New climate emulator generates high-resolution extreme weather data at 1/300th the cost of traditional models

Ai2 has released HiRO-ACEv2, a climate emulator that generates global weather data at 3 km resolution for roughly 1/300th the computational cost of X-SHiELD, a traditional high-resolution atmospheric model. The system captures extreme rain, wind, and pressure events with enough accuracy to match target statistics within 3%, even for weather so rare it appears once in a million data points. For researchers in climate risk, infrastructure planning, and disaster preparedness, this means multi-year high-resolution simulations become affordable for the first time.

How the two-part system works

HiRO-ACEv2 combines two models. ACE runs at a coarse 100 km resolution to simulate global climate patterns over long periods. HiROv2 then sharpens those outputs into 3 km detail. The previous version downscaled only rainfall. HiROv2 now predicts multiple high-resolution fields simultaneously - rain, wind, and pressure - which lets it capture a broader range of extreme weather.

The training process changed to improve performance on the most intense events. When fed coarsened X-SHiELD data, HiROv2 reproduces high-resolution statistics to within 3% accuracy. It estimates the total energy of tropical cyclones, including storms it never saw during training. The model generates hurricanes with realistic eyes and rain bands, though winds and pressure around the eye remain somewhat more lopsided than in X-SHiELD. It also realistically produces strong winds flowing down mountain slopes.

A full year of global 3 km data

The team paired HiROv2 with an updated version of ACE and generated a complete year of global 3 km data. Over a 10-year simulation, ACE's hurricanes, average climate patterns, and overall distributions of rain, wind, and pressure aligned well with X-SHiELD. When HiROv2 worked from ACE's own simulations rather than real X-SHiELD data, the resulting statistics still closely matched the target model. Hurricane energy across the 10-year run was comparable to X-SHiELD's output over the same period.

Producing global high-resolution data costs roughly 300 times less than running X-SHiELD directly. If a researcher needs only one region, the savings can exceed a hundred times that, depending on the area's size. A streamlined version of HiROv2 now in development could make the system more than ten times faster still.

Current limits and next steps

HiRO learned exclusively from X-SHiELD, so it only works with the version of ACE tuned to match that model. It cannot yet produce high-resolution data for other climates, such as a warmer future scenario or the historical record in datasets like ERA5. The team plans to train on new X-SHiELD datasets, including one simulating a world 4°C warmer. Future outputs will add near-surface temperature, snow, and ice to support applications like snowpack prediction.

For researchers who want to inspect results without running inference, Ai2 built a web demo that displays pre-generated precipitation, wind, and sea-level pressure from both HiROv1 and HiROv2 alongside X-SHiELD data. The models are available for download, and tutorial notebooks walk through the process of running them. Professionals in science and research who want to build skills with these kinds of tools can find relevant training on the same site.

Why this matters for education, real estate, construction, and research

Affordable high-resolution climate data changes the calculus for anyone whose work depends on understanding extreme weather risk. Researchers can run ensembles of 10-year simulations to study storm frequency and intensity without waiting for supercomputer allocations. Real estate developers and construction firms evaluating flood or wind exposure for long-term projects gain access to granular local projections that were previously cost-prohibitive. Educators training the next generation of climate scientists can give students hands-on experience with 3 km global data rather than relying on heavily simplified models. The cost reduction - 300 times cheaper for global runs, and far more for regional studies - turns what was once a supercomputer-only capability into something a small research group can use.