Uncertainty

Introduction

Modern weather forecasting is not really about predicting one future. It is about estimating many plausible futures and judging how likely each one is. That shift matters because the atmosphere is chaotic: tiny differences in current conditions can grow into completely different outcomes days later. AI ensemble systems such as GenCast changed forecasting not mainly by producing a single “better guess”, but by making uncertainty itself faster, cheaper and more usable. [Google DeepMind]deepmind.googlegencast predicts weather and the risks of extreme conditions with sota accuracyGoogle DeepMindGenCast predicts weather and the risks of extreme…4 Dec 2024 — New AI model advances the prediction of weather uncertai… [ECMWF That may sound like a technical detail]ecmwf.intECMWFChaos and Weather PredictionThe weather is a chaotic system. Small errors in the initial conditions of a forecast grow rapidly, and…, but it has broader significance for scientific acceleration. Many real-world systems — disease spread, energy demand, financial stress, climate risk, even parts of biology — are probabilistic rather than perfectly predictable. AI weather ensembles provide one of the clearest demonstrations that machine learning can model distributions of possible futures at planetary scale. In the wider AI bloom argument, this matters because civilisation often advances not by eliminating uncertainty, but by understanding and managing it better.

Uncertainty illustration 1

Why weather is probabilistic rather than certain

Weather prediction became one of the classic examples of chaos theory because small measurement errors amplify over time. A tiny uncertainty in wind speed, humidity or pressure today can produce a very different storm track a week later. ECMWF, one of the world’s leading forecasting centres, describes weather as a chaotic system in which errors in initial conditions rapidly grow and limit predictability. [ECMWF]ecmwf.intECMWFChaos and Weather PredictionThe weather is a chaotic system. Small errors in the initial conditions of a forecast grow rapidly, and… [ECMWF]ecmwf.intECMWFIntroduction to chaos, predictability and ensemble forecastsThe ECMWF Ensemble Prediction System (EPS) provides a practical tool for…

This creates a basic problem for deterministic forecasting. A traditional forecast might say:

rain will hit London on Thursday afternoon
a hurricane will land near one coastline
temperatures will reach a certain level

But in reality there may be several plausible outcomes. A storm could veer east. Rain bands could weaken. Heat could intensify unexpectedly. A single “best guess” hides the spread of possibilities.

Meteorologists therefore moved toward ensemble forecasting decades ago. Instead of running one simulation, forecasting centres run many slightly different simulations with altered starting conditions or altered model assumptions. ECMWF’s ensemble system, for example, runs dozens of forecasts in parallel to estimate the range of plausible futures. [ECMWF]ecmwf.intfact sheet ensemble weather forecastingFact sheet: Ensemble weather forecasting23 Mar 2017 — An ensemble weather forecast is a set of forecasts that present the range of future… [arXiv]arxiv.orgThis paper has been written to mark 25 years of operational medium-range ensemble forecasting. The origins of the ECMWF Ensemble Predicti…

The result is not just a prediction but a probability distribution:

70% chance of heavy rainfall
20% chance of hurricane landfall further north
small but non-zero risk of an extreme temperature spike

That information is often more useful for real decisions than a single answer. Energy grids, shipping companies, emergency planners, airlines and farmers all need to know not only what is likely, but how uncertain the forecast is.

How AI ensemble forecasts differ from single best guesses

Early AI weather systems such as GraphCast mainly produced deterministic forecasts: one predicted atmospheric evolution from one starting state. These systems were already impressive because they could rival major numerical weather models while running dramatically faster. But they still inherited a core limitation: they produced one trajectory at a time. [GitHub]github.comGitHubgoogle-deepmind/graphcastGenCast: Diffusion-based ensemble forecasting for medium-range weather. This package provides four pretrai…

GenCast represented a more important conceptual step. Instead of producing one future weather state, it generated many possible future trajectories using probabilistic modelling. DeepMind described the system as a probabilistic ensemble forecaster capable of modelling weather uncertainty directly. [Google DeepMind]deepmind.googlegencast predicts weather and the risks of extreme conditions with sota accuracyGoogle DeepMindGenCast predicts weather and the risks of extreme…4 Dec 2024 — New AI model advances the prediction of weather uncertai… [Nature This matters because uncertainty is not an afterthought in weather science. It is part of the problem itself.]nature.comProbabilistic weather forecasting with machine learningby I Price · 2025 · Cited by 656 — Here we introduce GenCast, a probabilistic weat…

