AllenAI releases HiRO-ACEv2, a cost-efficient AI system for high-resolution extreme weather simulation

Allen Institute for AI (AllenAI) released HiRO-ACEv2, an updated version of its High Resolution Output Ai2 Climate Emulator (HiRO-ACE) system, designed to simulate extreme weather at 3 km resolution at a fraction of the cost of traditional high-resolution atmospheric models.

The system consists of two components: ACE, a coarse 100 km-resolution AI model, and HiRO, a downscaling model that sharpens ACE’s outputs into 3 km detail. HiRO-ACEv2 introduces HiROv2, an improved downscaling model that predicts multiple high-resolution fields simultaneously—including rainfall, wind, and pressure—rather than just rainfall, broadening its ability to capture a wider range of extreme weather events.

HiROv2 demonstrates improved skill in reproducing extreme events, matching X-SHiELD’s high-resolution statistics to within 3% even for events so rare they occur only once in a million data points. It also accurately estimates tropical cyclone energy, including for storms not seen during training, and generates hurricanes with realistic eyes and rain bands, though wind and pressure near storm eyes remain slightly more lopsided than in X-SHiELD.

AllenAI trained an updated ACE model alongside HiROv2. Over a 10-year simulation, the updated ACE’s hurricanes, average climate patterns, and distributions of rain, wind, and pressure aligned well with X-SHiELD, indicating that ACE’s coarse simulations provide a reliable foundation for producing long-term high-resolution climate data.

Using HiROv2 and the updated ACE, AllenAI generated a full year of global 3 km data. The system’s statistics closely matched X-SHiELD’s, even when HiROv2 operated on ACE’s own simulations rather than real X-SHiELD data. Over a 10-year run, the energy of HiRO-ACEv2’s hurricanes was comparable to X-SHiELD’s, and snapshots showed a variety of extreme events resembling those in X-SHiELD.

AllenAI reports that generating global high-resolution data with HiRO-ACEv2 costs roughly 300 times less than running X-SHiELD. For regional simulations, the savings can exceed a hundredfold, depending on the area size. The team is also developing a streamlined version of HiROv2 expected to be more than ten times faster.

HiRO-ACEv2 currently works only with the ACE version tuned to match X-SHiELD, limiting its applicability to other climates such as warmer futures or historical datasets like ERA5. AllenAI plans to train on additional X-SHiELD datasets, including one simulating a 4°C warmer world, and add outputs like near-surface temperature and snow and ice to expand use cases such as snowpack prediction.

The models are available on Hugging Face, accompanied by a web demo that lets users compare HiROv1 and HiROv2 outputs side-by-side with X-SHiELD, tutorial notebooks, and downloadable model weights for further analysis.