HiRO-ACEv2: Simulating extreme weather at 3 km resolution for a fraction of the cost
Hugging Face
Hugging Face launched HiRO-ACEv2, an AI system that makes 3 km climate data. It matches extreme storms closely, at about 1/300th the cost of X-SHiELD.
Based on reporting by Hugging Face — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Hugging Face has updated its climate emulator again. HiRO-ACEv2 is the new version of a two-part system: ACE makes coarse global climate output at 100 km resolution, then HiRO sharpens that into 3 km detail. The pitch is simple and pretty bold: you get storm-scale weather without paying storm-scale compute bills.
The big change is HiROv2, the downscaler. It no longer focuses on rainfall alone. It now predicts several high-resolution fields at once, which lets it track a wider range of extreme weather, and the training changes behind it improved its skill on the most intense rain, wind, and pressure values. On coarsened X-SHiELD data, it matched the target simulation’s high-resolution statistics to within 3% even for events so rare they show up only once in a million data points.
It also seems to handle the stuff people care about when they say “extreme weather” and not just “nice maps.” The model estimated tropical cyclone energy accurately, including storms it never saw in training, and it produced hurricanes with realistic eyes and rain bands. Winds and pressure around the eye were still a bit lopsided compared with X-SHiELD, but the system also generated strong downhill winds along mountain slopes in a realistic way.
Hugging Face says the updated ACE component also held up over a 10-year simulation, with hurricanes, average climate patterns, and the overall rain, wind, and pressure distributions lining up well with X-SHiELD. Put together, ACE and HiROv2 generated a full year of global 3 km data whose statistics closely matched X-SHiELD, even though the emulator was fed ACE’s own simulation rather than the original high-resolution run.
The cost story is the other half of the appeal. Hugging Face says global high-resolution output is roughly 300 times cheaper than running X-SHiELD, and the savings can be more than a hundred times greater if only one region is needed. The catch is that HiRO still depends on the ACE version tuned to X-SHiELD, so it can’t yet handle warmer futures or historical datasets like ERA5. That’s next on the list, along with extra outputs such as near-surface temperature, snow, and ice.
My take — AI-written commentary, not fact-checked reporting
This is the kind of AI climate work that actually earns the hype. Not because it’s magical, but because it is brutally practical: cheaper storms, finer detail, fewer excuses. The real test now is whether these models can move beyond one tuned climate and still keep their nerve.
Read more about this at: Hugging Face