Traditional ensemble systems achieve this by repeatedly running expensive physics simulations with slightly altered conditions. AI ensemble systems instead learn statistical patterns in how weather evolves across decades of atmospheric data. GenCast used a diffusion-model approach — related in some ways to methods used in image generation — to sample multiple plausible atmospheric futures from the same starting conditions. [arXiv]arxiv.orgThis paper has been written to mark 25 years of operational medium-range ensemble forecasting. The origins of the ECMWF Ensemble Predicti…

The practical consequence is speed and scale.

DeepMind reported that GenCast could generate a 15-day ensemble forecast in roughly eight minutes, while outperforming ECMWF’s ENS ensemble benchmark on most tested targets. [Google DeepMind]deepmind.googlegencast predicts weather and the risks of extreme conditions with sota accuracyGoogle DeepMindGenCast predicts weather and the risks of extreme…4 Dec 2024 — New AI model advances the prediction of weather uncertai… [Nature That computational efficiency changes what becomes feasible:]nature.comProbabilistic weather forecasting with machine learningby I Price · 2025 · Cited by 656 — Here we introduce GenCast, a probabilistic weat…

larger ensembles with more possible scenarios [ecmwf.int]ecmwf.intFact sheetEnsemble Weather Forecasting.inddThe uncertainty associated with every forecast means that different scenarios are possible, and the fore…
more frequent updates
wider access outside elite forecasting centres
faster experimentation with forecasting architectures
cheaper probabilistic forecasting for specialised applications

This is important because uncertainty estimation is computationally expensive. In conventional forecasting, each additional ensemble member means another large simulation. AI models reduce that cost enough that uncertainty modelling can become routine rather than exceptional.

Why probabilities often matter more than point forecasts

For many decisions, the tails of the distribution matter more than the average outcome.

A city preparing for flooding does not mainly care about the most likely rainfall total. It cares about the probability of catastrophic rainfall. An electricity operator balancing renewable energy needs estimates of wind variability. Emergency managers need to know whether a hurricane has a 5% or 25% chance of changing course toward a populated region.

[Ensemble forecasts help reveal those possibilities.]ecmwf.intfact sheet ensemble weather forecastingFact sheet: Ensemble weather forecasting23 Mar 2017 — An ensemble weather forecast is a set of forecasts that present the range of future…

Meteorologists sometimes visualise hurricane ensembles as “spaghetti plots”, where many possible storm tracks spread outward over time. Traditional deterministic forecasts can appear deceptively precise. Ensembles instead expose how confidence changes across time and geography. [WIRED]wired.comAI Hurricane Predictions Are Storming the World of Weather ForecastingThe AI models, developed by Nvidia, Huawei, and Google's DeepMind, forecasted its likely trajectory between Rhode Island and Nova Scotia…

AI ensembles potentially improve this in two ways simultaneously:

they may increase forecast skill
they make large ensembles cheaper to generate

That combination is important because better uncertainty estimates can create economic value even when average forecast accuracy improves only modestly. ECMWF research has argued that probabilistic forecasts often have greater decision value than deterministic ones because they allow risk-sensitive planning. [ECMWF]ecmwf.intQuantifying forecast uncertaintyDespite the increasing accuracy of weather forecasts, there is an element of uncertainty in all predictio…

In practice, this means AI weather systems are becoming tools for decision-making under uncertainty rather than just weather prediction engines.

Uncertainty illustration 2

GenCast as a broader scientific signal

The wider importance of GenCast is not only meteorological. It demonstrates a broader pattern in AI-assisted science: modelling distributions instead of single outcomes.

Many scientific systems behave more like weather than like clockwork. Biology, epidemiology, materials science and economics often involve branching possibilities rather than deterministic certainty. The future state of the system depends on incomplete observations, hidden interactions and stochastic processes.

