The Allen Institute for AI released HiRO-ACEv2, an AI model that downscales coarse global weather simulations into 3-kilometer-resolution output, according to a post on the Hugging Face blog. Ai2 said the model can reproduce rare extreme rain, wind and pressure events with high accuracy.
The model builds on Ai2's earlier ACE system, which ran at roughly 100-kilometer resolution, and generates global, multiyear climate datasets at about 1/300th the cost of X-SHiELD, a high-resolution atmospheric model researchers currently rely on for similar detail, Ai2 said.
Ai2 published the model weights, tutorial notebooks and a web demo at ace.allen.ai, making the full pipeline available for researchers to run their own regional or global simulations rather than relying on the company's own infrastructure.
High-resolution climate modeling has been limited by compute cost, not by scientific understanding of the physics involved. Cutting that cost by roughly 300 times turns a simulation once reserved for well-funded national labs into something a university lab or a startup can run directly.