The saving is on the simulation bill, not a win over observations
The Allen Institute for AI describes HiRO-ACEv2 on Hugging Face. It has two parts: ACE runs on a coarse grid of about 100 km, and HiRO sharpens that output to 3 km. The full system can generate global 3 km climate data at about one three-hundredth the cost of X-SHiELD. X-SHiELD is a high-resolution global atmospheric model that explicitly simulates storms and other small-scale weather.
The page time is 16:12 UTC on October 14, which is 00:12 in Beijing on October 15.
[1]
The 3% is against the simulation it learned from
Given coarsened X-SHiELD data, HiROv2 matches the target high-resolution statistics within 3%, including rain, wind, and pressure events so rare they occur once in a million data points. It also estimates the total energy of tropical cyclones, including storms it never saw in training, and generates hurricanes with eyes and rain bands. Winds and pressure around the eye stay somewhat more lopsided than in X-SHiELD. The post says strong winds flowing down mountain slopes also look realistic.
That is simulation against simulation. The post does not report an error against weather stations or satellite observations.
[1]A year driven by ACE's own coarse field still lines up
They also trained a matching ACE. Over a 10-year simulation, its hurricanes, average climate, and the overall distributions of rain, wind, and pressure line up with X-SHiELD. The post treats that as evidence that ACE's coarse runs can be a base for long high-resolution records.
When the two are joined, HiRO works from a coarse field ACE ran on its own, not from real X-SHiELD output. A full year of global 3 km data from that setup still closely matches X-SHiELD. Over 10 years, hurricane energy is comparable. If only one region is needed, the post says the savings can be more than a hundred times greater, depending on the area.
[1]A different climate is still out of reach
HiRO learned from X-SHiELD, so it only works with the ACE tuned to that model. The post says it cannot yet produce high-resolution data for a warmer future, or for a historical record such as ERA5. The next plan is to train on new X-SHiELD datasets, including a world 4°C warmer, and to add near-surface temperature, snow, and ice.
A slimmer HiROv2 could, the post says, be more than ten times faster still. That is work in progress, not a measured speed. The web demo shows precomputed precipitation, wind, and sea-level pressure. The post says the pictures are a visual check, not a substitute for the numbers in the technical report.
[1]