AI ensemble forecasting suggests machine learning may help in three broader scientific tasks:

estimating uncertainty directly rather than hiding it
generating many plausible trajectories cheaply
updating probabilistic beliefs rapidly as new data arrives

That is potentially important for scientific acceleration because real-world decision-making rarely waits for certainty.

Drug discovery, for example, may benefit from systems that estimate confidence intervals around molecular predictions rather than simply ranking compounds. Climate-risk planning may increasingly rely on AI systems that explore scenario spaces rapidly. Energy systems with large renewable inputs may need constant probabilistic balancing of uncertain supply and demand.

The key lesson from weather forecasting is that uncertainty is not merely noise around prediction. It is often the main object of prediction.

Why AI ensembles may change the economics of science

One reason weather forecasting became a frontier for AI is that it offers continuous feedback. Forecasts are tested against reality every day. But another reason is computational cost.

Traditional numerical weather prediction consumes enormous supercomputing resources because it repeatedly solves physical equations across a global grid. Ensemble forecasting multiplies those costs further because many simulations must run simultaneously. [ECMWF]ecmwf.intECMWFPredicting uncertainty in forecasts of weather and climateBy using ensemble forecasts as input to a simple decision-model analysis…

AI systems potentially compress this cost dramatically.

That matters beyond weather because expensive simulation limits experimentation in many sciences. If probabilistic modelling becomes cheaper, researchers can test more hypotheses, run more scenarios and iterate faster.

The effect resembles a broader trend in AI-assisted science:

lower-cost modelling
faster iteration loops
wider access to advanced computation
larger exploration of possibility spaces

In the optimistic AI bloom interpretation, this could contribute to a civilisation that becomes better at navigating uncertainty itself. Scientific progress often depends not on knowing everything with certainty, but on making better probabilistic judgments faster than before.

Uncertainty illustration 3

The limits and unresolved questions

The excitement around AI ensembles comes with important caveats.

Traditional forecasting systems still provide much of the physical infrastructure and training data that AI systems depend on. Many meteorologists expect hybrid systems rather than pure AI replacement. [The Guardian]theguardian.comGenCast is proficient in predicting day-to-day weather and extreme events up to 15 days ahead and surpasses ENS in forecasting hurricane…

There are also concerns about calibration and physical realism. A probabilistic model is only useful if its uncertainty estimates are trustworthy. An ensemble that looks confident when it should not be can become dangerous. Some researchers argue that deterministic AI systems still struggle to represent extreme-event uncertainty adequately without specialised probabilistic methods. [arXiv]arxiv.orgThis paper has been written to mark 25 years of operational medium-range ensemble forecasting. The origins of the ECMWF Ensemble Predicti…

Another open question is whether AI ensembles genuinely understand atmospheric dynamics or mainly interpolate from historical patterns. Extreme climate conditions outside historical experience may test those limits.

There is also a governance dimension. Better uncertainty forecasting could improve disaster preparedness, energy planning and climate resilience broadly. But unequal access to advanced forecasting systems could also concentrate advantages in wealthy states and corporations with greater computational resources.

These tensions mirror broader debates in the AI bloom discussion. Faster intelligence and prediction alone do not guarantee broad human flourishing. Institutions, access, transparency and public capacity still matter.

Why uncertainty forecasting matters for the long-term future

One of the deepest implications of AI ensemble forecasting is philosophical rather than technical. It reflects a shift from viewing science as the search for perfectly certain answers toward viewing it as increasingly powerful management of uncertainty.

Human civilisation already operates through probabilistic reasoning in finance, medicine, engineering and public health. But many of those systems remain slow, fragmented and computationally constrained.

AI weather ensembles show that machine learning may help civilisation reason about complex futures more dynamically:

updating forecasts continuously
exploring many futures simultaneously
estimating confidence directly
identifying low-probability high-impact risks earlier

That capability could matter profoundly in a more advanced civilisation. Long-term technological societies may depend increasingly on forecasting systems that help humans coordinate under uncertainty — whether for pandemics, climate risks, energy systems, infrastructure or eventually space settlement.

