Scaling Storm-Resolving Atmospheric AI Simulation to the Entire Planet

2026-06-30

AI Summary

Global atmospheric models struggle to efficiently simulate convection at kilometer scales, while high-resolution physical models incur prohibitive computational costs. This work proposes STRATA—the first autoregressive AI simulator for global storm-resolving modeling—leveraging a local spatiotemporal dynamics assumption to train on small spatial patches and seamlessly stitch predictions into global forecasts. Key innovations include a novel overlapping-tile hybrid framework for global kilometer-scale AI simulation, Stereographic Rotary Position Embedding (StereoRoPE) for grid-agnostic positional encoding, and a pixel-space anti-aliasing decoder to suppress tiling artifacts. Trained on only 17 days of high-frequency data at 4.9 km resolution, STRATA stably generates 24-hour global simulations, achieving 741 simulated days per day on 512 H100 GPUs with an energy efficiency of 48 simulated days per megawatt-hour—representing a 50-fold improvement over conventional physics-based models.

Abstract

Kilometer-scale convection shapes precipitation extremes, tropical organization, and cloud feedbacks, but most global atmospheric models approximate these processes at 25-100 km resolution. Global storm-resolving physics models resolve convective systems explicitly, but at a cost -- roughly one MWh per simulated day on exascale supercomputers -- that limits long-duration simulation. We introduce STRATA (Storm-resolving Tile-based autoRegressive Atmosphere Transformer Architecture), the first autoregressive AI emulator for global storm-resolving atmospheric dynamics. STRATA is trained on the highest-resolution atmospheric dataset yet used for global AI emulation: 17 days of SCREAM physics-model output at 4.9-km resolution (~25 million grid cells) sampled every 10 minutes. Our central premise is that on 10-minute timescales atmospheric dynamics are predominantly local, so training on small spatial tiles trades scarce global temporal samples for abundant local spatial samples and enables global rollout via overlapping-tile blending. STRATA combines 3D patch embedding and local 3D neighborhood attention, a novel Stereographic Rotary Position Embedding (StereoRoPE) for grid-invariant encoding, and a pixel-space de-aliasing decoder that suppresses patch-scale rollout artifacts. An iso-FLOP scaling study reveals that km-scale emulation requires ~10x more FLOPs per grid point than coarse-resolution AI weather models, consistent with the higher information density of convective-scale dynamics. Trained on only 17 days of data, STRATA produces stable 24-hour global rollouts with realistic km-scale dynamics across diverse regimes, though large-scale biases develop with lead time. It achieves 48 simulation days per megawatt-hour -- about 50 times better energy efficiency than the SCREAM physics model -- and 741 simulated days per wall-clock day at 512 H100 GPUs. Code and dataset are publicly available.

Problem

• storm-resolving
• atmospheric simulation
• convection
• global modeling
• computational cost

Innovation

• storm-resolving modeling
• autoregressive AI emulator
• Stereographic Rotary Position Embedding
• tile-based rollout
• kilometer-scale atmospheric simulation