How AI Is Changing Weather Prediction
From Equations to Learning: How AI Is Changing Weather Prediction
For most of modern meteorology, predicting tomorrow's weather has meant describing the atmosphere with mathematics and asking a computer to calculate what happens next.
Today, there is a second approach.
Instead of explicitly calculating every step of atmospheric evolution from equations written by scientists, a machine-learning system can study decades of previous atmospheric states and learn how one state tends to develop into another.
That sounds like a subtle change. It is not.
It represents one of the most important changes in numerical weather prediction since computers first became central to forecasting. By 2026, artificial intelligence is no longer simply an interesting weather-research experiment. AI forecasting systems are being operated alongside traditional numerical models by major forecasting centers, including the European Centre for Medium-Range Weather Forecasts (ECMWF) and NOAA in the United States.
But today's revolution did not appear overnight. Machine learning is the newest chapter in a much longer story of computers, statistics, observations and atmospheric physics.
Before AI: Turning the Atmosphere Into Numbers
Modern forecasting is built around numerical weather prediction, or NWP.
At its simplest, an NWP model begins with an estimate of the current state of the atmosphere, including variables such as temperature, air pressure, humidity and wind. The atmosphere is represented numerically, and equations describing its physical behavior are calculated forward through time.
This approach became enormously successful.
Better observations, satellites, improved atmospheric physics and increasingly powerful supercomputers steadily increased the accuracy of forecasts. Modern forecasting centers developed extraordinarily sophisticated models capable of representing the atmosphere on a global scale.
But numerical weather prediction has an unavoidable cost: computation.
The more detail meteorologists want to represent, and the more possible futures they want to examine, the more computing power they generally need.
Ensemble forecasting magnifies the problem. Instead of calculating one future atmosphere, forecasters calculate many slightly different versions. Comparing those forecasts helps reveal uncertainty. If nearly every model run produces a hurricane moving toward the same coastline, confidence is much greater than if the ensemble members scatter across hundreds of miles.
The technique is powerful, but performing all those calculations is expensive.
That computational problem eventually created an enormous opportunity for machine learning.
Machine Learning Was Already Part of Weather Forecasting
AI did not suddenly arrive in meteorology with today's neural networks.
Statistical techniques have long helped meteorologists interpret numerical forecasts. One important American example is Model Output Statistics, commonly called MOS. Rather than replacing the physical forecast model, statistical relationships could be used to translate model output into more useful local predictions.
The basic idea was important.
A computer model might consistently predict a particular variable slightly too high or too low at a particular location. Historical observations could reveal that relationship. Statistics could then help correct or interpret the raw model prediction.
This established an idea that survives in today's AI systems:
Past weather can teach us something about how to interpret future weather.
For decades, however, statistical methods generally sat downstream from numerical weather prediction. Physics created the forecast. Statistics helped refine it.
Machine learning would eventually challenge that division.
The Data Revolution
Several developments made the new approach possible.
Weather agencies accumulated enormous archives of atmospheric observations. Satellites continuously expanded the amount of information available about Earth. Computing improved. Graphics processors and specialized AI hardware made training very large neural networks practical.
Another particularly important development was reanalysis.
A reanalysis takes historical observations and processes them using a consistent modern forecasting system, creating a detailed reconstruction of the atmosphere over many years.
Instead of giving a machine-learning system a messy assortment of old weather reports, researchers could give it decades of structured descriptions of the atmosphere.
The problem suddenly began to resemble one that machine learning is exceptionally good at solving:
Given this state, what state usually comes next?
A neural network does not need to be explicitly programmed with every atmospheric relationship in the same manner as a conventional numerical weather model. During training, it adjusts huge numbers of internal parameters as it learns relationships within historical atmospheric data.
After training, forecasting becomes inference.
Give the model the atmosphere now.
Ask it to predict the atmosphere later.
Then repeat.
Deep Learning Enters the Forecast
During the late 2010s and early 2020s, increasingly sophisticated machine-learning weather systems began appearing.
Researchers experimented with convolutional neural networks, transformers, graph neural networks and other architectures. AI was applied to precipitation nowcasting, radar interpretation, forecast post-processing and increasingly ambitious global weather prediction.
The critical transition came when researchers stopped asking only whether machine learning could improve conventional forecasts.
They began asking whether it could predict the evolution of the global atmosphere itself.
A series of models demonstrated that the answer was increasingly yes.
One important milestone was Pangu-Weather, developed by researchers at Huawei Cloud.
Published in Nature in July 2023, Pangu-Weather used three-dimensional deep networks trained on 39 years of global data. Its researchers reported stronger deterministic results on reanalysis data across the variables they tested than ECMWF's operational Integrated Forecasting System. They also reported strong performance in areas including tropical-cyclone tracking.
