Extreme weather events are harrowing for everyone who endures them. Hurricanes devastate entire communities, blizzards lead to power and water outages, and tornadoes leave a wake of destruction.
But thanks to improvements in AI weather forecasting, advancements in collecting atmospheric data and a willingness to experiment, weather prediction is becoming more precise than ever, which is making it easier to anticipate extreme weather events and, ultimately, prepare for them.
Consider, for example, data from the European Centre for Medium-Range Weather Forecasts (ECMWF): In May 2026, it reported that its Artificial Intelligence Forecasting System (AIFS) produces 5% to 15% lower errors compared to its traditional physics-based system for most medium-range atmospheric and surface variables, with forecasts extending to 15 days. For precipitation, that improvement translates to one additional day of useful forecast lead time, ECMWF shared.
In the United States, the National Oceanic and Atmospheric Administration has reported similar success with AI-based weather forecasting models. In December 2025, it said those models provide “a significant advancement in forecast speed, efficiency and accuracy.”
“NOAA’s strategic application of AI is a significant leap forward in American weather model innovation,” NOAA Administrator Neil Jacobs said in a news release.
“These AI models reflect a new paradigm … in providing improved accuracy for large-scale weather and tropical tracks, and faster delivery of forecast products to meteorologists and the public.”
Harnessing Data in AI Weather Forecasting
Succeeding with any AI model requires having valuable data — and enough of it. Weather forecasting is leading the way in pace of innovation, plus a willingness to approach data collection and analysis in unique ways.
“The last five years have seen the biggest changes in the field of meteorology that have ever occurred,” Spire Director of Weather Prediction and AI Tom Gowan told The Forecast in an interview.
Spire has built a large constellation of small satellites that can see the entire globe at all times and listen for radio-frequency signals from Earth. The company also launched the first Hyperspectral Microwave Sounding (HyMS) satellite in early 2026. The HyMS satellite listens for background microwave radiation to detect very-high-resolution 3D snapshots of clouds, precipitation, temperature and moisture.
That raw data can be paired with public satellites, ground stations, aircraft and other sensors of atmospheric data. The science of creating a 3D view of what’s happening from that data is known as data assimilation. For short-term weather modeling systems, Spire’s team runs high-resolution physical models. With longer ranges that can go up to 45-day, subseasonal forecasts, they’ve developed AI weather models that can be tailored to aviation, utilities, retailers or other markets.
Gowan said that every weather model forecast was using traditional physics-based models, which improved alongside compute, more input data and a better understanding of the atmosphere. In recent years, researchers have looked at graph neural networks, training a structured AI system based on weather going back 60 years. The AI models then tried to predict the next six hours of data without knowing what actually happened during that six-hour period.
“Turns out, with enough data and training compute, they do phenomenally well,” Gowan reported. “In many ways, they basically outperform the traditional physics models at coarser resolutions, at global scales and they’re remarkably cheap. We’re talking 100 times cheaper to run these AI models than the traditional physics-based models.”
Cheaper costs allow researchers to do ensemble forecasts, where they can make tweaks and run hundreds of different simulations for a given forecast. AI can handle forecasting in these custom physics-based models using temperature, pressure, humidity and other pure weather data.
Now, the data assimilation process is also beginning to fall to AI. The models often understand relationships humans don’t recognize in moving from a satellite’s radiometer to actual weather data.
“Eventually, we’re going to get to this world where it’s end-to-end AI being used to predict the weather,” suggested Gowan, who pointed out that some forecasting — such as understanding the impacts of individual thunderstorms or navigating areas with complex terrain — is still better performed by high-resolution models, since AI doesn’t yet have the appropriate training data and still experiences core issues with high resolution.
There’s not much value in a single, deterministic forecast, Gowan added. There either needs to be a full ensemble for a given model or a combination of several deterministic models, such as the spaghetti plots meteorologists use when forecasting a hurricane’s path. This ensemble model approach is beneficial for most industries.
“Any field where you know you cannot perfectly predict it, there should be a level of uncertainty,” Gowan explained. “To get that level of uncertainty, you need to run a bunch of simulations that cover your level of uncertainty and see what the outcome would be.”
Closing the Data Gap
Typically, weather balloons only fly for two to five hours and are limited in where they can go. WindBorne Systems designs, builds and operates balloons that can capture atmospheric profiles over oceans and other remote places around Earth. Every balloon controls its altitude and can fly autonomously, remaining in the air for weeks at a time.
“The more data we’ve found from balloons, they help better calibrate and even improve the accuracy of satellite data profiles,” Karen Ye, WindBorne’s director of commercial growth, told The Forecast. “When we’re looking at forecast accuracy, our goal is to be able to close that data gap, specifically in the atmosphere.”
WindBorne’s latest forecasting model, WeatherMesh-6 (WM-6), introduced 24 new surface, soil and atmospheric parameters alongside five upper-air parameters at 25 pressure levels. WM-6 has the largest output catalog of any AI weather model to date. By modeling a full ensemble in latent space, each ensemble member is at full resolution with physical coherence across weather parameters. WM-6 uses this approach for scenario analysis, extreme-event detection, and distributing processed statistics like percentiles and precipitation exceedance probabilities.
