HiRO-ACEv2: Simulating extreme weather at 3 km resolution for a fraction of the cost

What it claims

HiROv2 is a downscaling model that maps 100 km ACE output to 3 km precipitation, wind and sea-level pressure. Version 1 produced rainfall only; v2 predicts these fields jointly. The headline 3% statistical match to X-SHiELD is measured when HiROv2 is fed coarsened X-SHiELD data, not ACE output. This is a downscaling-only test, so it isolates HiRO's skill from ACE's errors. The end-to-end system, with ACE running on its own, is described only qualitatively as "closely matching".

How it works

• ACE (retrained for v2): a 100 km AI emulator tuned to match X-SHiELD. Over a 10-year run, its hurricanes, mean climate and distributions of rain, wind and pressure "line up well" with X-SHiELD.
• HiROv2: sharpens ACE's coarse fields to 3 km. The gain on extremes is attributed to "changes to how we train HiROv2"; the post does not say what those changes are.
• HiRO is trained on X-SHiELD. That is why it works only with the X-SHiELD-tuned ACE.

Evidence

• Downscaling from coarsened X-SHiELD: high-resolution statistics match within 3%, including events at the one-in-a-million tail.
• Tropical cyclones: total storm energy is estimated accurately, including storms held out of training. Hurricanes show realistic eyes and rain bands.
• Mountains: strong downslope winds are generated realistically.
• Free-running ACE plus HiROv2: the team generated one year of global 3 km data, with statistics that "closely match" X-SHiELD. Over a 10-year run, hurricane energy is "comparable" to X-SHiELD over the same period. No numbers are given for this end-to-end setting.
• Cost: about 1/300th of X-SHiELD for global output. For a single region, the savings can be more than 100× greater still, depending on the area's size.

Detailed quantitative comparisons are deferred to a tech report that is not reproduced on this page.

How to use it

• Model weights are downloadable, and tutorial notebooks walk through running them. The page gives no install commands, model sizes or licence.
• A web demo shows pre-generated precipitation, 10 m wind and sea-level pressure from HiROv1, HiROv2 and X-SHiELD side by side. You can zoom to your own region to check quality there before running any inference.
• A streamlined HiROv2 that could be more than 10× faster is in progress but not released.

Limits

• Climate scope: the system covers only the X-SHiELD-tuned ACE climate. It cannot downscale warmer-future scenarios or historical reanalysis such as ERA5. A run on a 4°C-warmer X-SHiELD dataset is planned, not done.
• Missing outputs: near-surface temperature and snow/ice are not produced yet, so it is not usable for snowpack work today.
• Hurricane cores: winds and pressure around the eye are more lopsided than in X-SHiELD. Treat eyewall-scale asymmetry with caution.
• End-to-end accuracy: the 3% figure does not apply to ACE-driven runs. For those, the page offers only qualitative agreement.