The significance extended beyond accuracy.
Machine learning offered the possibility of producing sophisticated forecasts without performing the enormous set of physical calculations required during every conventional model run.
Training the AI could be computationally expensive.
Running the trained model could be remarkably fast.
GraphCast Changes the Conversation
Later in 2023, Google DeepMind brought considerably more attention to the field with GraphCast.
GraphCast approached the atmosphere as an interconnected global system using graph neural networks. Rather than treating weather simply as pixels on a rectangular map, the architecture represented relationships across Earth's atmosphere using a multidimensional graph.
The results were striking.
Published in Science, GraphCast generated global 10-day forecasts at 0.25-degree resolution in under a minute. Its researchers reported that it significantly outperformed a leading operational deterministic system on 90 percent of 1,380 verification targets. The researchers also demonstrated useful performance in predicting phenomena including tropical-cyclone tracks, atmospheric rivers and extreme temperatures.
The speed was almost as important as the accuracy.
Traditional numerical weather prediction requires enormous computing systems repeatedly solving physical equations across millions of atmospheric points.
Once trained, GraphCast learned a computational shortcut through that problem.
It had effectively learned patterns describing how the global atmospheric state evolves.
That does not mean GraphCast had discovered a replacement for atmospheric physics. Its knowledge ultimately came from atmospheric data generated and reconstructed through decades of observations and meteorological science.
But it demonstrated something profound:
A machine-learning model could learn enough about atmospheric evolution to produce highly competitive global forecasts without reproducing the entire traditional numerical calculation every time a forecast was made.
From Research to the Forecast Center
For a while, AI weather models were easy to describe as impressive research demonstrations.
That description no longer works.
On February 25, 2025, ECMWF made its Artificial Intelligence Forecasting System, or AIFS, operational alongside its traditional physics-based Integrated Forecasting System.
According to ECMWF, AIFS outperformed state-of-the-art physics-based models on many measures, including tropical cyclone tracks, while reducing energy use for producing a forecast by approximately 1,000 times. Importantly, ECMWF did not throw away its traditional forecasting system. The AI and physics-based systems operate together.
Then came another important step.
Weather is inherently uncertain, so operational forecasting depends heavily on ensembles rather than a single predicted future. In July 2025, ECMWF made AIFS ENS, its AI ensemble system, operational. It generates 51 different forecasts with small variations to represent a range of possible future atmospheric states. ECMWF reported that AIFS ENS could generate forecasts more than 10 times faster than its physics-based system while again reducing energy consumption by approximately a factor of 1,000.
The important word in both developments is operational.
AI weather forecasting had crossed the boundary from experimental research into systems used by a major forecasting center.
NOAA Joins the Operational AI Era
The same transition has taken place in the United States.
In December 2025, NOAA announced a suite of operational AI-driven global models.
The Artificial Intelligence Global Forecast System, or AIGFS, provides AI-based deterministic forecasts. NOAA says a 16-day AIGFS forecast uses only 0.3 percent of the computing resources required by the operational GFS and completes in approximately 40 minutes.
The Artificial Intelligence Global Ensemble Forecast System, or AIGEFS, extends the approach to probabilistic forecasting. NOAA reported that its early results improved on the traditional GEFS and extended useful forecast skill by 18 to 24 hours.
Perhaps most interesting is Hybrid-GEFS, or HGEFS.
Instead of asking forecasters to choose between artificial intelligence and conventional physics, HGEFS combines AI-based AIGEFS forecasts with NOAA's physics-based Global Ensemble Forecast System. NOAA says testing showed that the hybrid ensemble consistently outperformed the AI-only and physics-only ensemble systems.
That may offer one of the clearest indications of where weather prediction is heading.
The future may not be AI versus physics. It may be AI plus physics.
What the AI Actually Learns
It is tempting to describe these systems as pattern-matching machines, but that description can undersell what is happening.
The atmosphere contains enormous numbers of interconnected relationships.
Temperature affects pressure. Pressure differences drive winds. Winds transport heat and moisture. Water changes phase. Clouds form. Solar radiation heats the surface unevenly. Oceans exchange energy with the atmosphere. Mountains alter airflow. Storm systems interact with jet streams across thousands of miles.
Traditional numerical weather prediction attempts to represent those processes explicitly through physics and numerical approximation.
Machine-learning models approach the problem differently.
During training, they encounter enormous numbers of examples of atmospheric evolution. Their internal parameters gradually encode relationships that allow them to transform one atmospheric state into a plausible future state.