Weather forecasting therefore matters not just because storms are important, but because it offers a visible, measurable example of something larger: AI systems learning to represent uncertainty itself as a first-class scientific object. [Google DeepMind]deepmind.googlegencast predicts weather and the risks of extreme conditions with sota accuracyGoogle DeepMindGenCast predicts weather and the risks of extreme…4 Dec 2024 — New AI model advances the prediction of weather uncertai… [2wmo.int]wmo.intimo prize lecture 2024 ensemble weather and climate prediction from origins aiIMO Prize Lecture 2024: Ensemble Weather and Climate…5 Nov 2024 — The key purpose of an ensemble forecast system is to help determine…

Endnotes

Source: deepmind.google
Title: gencast predicts weather and the risks of extreme conditions with sota accuracy
Link: https://deepmind.google/blog/gencast-predicts-weather-and-the-risks-of-extreme-conditions-with-sota-accuracy/
Source snippet
Google DeepMindGenCast predicts weather and the risks of extreme...4 Dec 2024 — New AI model advances the prediction of weather uncertai...
Source: ecmwf.int
Link: https://www.ecmwf.int/en/elibrary/79859-chaos-and-weather-prediction
Source snippet
ECMWFChaos and Weather PredictionThe weather is a chaotic system. Small errors in the initial conditions of a forecast grow rapidly, and...
Source: ecmwf.int
Link: https://www.ecmwf.int/en/learning/training/introduction-chaos-predictability-and-ensemble-forecasts
Source snippet
ECMWFIntroduction to chaos, predictability and ensemble forecastsThe ECMWF Ensemble Prediction System (EPS) provides a practical tool for...
Source: ecmwf.int
Title: fact sheet ensemble weather forecasting
Link: https://www.ecmwf.int/en/about/media-centre/focus/2017/fact-sheet-ensemble-weather-forecasting
Source snippet
Fact sheet: Ensemble weather forecasting23 Mar 2017 — An ensemble weather forecast is a set of forecasts that present the range of future...
Source: arxiv.org
Link: https://arxiv.org/pdf/1803.06940
Source snippet
This paper has been written to mark 25 years of operational medium-range ensemble forecasting. The origins of the ECMWF Ensemble Predicti...
Source: ecmwf.int
Link: https://www.ecmwf.int/en/research/modelling-and-prediction/quantifying-forecast-uncertainty
Source snippet
Quantifying forecast uncertaintyDespite the increasing accuracy of weather forecasts, there is an element of uncertainty in all predictio...
Source: github.com
Link: https://github.com/google-deepmind/graphcast
Source snippet
GitHubgoogle-deepmind/graphcastGenCast: Diffusion-based ensemble forecasting for medium-range weather. This package provides four pretrai...
Source: nature.com
Link: https://www.nature.com/articles/s41586-024-08252-9
Source snippet
Probabilistic weather forecasting with machine learningby I Price · 2025 · Cited by 656 — Here we introduce GenCast, a probabilistic weat...
Source: arxiv.org
Title: arXiv Gen Cast: Diffusion-based ensemble forecasting for medium-range weather
Link: https://arxiv.org/abs/2312.15796
Source: developers.google.com
Link: https://developers.google.com/weathernext/guides/research
Source snippet
Google for DevelopersWeather research | WeatherNextNov 17, 2025 — GenCast. GNN/Transformer-based probabilistic (diffusion) model for ense...
Source: wired.com
Title: AI Hurricane Predictions Are Storming the World of Weather Forecasting
Link: https://www.wired.com/story/ai-hurricane-predictions-are-storming-the-world-of-weather-forecasting
Source snippet
The AI models, developed by Nvidia, Huawei, and Google's DeepMind, forecasted its likely trajectory between Rhode Island and Nova Scotia...
Source: ecmwf.int
Link: https://www.ecmwf.int/en/elibrary/75907-predicting-uncertainty-forecasts-weather-and-climate
Source snippet
ECMWFPredicting uncertainty in forecasts of weather and climateBy using ensemble forecasts as input to a simple decision-model analysis...
Source: ecmwf.int
Link: https://www.ecmwf.int/en/research/modelling-and-prediction
Source snippet
ECMWFModelling and PredictionAll our forecasts and reanalyses use a numerical model to make a prediction. We have developed our own atmos...
Source: arxiv.org
Link: https://arxiv.org/abs/2511.17176
Source snippet
arXivOn the Predictive Skill of Artificial Intelligence-based Weather Models for Extreme Events using Uncertainty QuantificationNovember...
Source: arxiv.org
