Ai2's AI Maps Hurricanes and Downpours at 3 km Resolution for 300 Times Less Than a Physics Model
Maps of hurricanes and heavy rainfall at 3 km resolution are usually produced with the physics-based X-SHiELD model on supercomputers, which is expensive. Ai2 trained an AI system: ACE builds a rough picture at 100 km, and HiROv2 adds the details, much like upscaling a photo. The result costs 300 times less, and statistics for rare storms differ from X-SHiELD by no more than 3%. That means researchers can run dozens of scenarios instead of one or two.
AI-processed from Ai2 (Allen Institute); edited by Hamidun News
To understand how hurricanes, heavy rain and storm winds will change, climatologists need detailed weather maps covering many years ahead. "Detailed" here means a 3 km grid: at that scale, individual thunderstorms, the eye of a hurricane and rain bands are visible. Such data is usually produced by X-SHiELD, a physical model of the atmosphere that explicitly simulates storms and other small-scale processes, and each run of it is very expensive. The Allen Institute for AI (Ai2) has presented HiRO-ACEv2, an updated AI system that produces similar data at roughly 300 times lower cost.
How the system works
The idea will be familiar to anyone who has seen a neural network sharpen a photo: first there is a blurry image, then the details are reconstructed from it. HiRO-ACE consists of two models. ACE simulates the planet's climate roughly, at a resolution of about 100 km. HiRO then takes that coarse picture and "increases the resolution" to 3 km.
X-SHiELD served as the teacher: the system learns to reproduce its data. But a single run costs hundreds of times less than a calculation with the physical model.
The main change in version v2 is the new HiROv2 model. The first version could only predict precipitation. Now the model outputs several fields at once: rain, wind and pressure, so the data can show far more types of extreme weather. In addition, the team changed the training scheme so that the model better reproduces the strongest values, not just ordinary days.
What the numbers show
When HiROv2 is fed coarsened X-SHiELD data, its output statistics match the reference within 3%. This holds even for events so rare that they occur once per million data points. The model accurately estimates the total energy of tropical cyclones, including storms that were absent from the training set. Hurricanes come out with realistic eyes and rain bands. There is one caveat: the wind and pressure around the eye come out more "skewed" than in X-SHiELD. The model also reproduces strong winds flowing down mountain slopes.
ACE was updated separately. In a ten-year run, its hurricanes, average climate patterns and distributions of rain, wind and pressure agree well with X-SHiELD. This matters because in practice HiRO does not work with "real" coarse data, but with what ACE itself calculates.
Combining both models, the researchers obtained a full year of global data at 3 km resolution. Its statistics are close to X-SHiELD. In a ten-year simulation, hurricane energy is comparable to the reference, and the output also shows other extreme events similar to those in the physical model.
Why it matters for practitioners, and the limitations
The main gain is the number of scenarios. When a single simulation is equivalent to a separate supercomputing project, only a few runs are feasible. If the cost falls roughly 300-fold, researchers can generate many versions of a decade-long climate and estimate the probabilities of rare events. Insurers, energy companies, urban planners and risk specialists need such scenario sets more than a single attractive map. If only one region is of interest, the savings may be even more than a hundredfold, depending on the size of the area. Ai2 is also preparing a simplified version of HiROv2 that should run at least ten times faster still.
The limitations are significant. HiRO was trained on X-SHiELD, so it only works with the version of ACE tuned to that model. It does not yet produce data for other climates, such as a warmer future or the ERA5 historical archive. Moreover, the quality in the publication is assessed by comparison with X-SHiELD, not with direct observations.
Plans include training on new X-SHiELD datasets, including a world 4 °C warmer, as well as adding near-surface temperature, snow and ice. This would open the way, for example, to forecasting snow cover.
The results can be viewed without running any code: Ai2 has built a visualizer where precipitation, wind and pressure from HiROv1 and HiROv2 are shown side by side with X-SHiELD, and users can zoom into their own region and judge quality by eye. For deeper analysis, the models are openly available, with instructions for running them.
- Aug 31, 2026Trump's Cuts to Weather Data Threaten Less Reliable Forecasts
- Aug 29, 2026First Street Study: 80% of Data Centers Are Vulnerable to Floods, Wildfires and Hurricanes
- Aug 11, 2026DeepMind's WeatherNext Predicts Hurricanes Earlier Than Existing Forecast Models
- May 23, 2026AI Weather Forecasting: WeatherNext Helped Save Lives During Hurricane Melissa
Key Questions Answered
How the system works?
The idea will be familiar to anyone who has seen a neural network sharpen a photo: first there is a blurry image, then the details are reconstructed from it. HiRO-ACE consists of two models. ACE simulates the planet's climate roughly, at a resolution of about 100 km. HiRO then takes that coarse picture and "increases the resolution" to 3 km. X-SHiELD served as the teacher: the system learns to reproduce its data. But a single run costs hundreds of times less than a calculation with the physical model.
What the numbers show?
When HiROv2 is fed coarsened X-SHiELD data, its output statistics match the reference within 3%. This holds even for events so rare that they occur once per million data points. The model accurately estimates the total energy of tropical cyclones, including storms that were absent from the training set. Hurricanes come out with realistic eyes and rain bands. There is one caveat: the wind and pressure around the eye come out more "skewed" than in X-SHiELD.
Why it matters for practitioners, and the limitations?
The main gain is the number of scenarios. When a single simulation is equivalent to a separate supercomputing project, only a few runs are feasible. If the cost falls roughly 300-fold, researchers can generate many versions of a decade-long climate and estimate the probabilities of rare events. Insurers, energy companies, urban planners and risk specialists need such scenario sets more than a single attractive map.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.