January 21, 2026
W. Andre Perkins, Oliver Watt-Meyer, and Christopher S. Bretherton - Ai2
Imagine having access to a suite of information about potential future climate conditions in your own local area. What is the range of expected seasonal snowfall in the mountains that provide drinking water and electricity to your city? What might the heaviest summertime rainfall events look like for your city’s necessary stormwater system upgrades? Are there any shifts in the character of tropical storms impacting your coastal region in the coming decades?
These are all difficult problems to address. Answering these questions requires simulating the atmosphere at scales fine enough to capture the storms and terrain interactions that matter most. Typical climate models run at coarse resol…
January 21, 2026
W. Andre Perkins, Oliver Watt-Meyer, and Christopher S. Bretherton - Ai2
Imagine having access to a suite of information about potential future climate conditions in your own local area. What is the range of expected seasonal snowfall in the mountains that provide drinking water and electricity to your city? What might the heaviest summertime rainfall events look like for your city’s necessary stormwater system upgrades? Are there any shifts in the character of tropical storms impacting your coastal region in the coming decades?
These are all difficult problems to address. Answering these questions requires simulating the atmosphere at scales fine enough to capture the storms and terrain interactions that matter most. Typical climate models run at coarse resolutions (e.g., 100 km grid resolution) to affordably run decades to centuries of climate evolution, but the tradeoff is that they fail to resolve details necessary for local impacts. Global "storm-resolving" models running at 3-kilometer grid resolution can capture these important details, but they’re extraordinarily expensive—a single decade-long simulation can take months of time and consume approximately 21 years of average U.S. household electricity. This computational barrier has limited the utility of fully global kilometer-scale climate simulation for downstream applications.
In collaboration with NOAA’s Geophysical Fluid Dynamics Laboratory (GFDL), we’ve developedHiRO-ACE, a two-stage AI framework that provides a pathway to dramatically more accessible high-resolution climate information. HiRO-ACE is trained to emulate outputs from a 10-year atmospheric simulation from X-SHiELD, GFDL’s state-of-the-art 3 km global storm-resolving model. As such, HiRO-ACE can generate decades of 3 km resolution precipitation data for any region of the globe in a single day on one GPU, opening new possibilities for research on local climate adaptation planning and extreme event risk assessment.
How HiRO-ACE works
HiRO-ACE combines two specialized AI components that work in tandem:
**ACE2S (Stochastic Climate Model Emulator). **First, we run ACE2S, a new stochastic variant of ourAi2 Climate Emulator, at 100 km resolution. ACE2S generates global atmospheric conditions (temperature, humidity, winds, precipitation) rollouts in 6-hour steps. The "stochastic" aspect is crucial: rather than producing smoother outputs typical of past deterministic AI-based atmospheric models, ACE2S preserves the grid-scale precipitation characteristics that physics-based models exhibit.
**HiRO (High Resolution Output Downscaler). **Then, HiRO takes the coarse 100 km precipitation and wind fields from ACE2S and generates 3 km precipitation for any region of interest. This is a 32× downscaling; HiRO transforms a fuzzy blob of regional rainfall into the detailed spatial structure of individual tropical cyclones, mountain rain events, and convective thunderstorms.
Both components are trained on the same data from a 10-year X-SHiELD simulation, ensuring they work together seamlessly without additional fine-tuning. Importantly, both are probabilistic: users can generate ensembles to quantify uncertainty at both the weather scale (ACE2S) and the storm scale (HiRO).
From coarse blobs to high-fidelity storms, accurate fine-scale details and statistics
Overall, HiRO-ACE captures the important details of the parent X-SHiELD simulation from single-event extremes to long-term averages. It accurately reproduces the probability of rainfall rates across the globe through the 99.99th percentile, and it generates realistic storm structures for a variety of high-impact atmospheric events such as tropical cyclones, atmospheric rivers interacting with complex terrain, or heavy convective thunderstorms. For climate applications, HiRO-ACE displays low time-mean biases compared to the original X-SHiELD simulation it is emulating, with relative errors less than 10% almost everywhere. It recovers the details of time-mean precipitation in regions of complex topography where a few km can change the rainfall amounts by 50% or more.
Finally, as in our team’s previous ACE releases, the major leap in capability is one of efficiency. On a single NVIDIA H100 GPU, the ACE2S climate emulator simulates ~1,500 years in a single day, while HiRO can generate a year of downscaled outputs over a single region in about 45 minutes. With this, a researcher could generate decades of regional 3 km precipitation in a day, which would take months with the original physics-based simulation.
A climate emulator for the broader community
Since its debut, the Ai2 Climate Emulator (ACE) has been adopted by a growing set of academic researchers and governmental modeling centers interested in exploring how AI-based emulation can complement traditional climate models.
Academic users who use ACE emphasize its advantages for research. “Our group has made extensive use of ACE2 because it offers a unique combination of speed and fidelity,” says Elizabeth Barnes, Professor and Dalton Family Chair in Environmental Data Science & Sustainability at Boston University. “It has enabled us to study processes like tropical cyclone formation in ways that would be prohibitively expensive with traditional climate models. This allowed us to pinpoint the causes of periods of reduced tropical cyclone activity in the far past.”
With the release of the HiRO-ACE model, which introduces high-resolution downscaling capabilities, the potential applications of ACE are expanding beyond its initial academic and government user base. HiRO-ACE is designed to translate large-scale climate signals into more localized information, an area of strong interest for practitioners engaged in climate risk assessment and impact analysis.
That potential is already attracting attention. Josh Hacker, Chief Science Officer at Jupiter Intelligence, views HiRO-ACE as a promising addition to the climate analytics toolbox. “This is a powerful and exciting technology,” Hacker says. “The ACE2 model has allowed us to assess the probability of multiple extreme event types with various durations and spatial extents—something not possible using prior methods. HiRO-ACE will bring this capability to the local scale that stakeholders need.”
Taken together, these perspectives reflect a moment of exciting expansion for ACE: a system shaped by close collaboration with academic and governmental users, and now expanding into a broader ecosystem where cutting-edge climate science directly informs real-world risk, resilience, and adaptation decisions.
Try it yourself
The HiRO-ACE paper is available onarXiv, and our models and code are available for download on Hugging Face and GitHub, respectively. For those interested in the broader suite of Ai2 climate emulators, check out our blog posts on ACE2 and SamudrACE.