The company has collaborated with several organizations, including the Gates Foundation, to improve life in areas with less in-depth data collection methods. For example, ground-based weather stations in Africa have fallen by more than 75% over the past 45 years. Long-duration weather balloons enable rapid redeployment, data continuity and more extensive coverage in the region. As a result, farmers can avoid guessing what kind of weather they might encounter, while roads and flight paths can better work around dangerous events that can cause delays.
“Weather impacts all industries, all verticals,” said Ellie Yoon, head of marketing and communications at WindBorne Systems.
Yoon shared how the WindBorne team had gotten feedback about balloons being bad for the environment. An R&D engineer wondered about changing the ballast at the bottom into a buoy. “If the balloon lands in the ocean, suddenly they’re capturing ocean data like salinity or wave height,” she said.
Added Ye, “Frankly, the more data, the better — especially at such a low and affordable cost structure that we have today.”
A Strategic Approach to AI Benchmarking
Dr. Pedram Hassanzadeh, an associate professor at the University of Chicago, runs a research group focused on building better AI models and understanding how they work, why they’re successful and why they fail. For instance, AI models don’t encounter the “butterfly effect,” where one change from the past can impact the future. Therefore, models are actually more accurate in predicting what will occur tomorrow than what happened yesterday.
“AI solving a problem or giving you a prediction is one thing,” Hassanzadeh told The Forecast. “One thing we’re very interested in is the question of interpretability. How come these models can do this with just 40 years of data and be so good?”
As models continue to develop, Hassanzadeh’s team is promoting human-centered benchmarking. While pointing to a monetary amount is nice, there are other ways to track performance.
That includes understanding how information will be conveyed to the end user. An increase in Continuous Ranked Probability Score helps researchers, but it won’t mean much to farmers. They’re more interested in how a monsoon or other extreme weather event will impact the crop harvest that is the basis of their livelihood.
“The best measure of success is whether the farmer changed their behavior,” Hassanzadeh said. “You gave them a forecast about, ‘Hey, the rain is going to happen next week.’ [You want to know] whether that led to them changing what they planned, or when they planned, or whether they hire help now or wait.”
This approach is valid across industries. Did a logistics company reroute its supply chain to avoid a slowdown due to weather or closure from a political stalemate? Did doctors at a hospital change how they review patient X-rays to incorporate AI monitoring and tracking?
Building and Managing Next-Gen Distributed AI Systems
Even when there’s an effective solution, Hassanzadeh encourages regularly testing models and approaches. Organizations study and collect lots of data, often over many years. While it may not always be enough for accurate forecasting, sitting on data without testing its efficacy won’t lead to progress, he stressed.
“You never know whether you have enough data until you try. Whether 40 years of [high-quality observation-based] data was enough, we didn’t know until we tried,” Hassanzadeh said. “These AI models are really impressive in what they can do. But also, they’re not magical.”
Yoon echoed that willingness to tinker and experiment. She highlighted how WindBorne’s engineering team has evolved its balloons to do previously unachievable or untestable tasks, such as flying into the eyes of Typhoon Dolphin and Typhoon Saudel, or capturing atmospheric changes by flying balloons into the path of totality during an eclipse over Spain.
“Being at a company where people are constantly trying to optimize — truly in the mind of an engineer, seeing how the smallest revision can make the balloon more robust and more capable — has been really fun to watch,” Yoon said. “People are always testing the boundaries of what works.”
With renewable energy sources such as the sun and wind being used more frequently, companies are using forecasting to try and wrangle what traditionally have been volatile costs. Gowan said that many organizations are building their own weather insights and applications, and even doing their own modeling.
However, one thing AI models can’t commoditize is data. While anyone can utilize publicly available information, Spire’s proprietary data can, for example, help ERCOT in Texas or a country in Europe understand how much solar energy they can produce in a given timeframe. AI helps expedite that process even further.
“We’re doing a big push to completely replace this traditional data assimilation with an AI system that just takes in the raw data off the satellites and converts it into a real weather analysis where we know the temperature, the pressure and the humidity at every grid point in the world,” Gowan said.
Hassanzadeh agreed that AI will continue improving, though he urged people to consider all angles to maximize a model’s performance.
“There is a tradeoff here. Do you just want a good forecast, or do you want something that really obeys physics? Is there a way to have both?” he asked. “It’s not just brute forcing with bigger and bigger computers to train the models. You also have to think about the physics and the math, and what are the things the AI model is not seeing in the training data. That’s a really exciting next frontier.”
Related:
Joey Held is a writer, podcaster and author based in Austin, Texas, and the founder of Fun Fact Friyay and Wacky Travel. Connect with him on LinkedIn or BlueSky.
© 2026 Nutanix, Inc. All rights reserved. For additional information and important legal disclaimers, please go here.
Related Articles
