HiRO-ACEv2: Simulating extreme weather at 3 km resolution for a fraction of the cost

AI2 Model Slashes Cost of High-Resolution Climate Simulation

The Allen Institute for AI (AI2) has released HiRO-ACEv2, an updated AI-powered climate emulation system that can generate global weather data at a fine 3 km resolution. The system's main contribution is its economic efficiency; it produces detailed simulations of extreme rain, wind, and pressure events for roughly 1/300th the cost of traditional physics-based models like X-SHiELD. This cost reduction makes high-fidelity, long-term climate research substantially more accessible to a wider range of institutions.

Key Capabilities and Technical Details

The system functions using a two-part process: a coarse AI model named ACE operates at a 100 km resolution, and a new downscaling model, HiROv2, sharpens that output to the final 3 km detail. The key improvements in this second version focus on simulating extreme events with greater accuracy.

• Multi-Field Prediction: Unlike its predecessor, HiROv2 can predict several high-resolution variables at once, including rainfall, wind, and pressure.
• Extreme Event Fidelity: The model matches the target simulation's statistics to within 3%, even for rare events that occur once in a million data points.
• Cost Efficiency: Global data generation is ~300 times cheaper than X-SHiELD, with regional savings potentially being over one hundred times greater.
• Realistic Storms: It accurately estimates tropical cyclone energy and generates hurricanes with realistic structures, such as eyes and rain bands.

While HiRO-ACEv2 provides a powerful and inexpensive tool, its current utility is constrained by its training data. Because the model learned from X-SHiELD, it can only reproduce climates consistent with that source model and cannot yet generate data for other scenarios, such as a warmer future or historical records from datasets like ERA5. Future work aims to address this by training on new simulations, including one of a world 4°C warmer, and by adding outputs like temperature and snow.

The primary value of systems like HiRO-ACEv2 lies in shifting the core challenge of climate modeling from a dependency on massive computational budgets to a focus on acquiring diverse, high-quality training data to simulate different climate futures.