comparable measurement

Archive CRPS/Brier/error tables by variable, region and lead time for contemporaneous systems; preserve initialization, resolution, ensemble size and uncertainty. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Robustness

20%estimated progress

What we track

Skill in poorly observed regions, unusual seasons, and rare events, with uncertainty and missing-data handling reported.

Why this estimate?

An operational initialization change exposes a concrete adaptation challenge. Reliable self-detection and correction of such shifts across environments are not established.

Advancements

Study finds AI models underestimate Storm Ciaran's peak winds5 December 2023. A comparison of FourCastNet, Pangu-Weather, GraphCast and FourCastNet-v2 on Storm Ciaran found the models accurately capture the cyclone's synoptic-scale structure, but all of them underestimate the peak amplitude of the storm's winds and none capture the sharp bent-back warm frontal gradient.
FourCastNet trained without strong cyclones cannot forecast Category 5 storms19 October 2024. Researchers trained FourCastNet on 1979-2015 data with Category 3-5 tropical cyclones removed and tested it on 2018-2023 Category 5 storms. Global accuracy was similar, but the model could not accurately forecast Category 5 cyclones, showing it does not extrapolate from weaker storms; it did generalize across ocean basins. Published in PNAS (DOI 10.1073/pnas.2420914122).
Study: ECMWF's HRES beats AI models on record-breaking heat, cold and wind21 August 2025. The preprint 'Numerical models outperform AI weather forecasts of record-breaking extremes' finds that ECMWF's physics-based HRES consistently outperforms GraphCast, Pangu-Weather and Fuxi on record-breaking heat, cold and wind events at nearly all lead times. The AI models underestimate the frequency and intensity of records, with errors growing as records are exceeded by larger margins.
ECMWF retires external AI-model forecastsMay 11, 2026. ECMWF's May 2026 discussion describes retiring external AI-model forecasts as the operational initialization system changed.

Read the dated findings and sources

Status: Operational distribution-shift issue documented. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

May 11, 2026: ECMWF's May 2026 discussion describes retiring external AI-model forecasts as the operational initialization system changed. ECMWF: farewell to external AI models.

Interpretation and tracking value

A model's historical skill may not survive an upstream data change; this is not a universal quantitative robustness score.

Next comparable measurement

Test before and after initialization changes and by poorly observed region, unusual season and rare event; retain missing-data procedures and uncertainty. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note