Title: arXiv A Practical Probabilistic Benchmark for AI Weather Models
Link: https://arxiv.org/abs/2401.15305
Source: wmo.int
Title: imo prize lecture 2024 ensemble weather and climate prediction from origins ai
Link: https://wmo.int/media/magazine-article/imo-prize-lecture-2024-ensemble-weather-and-climate-prediction-from-origins-ai
Source snippet
IMO Prize Lecture 2024: Ensemble Weather and Climate...5 Nov 2024 — The key purpose of an ensemble forecast system is to help determine...
Source: ecmwf.int
Link: https://www.ecmwf.int/
Source snippet
We are both a research institute and a 24/7 operational service, producing global numerical...
Source: ecmwf.int
Title: Fact sheet
Link: https://www.ecmwf.int/sites/default/files/medialibrary/2017-03/ecmwf-fact-sheet-ensemble-forecasting.pdf
Source snippet
Ensemble Weather Forecasting.inddThe uncertainty associated with every forecast means that different scenarios are possible, and the fore...
Source: ecmwf.int
Title: 16927 chaos and weather prediction
Link: https://www.ecmwf.int/sites/default/files/elibrary/2002/16927-chaos-and-weather-prediction.pdf
Source snippet
Chaos and weather prediction January 2000by R Buizza · Cited by 82 — Ensemble prediction is a feasible method to integrate a single, dete...
Source: nature.com
Title: And it does so faster than the best
Link: https://www.nature.com/articles/d41586-024-03957-3
Source snippet
DeepMind AI weather forecaster beats world-class systemby A Soliman · 2024 · Cited by 8 — Artificial-intelligence model provides forecast...
Source: nature.com
Link: https://www.nature.com/articles/s41612-025-01009-9
Source snippet
A fast physics-based perturbation generator of machine...by J Pu · 2025 · Cited by 13 — We propose a fast, physics-constrained perturbat...
Source: developers.google.com
Title: projects gcp public data weathernext assets 126478713 1 0
Link: https://developers.google.com/earth-engine/datasets/catalog/projects_gcp-public-data-weathernext_assets_126478713_1_0
Source snippet
· The dataset contains both real-...Read more...
Source: deepmind.google
Link: https://deepmind.google/science/weathernext/
Source snippet
WeatherNext 2Our most accurate AI weather forecasting technology · Efficient and reliable forecasts · Accuracy · Speed and efficiency · T...
Source: youtube.com
Link: https://www.youtube.com/watch?v=pWxkf7cGI9k
Source snippet
Google DeepMind's GenCast is Revolutionizing Weather Forcasting...
Source: youtube.com
Title: Google Deep Mind’s Gen Cast is Revolutionizing Weather Forcasting
Link: https://www.youtube.com/watch?v=N1ikK-E-mqw
Source snippet
snowbrains.com...
Source: youtube.com
Link: https://www.youtube.com/watch?v=i5nUrT5WNe0
Source snippet
ECMWF IFS Cycle 50r1 and AIFS v2 webinar: Data access and format, testing and practicalities...
Source: youtube.com
Link: https://www.youtube.com/watch?v=RMN1O0nLbbE
Source: theguardian.com
Link: https://www.theguardian.com/science/2024/dec/04/google-deepmind-predicts-weather-more-accurately-than-leading-system
Source snippet
GenCast is proficient in predicting day-to-day weather and extreme events up to 15 days ahead and surpasses ENS in forecasting hurricane...
Source: medium.com
Link: https://medium.com/%40sergiosear/gencast-redefining-weather-forecasting-with-next-generation-technology-dbab3376e2d2
Source snippet
GenCast: Redefining Weather Forecasting with Next-...GenCast is an innovative probabilistic weather forecasting model developed by DeepMind...
Source: youtube.com
Link: https://www.youtube.com/watch?v=NLhRUun2iso
Source snippet
ECMWF 25 YEARS OF ENSEMBLE PREDICTIONMIT on Chaos and Climate: From Determinism to Probability in Numerical Weather Prediction. Earth, At...
Source: Wikipedia
Title: Ensemble forecasting
Link: https://en.wikipedia.org/wiki/Ensemble_forecasting
Source snippet
Ensemble forecastingEnsemble forecasting is a method used in or within numerical weather prediction. Instead of making a single foreca...
Source: datacamp.com
Link: https://www.datacamp.com/blog/gencast
Source snippet
Inside Google's GenCast: Learn About AI in Forecasting29 Jan 2025 — Discover how GenCast makes use of machine learning and probabilistic...
Source: ftp.bills.com.au
Link: https://ftp.bills.com.au/lunar-tips/ecmwf-ensemble-weather-prediction-system-explained-1767648833
Source snippet
Ensemble: Weather Prediction System Explained - Ftp6 Jan 2026 — It allows us to quantify the uncertainty in the forecast, which is crucia...

