HiRO-ACE brings climate AI to 3-kilometer simulations

HiRO-ACE can generate decades of climate data at 3-kilometer resolution in a single day using just one GPU. The system from Ai2 and NOAA aims to make a task accessible that previously required months of computation and enormous amounts of electricity.

Resolution matters because a climate model with 100-kilometer cells may indicate that a region will be wetter, but it cannot clearly explain what will happen in a city, valley, or coastal area. At that scale, details such as the interaction between a storm and a mountain, or the exact structure of a tropical cyclone, are lost.

Global models operating at a resolution close to 3 kilometers can represent those phenomena. The problem is the cost: according to Ai2, a ten-year simulation can take months and consume an amount of electricity equivalent to 21 years of consumption by an average US household.

Two AI models working together

HiRO-ACE aims to reduce that cost by imitating the results of X-SHiELD, the 3-kilometer global storm model developed by NOAA's GFDL laboratory. Its training is based on a ten-year atmospheric simulation from X-SHiELD.

The system combines two components:

• ACE2S generates global atmospheric conditions at 100 kilometers, including temperature, humidity, wind, and precipitation, in six-hour steps.
• HiRO takes that data and turns it into precipitation at 3-kilometer resolution for a specific region. That is a 32-fold increase in resolution.

ACE2S adds an important feature: it is probabilistic. Instead of always producing the same smoothed prediction, it can generate different scenarios and better preserve rainfall variability. HiRO also works probabilistically, allowing researchers to create simulation ensembles and measure uncertainty.

What it achieves and what it does not mean

In the tests described by Ai2, HiRO-ACE reproduces X-SHiELD's precipitation patterns up to the 99.99th percentile, a benchmark that makes it possible to study extremely intense rainfall. It also represents structures associated with tropical cyclones, atmospheric rivers over complex terrain, and convective storms.

The model records relative errors below 10% in nearly all areas when precipitation averages are compared with the original simulation. In mountainous areas, where a few kilometers can change accumulated rainfall by 50% or more, it recovers much of the detail lost at coarser resolutions.

There is one important qualification: HiRO-ACE emulates X-SHiELD. It does not automatically replace real-world observations or guarantee that all of its results are correct. Its usefulness depends on the quality of the physical model it is trying to reproduce and on how its scenarios are used.

What changes for local planning

On an NVIDIA H100 GPU, ACE2S can simulate about 1,500 years in one day. HiRO needs around 45 minutes to generate one year of detailed data for a region. Together, they make it possible to produce decades of precipitation data at 3 kilometers in a day instead of waiting months.

That jump could help with specific decisions:

• Size drainage systems for heavier rainfall.
• Estimate changes in mountain snowpack that supplies water and electricity to a city.
• Assess how tropical cyclones could affect a coastal area.
• Compare multiple scenarios before designing adaptation projects.

Ai2 has published the HiRO-ACE paper on arXiv and offers the models on Hugging Face and the code on GitHub. The next thing to watch is how it performs outside the ten-year simulation used for training and what results it produces when its data informs real-world decisions about climate risk.