The Allen Institute for AI (Ai2) has released HiRO-ACEv2, an updated AI system for simulating global weather and climate at 3 kilometer resolution, which it says can generate the data at about one-three-hundredth of the cost of the physics-based model it is compared with. The announcement, dated October 14, focuses on a hard problem for emulators: getting rare, extreme events right.

Two models working together

HiRO-ACE has two parts, according to Ai2. ACE is an AI climate emulator that runs at a coarse resolution of about 100 kilometers. HiRO, short for High Resolution Output, is a second model that sharpens those coarse outputs into fine 3 km detail. The coarse model keeps long simulations stable and cheap; the downscaling model adds local structure such as storms, wind patterns near mountains and intense rainfall.

The big change in version 2 is the new downscaling model, HiROv2. Ai2 says it can now predict several high-resolution fields at once instead of only rainfall, which lets it capture more kinds of extreme weather, and that it has improved skill at extreme events. Changes to training improved its ability to reproduce the most intense rain, wind and pressure values.

The numbers Ai2 reports

All figures below are from Ai2’s own evaluation against X-SHiELD, a high-resolution global atmospheric model that explicitly simulates storms and other small-scale weather:

  • Given coarsened X-SHiELD data, HiROv2 matches the target simulation’s high-resolution statistics to within 3%, even for events so rare they occur once in a million data points.
  • It estimates the total energy of tropical cyclones accurately, including storms it did not see during training, and generates hurricanes with realistic eyes and rain bands, though winds and pressure around the eye remain “somewhat more lopsided” than in X-SHiELD.
  • It realistically generates strong winds flowing down mountain slopes.
  • Over a 10-year simulation, the updated ACE produces hurricanes, average climate patterns and distributions of rain, wind and pressure that line up well with X-SHiELD.
  • Combining the two, Ai2 generated a full year of global 3 km data whose statistics closely match X-SHiELD’s, even though HiRO was working from ACE’s own simulation rather than real X-SHiELD data.
  • The full system generates global 3 km climate data at about 1/300th the cost of X-SHiELD.

These are developer-reported results; independent evaluations on observed extremes would be the real test, because matching another model is not the same as matching reality.

Why emulators are interesting

Traditional climate models solve physics equations on supercomputers. At 3 km resolution they can resolve storm-scale processes, but a single multi-decade run can take enormous resources, so researchers run few of them. An emulator trained on those simulations can produce many more samples cheaply. That helps study rare events: a once-in-a-century storm only shows up in a few simulated years, so more simulated years means better estimates of risk for infrastructure, insurance and emergency planning. Our overview of AI for climate and the environment covers the wider field, and how AI is accelerating science puts weather modeling in context.

Limits to keep in mind

An emulator learns from its training simulations, so it inherits their biases and can struggle with conditions outside its training range, such as a future climate warmer than anything it has seen. Ai2 itself notes the remaining mismatch in storm-eye wind and pressure. The value of the approach is speed and volume, and it complements rather than replaces physics models, which are still needed to generate training data and to validate results.

What to watch next

  • Independent tests of HiRO-ACEv2 against observations of real extreme events.
  • Public release details for model weights and data, which Ai2 and Hugging Face describe on their blogs.
  • Use by insurers, utilities and city planners for risk estimates.
  • Performance in warmer-climate scenarios beyond the training distribution.
  • Competing emulators from weather and climate labs.

Key takeaways

  • HiRO-ACEv2 pairs a 100 km emulator (ACE) with a 3 km downscaler (HiROv2).
  • Ai2 says it matches X-SHiELD high-resolution statistics within 3%, even for one-in-a-million events.
  • The full system reaches about 1/300th of X-SHiELD’s cost for global 3 km data.
  • Hurricane structure is good but wind and pressure around the eye are not perfect.
  • Results come from Ai2 and compare against another model, not directly against observations.

Frequently asked questions

What is an emulator?

An AI model trained on the output of a physics-based simulation so it can reproduce similar results much faster and more cheaply.

Does this replace weather forecasts?

No. It is aimed at generating climate-scale data and studying extremes, not issuing day-to-day forecasts.

Why does the 3 km resolution matter?

At that scale a model can represent storms and mountain winds that coarser grids smooth over.