Additional References

Source: reddit.com
Link: https://www.reddit.com/r/machinelearningnews/comments/1h7v850/google_deepmind_opensources_gencast_a_machine/
Source snippet
A Machine Learning-based Weather Model that can Predict...Researchers from Google DeepMind released GenCast, a probabilistic weather for...
Source: facebook.com
Link: https://www.facebook.com/datacampinc/posts/gencast-google-deepminds-weather-forecasting-model-uses-advanced-machine-learnin/1055219626640864/
Source snippet
GenCast, Google DeepMind's weather forecasting model...GenCast can predict weather conditions and extreme events up to 15 days in advanc...
Source: qts-ltd.com
Link: https://www.qts-ltd.com/gencast-ultra-advanced-weather-forecasting/
Source snippet
GenCast: Ultra Advanced Weather ForecastingGoogle DeepMind has introduced GenCast, an advanced AI model designed to revolutionise weather...
Source: linkedin.com
Link: https://www.linkedin.com/posts/mckabue_probabilistic-weather-forecasting-with-machine-activity-7414169010176454657-nZ84
Source snippet
Probabilistic weather forecasting with machine learningGoogle DeepMind's new AI model, GenCast, is a significant breakthrough in meteorol...
Source: linkedin.com
Link: https://www.linkedin.com/pulse/gencast-google-deepminds-ai-model-precise-efficient-weather-2ot5c
Source snippet
GenCast: Google DeepMind's AI Model for Precise...Probabilistic Ensemble Forecasting - Unlike deterministic models offering a single out...
Source: genre.com
Title: gen ai and its implications for weather and climate risk management en
Link: https://www.genre.com/us/knowledge/publications/2025/september/gen-ai-and-its-implications-for-weather-and-climate-risk-management-en
Source snippet
Generative Artificial Intelligence and Its Implications for...Sep 15, 2025 — Gen AI is changing how insurers manage weather and climate...
Source: openreview.net
Link: https://openreview.net/forum?id=wTIzpqX121
Source snippet
ral network model for weather forecasting, capturing uncertainty by generating ensemble forecasts.Read more...
Source: researchgate.net
Title: 386439155 Probabilistic weather forecasting with machine learning
Link: https://www.researchgate.net/publication/386439155_Probabilistic_weather_forecasting_with_machine_learning
Source snippet
Probabilistic weather forecasting with machine learningDec 4, 2024 — Here we introduce GenCast, a probabilistic weather model with greate...
Source: sciencemediacentre.es
Title: new machine learning model outperforms current weather forecasts
Link: https://sciencemediacentre.es/en/new-machine-learning-model-outperforms-current-weather-forecasts
Source snippet
New machine learning model outperforms current weather...Dec 4, 2024 — The model, called GenCast, outperforms the most efficient traditi...
Source: onyxaero.com
Title: probabilistic weather forecasting with machine learning
Link: https://onyxaero.com/news/probabilistic-weather-forecasting-with-machine-learning/
Source snippet
23 Dec 2024 — A new machine learning model from Google's DeepMind called GenCast demonstrates significant advances in weather forecasting...