GraphCast, for example, was trained directly from reanalysis data. Pangu-Weather was trained using 39 years of global atmospheric data. ECMWF says AIFS learned from about 40 years of its ERA5 reanalysis before further development for operational forecasting.
The result is an interesting inversion of traditional forecasting.
A traditional model essentially says:
We know the physics. Calculate what happens next.
A machine-learning model says:
We have seen how the atmosphere evolved before. Learn the relationships and predict what happens next.
AI Still Needs Traditional Meteorology
The apparent opposition between those approaches can be misleading.
Today's AI forecasting systems remain deeply connected to traditional meteorology.
Before any forecast can begin, scientists need to know the current state of the atmosphere. That requires observations from satellites, aircraft, ships, buoys, weather stations and other instruments.
Those observations must then be turned into a coherent three-dimensional description of the atmosphere.
ECMWF notes that AIFS uses the same initial conditions as its physics-based IFS. Those initial conditions combine a previous short-range forecast with roughly 60 million quality-controlled observations.
Its ensemble AI system likewise continues to depend on physics-based data assimilation for initial conditions. ECMWF explicitly says its conventional system remains indispensable for high-resolution fields and coupled Earth-system processes and that it is investigating hybrid approaches combining the strengths of both systems.
AI, in other words, has not made conventional meteorology obsolete.
It has become another powerful layer within it.
Why Speed Changes More Than the Computer Bill
Much of the discussion around AI weather forecasting focuses on computational efficiency, but speed has scientific value beyond reducing electricity consumption.
Weather forecasting is fundamentally a problem of uncertainty.
A single forecast represents one possible atmospheric future. An ensemble represents many.
If forecasts become dramatically cheaper to calculate, forecasting centers can potentially explore uncertainty much more extensively. More model runs, different initial conditions and combinations of different forecasting architectures can provide a richer picture of what might happen.
Instead of saying:
This is what the weather will be.
The system can increasingly help answer:
What are the plausible outcomes, and how likely is each one?
That distinction becomes especially important for hurricanes, floods, winter storms and other high-impact events where the difference between two plausible tracks can mean radically different consequences for communities.
The Meteorologist Remains in the Loop
None of this eliminates the forecaster.
Weather prediction has never simply been the output of a computer model.
Meteorologists compare models, study observational trends, identify model biases, examine ensemble probabilities and interpret what those signals mean for particular places and populations.
AI adds another extremely powerful source of guidance.
It may also change the nature of forecasting work.
A June 2026 paper led by researchers associated with ECMWF argues that machine learning and related digital technologies could reshape much of the weather-forecasting pipeline, including model development, observations, data management, verification and the creation of forecast services. At the same time, the authors emphasize the continuing importance of scientific understanding, operational reliability and human expertise.
That is an important distinction.
Predicting atmospheric variables and forecasting weather for people are related problems, but they are not identical.
A neural network may predict pressure, temperature, precipitation and wind.
Humans still have to ask what those predictions mean.
Weather Prediction in 2026
As of 2026, AI weather forecasting has moved remarkably quickly from experimental demonstrations into operational meteorology.
ECMWF operates both deterministic and ensemble versions of AIFS alongside its traditional physics-based systems.
NOAA operates AI-based global deterministic and ensemble systems, as well as a hybrid ensemble combining machine-learning and traditional forecasts. In July 2026, NOAA's Project EAGLE updated its global AI system with improvements aimed particularly at tropical-cyclone intensity and precipitation forecast skill.
At the same time, traditional numerical weather prediction remains fundamental.
That makes the current moment less a replacement of one technology by another than a convergence.
For decades, forecasting progressed mainly by improving observations, improving physical models and giving those models larger computers.
Now meteorology has another lever.
It can learn from the atmosphere's history.
Physics tells the computer how scientists understand the atmosphere to work. Machine learning extracts relationships from how the atmosphere has actually behaved. Observations tell both systems where the atmosphere is now.
The emerging forecasting system combines all three.
And that may ultimately be the most important contribution of artificial intelligence to weather prediction.
The computer is no longer limited to calculating equations that scientists have explicitly given it. It can also learn statistical representations of atmospheric behavior from decades of data.
Weather prediction is therefore entering an unusual period in its history. The old forecasting system is not disappearing. Instead, physics-based simulation, observation, statistics and machine learning are beginning to operate together.
After decades of teaching computers the equations of the atmosphere, meteorologists are now doing something new. They are also letting the computers study the weather itself.
Aaron Rose is the founder and publisher of Tech Reader Magazine, covering system architecture, cloud platforms, artificial intelligence and emerging technology.