HiRO-ACEv2 Simulates Extreme Weather at 3 km Resolution at a Fraction of the Cost
HiRO ACEv2 Simulates Extreme Weather at 3 km Resolution at a Fraction of the Cost HiRO ACEv2 is the latest version of the High Resolution Output Ai2 Climate Emulator (HiRO ACE)....
By AI Engineering Team
HiRO-ACEv2 Simulates Extreme Weather at 3 km Resolution at a Fraction of the Cost
HiRO-ACEv2 is the latest version of the High Resolution Output Ai2 Climate Emulator (HiRO-ACE). It is designed to reproduce rare rain, wind, and pressure events while generating global climate data at 3 km resolution at approximately 1/300th the cost of X-SHiELD, a high-resolution global atmospheric model that explicitly simulates storms and other small-scale weather.
HiRO-ACE, introduced earlier in 2026, combines two models:
- ACE, an AI model that operates at a coarse 100 km resolution.
- HiRO, a downscaling model that converts those coarse outputs into detailed 3 km fields.
The main change in HiRO-ACEv2 is HiROv2. Unlike the earlier version, it predicts several high-resolution fields simultaneously rather than focusing only on rainfall. This allows it to represent a broader range of extreme weather. Changes to the training process also improved its ability to predict unusually intense events.
Reproducing extreme weather
The HiROv2 training changes improved its reproduction of the most extreme rain, wind, and pressure values. When supplied with coarsened X-SHiELD data, the model matches the target simulation's high-resolution statistics within 3%, including events that occur only once in a million data points.
HiROv2 also estimates the total energy of tropical cyclones accurately, including storms that were not present in its training data. It generates hurricanes with realistic eyes and rain bands, although winds and pressure near the eye remain somewhat more asymmetrical than those in X-SHiELD. The model also produces realistic strong winds flowing down mountain slopes.
An updated version of ACE was trained for the system as well. Across a 10-year simulation, its hurricanes, average climate patterns, and overall rain, wind, and pressure distributions correspond closely with X-SHiELD. These results indicate that ACE's coarse simulations can provide a foundation for generating long-term, high-resolution climate data.
When ACE and HiROv2 are combined, the system can generate a complete year of global data at 3 km resolution. The resulting statistics closely match X-SHiELD, even though HiRO uses a simulation produced independently by ACE rather than actual X-SHiELD output. Over a 10-year run, the energy of HiRO-ACEv2's hurricanes is comparable to that of X-SHiELD's hurricanes during the same period. Snapshots from the run also show several other extreme events resembling those in X-SHiELD.
Lower-cost high-resolution climate data
HiRO-ACEv2 provides a way to generate multiple versions of high-resolution rain, wind, and pressure data for extreme events across periods of 10 years or more. Producing global high-resolution data with the system costs roughly 300 times less than running X-SHiELD. For a single region, the savings can exceed a factor of 100, depending on the area's size.
A streamlined version of HiROv2 is also being developed and could make the model more than 10 times faster.
The system currently has an important limitation: because HiRO was trained on X-SHiELD, it works only with the version of ACE configured to match X-SHiELD. It cannot yet generate high-resolution data for other climates, including a warmer future climate or historical records from datasets such as ERA5.
Planned work includes training the system on additional X-SHiELD datasets, including one representing a world that is 4°C warmer. Future versions are also expected to add fields such as near-surface temperature, snow, and ice, enabling applications such as snowpack prediction.
Visualizing and using HiRO outputs
A web demo displays pre-generated precipitation, wind, and sea-level pressure outputs from HiROv1 and HiROv2 alongside X-SHiELD, the simulation used for training. Users can compare the outputs visually and zoom into particular regions.
The models are also available for download, and tutorial notebooks explain how to run them for more detailed analysis.