AI emulator mimics a storm-resolving weather model at about 1/300th the cost

The Allen Institute for AI (Ai2) has released HiRO-ACEv2, an AI system that generates global weather data at 3-kilometer resolution, including hurricanes and extreme rainfall, at what its developers say is roughly 1/300th the cost of the physics-based model it imitates. The institute announced the system on October 14, 2026, alongside a technical report posted to arXiv the day before.

HiRO-ACEv2 is an emulator: a machine-learning model trained to reproduce the output of a conventional simulation. Its target is X-SHiELD, a global storm-resolving model from the National Oceanic and Atmospheric Administration’s Geophysical Fluid Dynamics Laboratory (GFDL) that, in Ai2’s words, “explicitly simulates storms and other small-scale weather.” The report’s 10 authors are from Ai2, NVIDIA and GFDL.

Two models in a chain

The system works in two stages. The first, ACE (the Ai2 Climate Emulator), steps the global atmosphere forward six hours at a time on a coarse grid of about 100 km. It was pretrained on roughly 90 years of data from a 100 km version of the SHiELD model, then fine-tuned on nine years of X-SHiELD output coarsened to the same grid.

The second stage, HiROv2, is a diffusion model that sharpens those coarse fields to 3 km, roughly a 32-fold increase in resolution, working on patches 16 degrees on a side. The earlier version, HiRO-ACEv1, produced only rainfall. The new one also outputs winds 10 meters above the surface and sea level pressure.

The training data came from an 11-year X-SHiELD simulation starting on January 1, 2013, run on a grid of about 3.25 km. The team trained on 2014 to 2022 and held back 2023 for testing.

Speed and accuracy

The coarse stage can produce a year of global data in under five minutes, the report says, while downscaling a full global year to 3 km takes about 2.75 days on a single B200 graphics processing unit (GPU). On eight B200s, a global 3 km year finishes about 33 times faster than X-SHiELD running on 27,648 AMD EPYC 9654 processor cores, and the authors put the energy cost at roughly 0.3% of a year of X-SHiELD.

Users who need only one region save more. A single 16-degree patch for a year takes 12 minutes on a B200, 15 minutes on an H100 and 30 minutes on an L40S, according to the report, which says regional use can cut costs by up to about 250 times.

On accuracy, the authors report that the emulator “closely matches the high-resolution reference distributions of precipitation and near-surface winds out to their 99.9999th percentile values.” In their conclusions, they put all output fields within 3% of the reference at that extreme percentile. Ai2’s post describes these as events that occur “only once in a million data points.”

For tropical cyclones, the root-mean-square error in integrated kinetic energy, a measure of a storm’s total wind energy, was 0.7 terajoules across the 2023 test set and under 1 TJ even for the strongest storms. The coarse model produced 69.0 ± 6.6 tropical cyclone tracks per year, against 75.3 ± 6.8 in X-SHiELD, with fewer weak storms and more strong ones. Peak winds in the strongest storms, above roughly 40 meters per second, tend to be underpredicted, the report says.

Why it matters

Running a storm-resolving model such as X-SHiELD ties up tens of thousands of processor cores, as the report’s comparison shows. Ai2 presents HiRO-ACEv2 as “a flexible and inexpensive way to generate many versions of high-resolution rain, wind, and pressure data for extreme events over 10 years or more.” In a 10-year run, the institute says, hurricanes, average climate patterns and the distributions of rain, wind and pressure “all line up well with X-SHiELD.”

The model weights are available on Hugging Face and the code is on GitHub, so other groups can run the emulator for their own regions. Ai2 says it is working on a streamlined version of HiROv2 that “could make it more than ten times faster still,” and on new X-SHiELD datasets, including one simulating a world 4°C warmer.

The caveats

The technical report is a preprint and has not been peer reviewed. The system reproduces X-SHiELD’s own simulated climate, not observations, and the authors say it cannot yet produce data for a warmer climate or for historical conditions such as those in the ERA5 reanalysis.

The authors list other limits. Each time step is generated independently, which limits how coherently storms evolve along their tracks. Downscaled hurricane eyes are less radially symmetric than in X-SHiELD, and rain bands are less coherent. The coarse model has rainfall biases over the tropical west Pacific and South America, and the outputs do not yet include near-surface temperature or frozen precipitation such as snow.

Ai2 is supported by the Allen Fund for Science and Technology, according to the report, and the authors thank GFDL for running the X-SHiELD simulation. The report states no conflicts of interest.

Corrections and tips: email the editor using the address on the Contact page.