HiRO-ACEv2 Simulates Extreme Weather at 3km Resolution 300x Cheaper

HiRO-ACEv2 3km resolution climate emulator

Researchers at the Allen Institute for AI just dropped HiRO-ACEv2, an updated climate emulator that churns out global weather data at three-kilometer resolution for roughly one three-hundredth the compute cost of the high-res atmospheric model X-SHiELD. Announced October 14, 2026, the system stitches together a coarse global model called ACE running at one hundred kilometers with a downscaling component, HiROv2, that sharpens outputs to three kilometers. This new version expands the downscaler’s repertoire beyond rainfall to include wind and pressure fields, and it shows markedly better skill at reproducing extreme events.

The original HiRO-ACE debuted earlier in 2026. Its two-part architecture let researchers run long climate simulations at coarse resolution and then enhance only the regions or variables they needed at fine scale. HiRO-ACEv2 keeps that structure but swaps in a downscaling model trained to output several high-resolution fields simultaneously. That tweak lets the system capture a wider range of extreme weather, from intense downpours to damaging wind gusts and deep pressure drops. When tested against coarsened X-SHiELD data, HiROv2 matches the target simulation's high-res statistics to within three percent, even for events so rare they show up only once in a million data points.

HiRO-ACEv2 tropical cyclone simulation accuracy

The system's handling of tropical cyclones really stands out. HiROv2 accurately estimates the total energy of hurricanes, including storms it never saw during training. Generated storms sport realistic eyes and rain bands, though the developers admit winds and pressure around the eye remain somewhat more lopsided than in X-SHiELD. The model also produces believable downslope windstorms flowing off mountain ranges, a phenomenon coarse models typically miss.

The coarse model ACE got its own update for this release. Over a ten-year simulation, its hurricane stats, mean climate patterns, and overall distributions of rain, wind, and pressure align well with X-SHiELD. That alignment matters because HiROv2 operates on ACE's output rather than real X-SHiELD data in production. When the full HiRO-ACEv2 pipeline runs for a full year globally, its statistics closely match X-SHiELD's despite the downscaler never seeing the high-res truth during that run. A ten-year integration shows hurricane energy comparable to X-SHiELD over the same period, and snapshots reveal a variety of extreme events that resemble the reference simulation.

HiRO-ACEv2 compute cost savings vs X-SHiELD

Cost savings are the practical headline. Producing global three-kilometer data with HiRO-ACEv2 costs about three hundred times less than running X-SHiELD directly. If a researcher only needs one region, the savings can exceed a hundred times depending on the area's size. The team is also working on a streamlined version of HiROv2 that could make inference more than ten times faster still.

Limitations remain, though. Because HiROv2 learned from X-SHiELD, it only works with the specific version of ACE tuned to match that simulator. The system can't yet produce high-res data for other climates, such as a warmer future scenario or historical reanalysis datasets like ERA5. The developers plan to train on new X-SHiELD datasets, including one simulating a world four degrees Celsius warmer, so the emulator can handle different climate states. They also intend to add near-surface temperature, snow, and ice outputs, which would enable applications like snowpack prediction.

HiRO-ACEv2 Hugging Face download access

For builders and researchers, the models are available now on Hugging Face under the identifier allenai/HiRO-ACEv2. A web demo lets users compare pre-generated precipitation, wind, and sea-level pressure outputs from both HiRO versions side by side with X-SHiELD, including the ability to zoom into any region. Tutorial notebooks walk through running the models locally. Anyone working on climate risk assessment, renewable energy siting, or regional adaptation planning can start experimenting with high-resolution ensembles that were previously out of reach computationally.

Source: Hugging Face