Multi-Modal Foundation · allenai
HiRO-ACE
Climate simulation framework generating 3 km precipitation fields from 100 km inputs using diffusion downscaling.
What HiRO-ACE is
HiRO-ACE is an artificial intelligence framework developed by the Allen Institute for AI (allenai) for simulating atmospheric variability and generating fine-scale precipitation outputs. The system combines two components: ACE2S, a stochastic atmospheric emulator from the Ai2 Climate Emulator family, and HiRO (High Resolution Output), a diffusion model that downscales coarse climate data into kilometer-scale fields.
The framework addresses the computational expense of producing multi-decade, high-resolution climate simulations. By running an efficient 100 km emulator and downscaling to 3 km resolution, it enables scientists to generate 6-hourly averaged surface precipitation fields over arbitrary global regions across decades of simulated climate.
How it works
HiRO operates as a diffusion-based downscaling model that takes 100 km resolution input fields and outputs 3 km resolution surface precipitation rates. It is trained on 16x16 1-degree patches of coarsened data from the X-SHiELD climate model spanning 2014 to 2022 (a 9- to 10-year training set).
ACE2S acts as a stochastic generative emulator of 100 km coarsened X-SHiELD output, enforcing exact conservation of global dry air mass and moisture. Outputs from ACE2S serve as the coarse input driving the HiRO diffusion process without observed quality degradation.
Key facts
- Framework architecture
- Stochastic emulator (ACE2S) paired with a diffusion downscaling model (HiRO)
- Input resolution
- 100 km
- Output resolution
- 3 km
- Output variable
- 6-hourly averaged surface precipitation rate
- Training data
- 9 to 10 years of coarsened X-SHiELD simulation data (2014–2022)
- Software library
- fme
What you can use it for
- Kilometer-scale precipitation downscaling
- Downscaling 100 km resolution climate emulator outputs to 3 km, 6-hourly averaged surface precipitation rates across specified global or regional domains.
- Decadal atmospheric variability emulation
- Simulating multi-decade atmospheric dynamics with exact conservation of global moisture and dry air mass using the ACE2S emulator.
- Ensemble generation of local precipitation
- Generating stochastic ensembles of local surface precipitation variability to capture fine-scale precipitation distributions up to the 99.99th percentile.
Run it here
We list this model but cannot run it yet: no pipeline_tag — needs a human.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Reproduces the surface precipitation rate distribution of X-SHiELD out to the 99.99th percentile while capturing topography-related rainfall characteristics.
- Guarantees exact conservation of global dry air mass and total moisture within the ACE2S emulator.
- Supports stochastic generation for straightforward ensemble production of small-scale atmospheric variability.
- Exhibits low time-mean biases relative to its 10-year X-SHiELD training data across most variables.
- Integrates seamlessly between ACE2S coarse-scale simulation and HiRO downscaling without observed degradation.
Watch out for
- Trained exclusively on 9 to 10 years of X-SHiELD climate model output (2014–2022) and is not expected to generalize to past or future sea surface temperatures or carbon dioxide conditions.
- Produces features representative of the X-SHiELD simulation rather than observational datasets, making it unsuitable for operational weather forecasting.
- Cannot operate outside of 88°S to 88°N, and was not trained or tested outside of 66°S to 70°N.
- Downscaling areas larger than the 16x16 1-degree training patch size causes boundary discontinuities or blending artefacts in overlapping zones.
- Overestimates tropical cyclone frequency compared to base X-SHiELD runs and misses the strongest isolated tropical convection (representative of the 99.9999th percentile).
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · apache-2.0
The apache-2.0 licence carries terms worth reading before you depend on it.
Questions
Can I use HiRO-ACE commercially?
Check the licence before commercial use. The apache-2.0 licence carries terms worth reading before you depend on it.
What is the relationship between HiRO and ACE2S?
ACE2S is a stochastic atmospheric emulator that models climate dynamics at a 100 km scale, whereas HiRO is a diffusion model that takes those 100 km fields and downscales them to 3 km surface precipitation rates.
Can this framework be used for operational weather forecasting?
No. The framework is trained strictly on synthetic output from the X-SHiELD climate model rather than reanalysis data or real-world observations, making it inappropriate for weather forecasting.
What physical conservation constraints are preserved?
The ACE2S model enforces exact conservation of both global dry air mass and global moisture.
What geographic domains can HiRO downscale?
HiRO can downscale arbitrary regions between 88°S and 88°N, though it was only trained and evaluated between 66°S and 70°N. Areas larger than 16x16 degrees require patching, which can introduce edge artefacts.
How to cite it
arXiv:2512.18224
Written from HiRO-ACE's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.