Climate models that can see individual storms are wonderful, and also expensive enough that most researchers can only admire them from a distance. Ai2 has released HiRO-ACEv2, an AI system that produces global 3 km climate data at about 1/300th the cost of X-SHiELD, the high-resolution physics model it learned from. The part that matters is that it does not only get the average weather right. It reproduces the rare, ugly extremes that people actually build flood defenses around.
What the system is
HiRO-ACE, the High Resolution Output Ai2 Climate Emulator, was introduced earlier in 2026. It has two parts that do different jobs.
- ACE is an AI model that simulates climate at a coarse 100 km resolution.
- HiRO takes those coarse outputs and sharpens them into 3 km detail.
X-SHiELD is the reference point throughout. It is a high-resolution global atmospheric model that explicitly simulates storms and other small-scale weather, which is exactly why it is costly to run. HiRO was trained on it, and everything in the new release is judged against it.
What changed in version two
The biggest change is the downscaler. HiROv2 replaces the original with two upgrades.
First, it predicts several high-resolution fields at once. The earlier version handled rainfall only. The new one covers rain, wind and pressure together, which widens the range of extreme weather it can describe.
Second, it is better at the extremes themselves. Ai2 says this was the main goal, and that changes to how HiROv2 is trained improved its handling of the most intense rain, wind and pressure values.
How good are the extremes?
The headline number comes from a controlled test. When HiROv2 is given coarsened X-SHiELD data, its high-resolution statistics match the target simulation to within 3%. That holds even for events so rare they show up only once in a million data points. Rare events are where emulators usually fall apart, so this is the figure to look at first.
Some specifics from the evaluation:
- Tropical cyclones. HiROv2 accurately estimates the total energy of these storms, including storms it never saw during training. It generates hurricanes with realistic eyes and rain bands.
- A caveat on those hurricanes. Winds and pressure around the eye are still somewhat more lopsided than in X-SHiELD. Ai2 says so plainly, which is more than most model announcements manage.
- Terrain. The model realistically generates strong winds flowing down mountain slopes.
The coarse model has to be good too
A sharp downscaler is no use if the coarse simulation underneath it wanders off. So Ai2 also trained an updated version of ACE for this system. Over a 10-year simulation, its hurricanes, average climate patterns, and overall distributions of rain, wind and pressure line up well with X-SHiELD. Ai2 reads that as evidence that ACE's coarse runs are a sound base for long-term high-resolution data.
Putting the two halves together
The real test is the full pipeline, because errors in the first stage can pile up in the second. Ai2 generated a full year of global 3 km data with HiROv2 and ACE working together. The statistics closely match X-SHiELD's, even though HiRO was working from a simulation that ACE ran on its own, not from real X-SHiELD data.
A 10-year run tells a similar story. The energy of HiRO-ACEv2's hurricanes is comparable to X-SHiELD's over the same period. Snapshots from that run show a variety of other extreme events that resemble those in X-SHiELD. Ai2 shows one example, a strong hurricane in the North Atlantic.
The cost argument
The economics are the point of the exercise. Producing global high-resolution data with HiRO-ACEv2 costs roughly 300 times less than running X-SHiELD. If a researcher needs only one region, the savings can be more than a hundred times greater, depending on the size of the area. Ai2 is also working on a streamlined version of HiROv2 that could be more than ten times faster still.
The practical use is volume. Ai2 pitches the system as a flexible, inexpensive way to generate many versions of high-resolution rain, wind and pressure data for extreme events over 10 years or more. Rare events are rare, so the way to study them is to have a lot of simulated years. That is hard to do when each year costs what a full physics run costs.
What it cannot do yet
The limitation is stated up front, and it is a real one. Because HiRO learned from X-SHiELD, it only works with the version of ACE tuned to match X-SHiELD. That means it cannot yet produce high-resolution data for other climates. A warmer future is out. So is the historical record in datasets like ERA5.
In other words, it is an excellent emulator of one particular simulated world. Whether it can emulate a different one is the next piece of work.
What comes next
Ai2 lists two directions.
- More climates. The team plans to train on new X-SHiELD datasets, including one simulating a world 4°C warmer, so the system can handle different climates.
- More outputs. Planned additions include near-surface temperature and snow and ice. Ai2 says these would open up uses such as snowpack prediction.
Try it without running anything
For people who want to look before they commit to downloading anything, Ai2 built a web demo. It shows pre-generated precipitation, wind and sea-level pressure output from HiROv1 and HiROv2 side by side with X-SHiELD. Users can zoom in on their own region of interest to judge how the model does there. Ai2 frames it as a quick visual check that complements the quantitative comparisons in its tech report.
The ACE data visualizer also pairs the two resolutions. One view shows simulated 10 m wind speed over Baja California, with ACE2 at 100 km on the main map and HiRO-ACEv2 at 3 km in an inset. Another shows a simulated storm over the northeastern Pacific, where the 3 km inset reveals finer structure inside the storm than the coarse map can.
For deeper analysis, the models are available to download, and tutorial notebooks walk through running them.
The takeaway
The interesting claim is not that an AI model can draw a pretty hurricane. It is that the statistics of once-in-a-million events survive the trip from a very expensive simulation to a very cheap one, at least within the world X-SHiELD describes. The catch is the last clause. Until the warmer-world training data arrives, this is a fast and affordable way to study one climate, and the question everyone will ask next is about the other ones.