HiRO-ACEv2 Climate Simulation High-Resolution Model: File List

HiRO-ACEv2 Climate Simulation High-Resolution Model: File List

At a Glance

Item Value
Repository allenai/HiRO-ACEv2
Publisher guide Ai2: models and licenses
Published 2026-10-13
License apache-2.0
Formats checkpoint (ckpt)
Paper arXiv:2610.17693
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

The Allen Institute for AI (Ai2) has released “HiRO-ACEv2," a generative model designed to increase the resolution of climate simulations. Rather than taking text or images as input, it is a diffusion model that takes 100km-resolution atmospheric data output from climate models (such as precipitation, wind speed, and sea level pressure) and generates high-definition 3km-resolution weather fields. It is reported to be capable of downscaling decades-long climate simulations for arbitrary regions on Earth.

HiRO-ACEv2 features a two-stage framework: first, a group of models called “ACE" (Ai2 Climate Emulator) simulates climate autoregressively at 100km resolution, mimicking atmospheric variability from daily to centennial scales, and then a diffusion model called “HiRO" (High Resolution Output) downscales the output to 3km resolution. The repository includes both the HiROv2 checkpoint for downscaling and the ACE2S-X-SHiELD checkpoint for input generation.

Specifications

  • Architecture: HiRO is a diffusion model that generates 3km resolution from 100km resolution input. ACE2S-X-SHiELD is a stochastic (generative) emulator that mimics X-SHiELD (high-resolution global climate simulation data provided by GFDL) at 100km resolution.
  • Output Variables: 6-hourly mean surface precipitation, 10m zonal and meridional winds (instantaneous values), and sea level pressure (instantaneous values). All are at 3km resolution with 6-hour intervals.
  • Latitude Range: Trained and validated within the 66S to 70N range; it cannot operate in regions exceeding 88S or 88N.
  • Spatial Patches: The downscaling patch size during training is 16×16 (1 degree). It is explained that downscaling larger regions results in discontinuities if patch overlap is 0, or blending artifacts if there is overlap.
  • License: apache-2.0
  • Training Data: Trained on 9 years of X-SHiELD data (2014-2022).
  • Distribution Format: Checkpoints (.ckpt). For inference, the fme package is pip-installed, and instructions are provided to update paths in YAML configuration files (ace2s_inference_config_global.yaml, hiro_downscaling_ace2s_pnw_output.yaml, etc.) and execute them.

Performance and Quality

According to the model card, HiROv2 reproduces the reference distributions of X-SHiELD for surface precipitation, 10m wind, and sea level pressure up to the 99.9999th percentile value, and is capable of generating realistic fields even in snapshots of events with strong regional impacts, such as typhoons and atmospheric rivers. It is also described as having low time-mean bias relative to X-SHiELD and being able to reproduce topography-related time-mean precipitation characteristics.

The abstract of the accompanying paper (arXiv:2610.17693) emphasizes that its inference computation cost is more than 300 times lower than conventional numerical Global Storm Resolving Models (GSRMs), allowing researchers to study the characteristics of typhoons and extreme events using large ensemble datasets. The paper reports that the tails of the precipitation and near-surface wind distributions (up to the 99.9999th percentile value) closely match the high-resolution reference distribution, and the reference distribution of typhoon intensity is roughly tracked. These are results presented by the publishers themselves, and a comparison table with independent third-party verification is not included in this material.

On the other hand, it is explicitly stated that ACE2S-X-SHiELD is trained on only 9 years of data and is not expected to generalize beyond the range of forcing conditions used during training (such as past and future SSTs or CO2), making it unsuitable for weather forecasting purposes (as it is trained solely on climate model output rather than reanalysis data). Regarding HiROv2, there is also a caveat that generalization to climates outside the 10-year range of the training data is not expected, and the features reproduced are representative of the X-SHiELD data rather than representing observational data itself.

Strengths and Use Cases

HiRO-ACEv2 serves as an extremely powerful tool for engineers and researchers who want to analyze the fine-scale structures of specific meteorological phenomena in detail. The model card emphasizes its ability to output realistic fields in reproducing the intensity of typhoons (tropical cyclones) and generating snapshots of extreme weather events with extremely large regional impacts (such as heavy rainfall and strong winds).

In particular, the ACE2S-X-SHiELD emulator improves the reproducibility of small-scale fluctuations, such as surface precipitation, compared to conventional deterministic models. It also satisfies the strict conservation laws of global dry air mass and moisture, making it suitable for simulations requiring long-term climate statistical stability. Taking advantage of its nature as a stochastic diffusion model, it can rapidly generate multiple slightly different scenarios (ensembles) from the same input conditions, making it suitable for evaluating the uncertainty of small-scale fluctuations.

HiROv2, which handles downscaling, excels at restoring detailed 3km-resolution precipitation, wind speed, and pressure distributions reflecting topographical effects from coarse 100km-resolution data. According to the paper, methods using multiple denoisers during inference, along with data sampling and loss weight optimization during training, have made it possible to accurately capture the statistical properties of rare extreme phenomena corresponding to the tails of the distribution.

This enables use cases such as running “thousands of patterns of ensemble simulations"—which was difficult with conventional numerical models due to computing time constraints—in a realistic timeframe to investigate the occurrence probabilities and intensity distributions of extreme events in detail. Because it can be executed by specifying an arbitrary region almost anywhere on the globe (within the latitude range of 66S to 70N), high-definition climate analysis focused on specific regions is possible.

Distributed Files

Weight files published in allenai/HiRO-ACEv2, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

File Size
ACE2S-X-SHiELD.ckpt 3.20GB
HiROv2.ckpt 1.38GB

How to Get It

Model weights and configuration files are available from the Hugging Face repository allenai/HiRO-ACEv2. The distribution format is checkpoints (.ckpt). This model is gated: false, and no special license agreement procedure on Hugging Face is required for download.

The fme library for climate emulators developed by Ai2 is required to run inference. Build the environment and run inference with the following steps:

  1. Download the repository. If necessary, also obtain subsets of forcing data and initial conditions for the target period.
  2. Run pip install fme to install dependencies.
  3. Update the paths (such as the location of checkpoints and datasets) in the ACE2S (100km emulator) configuration file ace2s_inference_config_global.yaml to match your environment.
  4. Update the paths in the HiROv2 (downscaling) configuration file hiro_downscaling_ace2s_pnw_output.yaml. At this time, set the input data path to point to the ACE2S output directory specified in Step 3.
  5. Run the accompanying script run-hiro-ace.sh to start the simulation.

Detailed documentation is available at Ai2 Climate Emulator Documentation, which also includes specific guides on downscaling inference.

Can You Run It Locally?

The publisher distributes this model as checkpoint (ckpt).

License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.

Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.

What to Read Next

Sources