Research map / Probabilistic and scientific modeling / Differential equations and PDEs
Scaling Storm-Resolving Atmospheric AI Simulation to the Entire Planet
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces STRATA, a tile-based autoregressive transformer emulator trained on 17 days of 4.9-km SCREAM output that performs stable global 24-hour storm-resolving rollouts with ~50x better energy efficiency than SCREAM.
Problem
Global storm-resolving atmospheric simulation is extremely expensive (~1 MWh per simulated day), and no AI emulator had achieved autoregressive rollout at both global and km-scale.
Approach
Train on small local spatial tiles (trading scarce global time samples for abundant spatial samples), then apply globally via overlapping-tile blending, combined with 3D patch embedding, local 3D neighborhood attention, a stereographic rotary position embedding for grid invariance, and a pixel-space de-aliasing decoder.
Key results
- STRATA achieves 48.1 SD/MWh on 16 H100s versus 0.78 SD/MWh for SCREAM, about 50x better energy efficiency.
- Throughput of 741 simulated days per wall-clock day on 512 H100 GPUs with 98.4% strong scaling efficiency from 8 to 512 GPUs.
- Km-scale emulation requires ~10x more FLOPs per grid point than coarse AI weather models; global mean rainfall biased low by up to 20%.
Surprising findings
- Km-scale emulation needs roughly 10x more FLOPs per grid point than coarse-resolution AI weather models, attributed to higher information density of 3D convection.
- Global mean rainfall is biased low by as much as 20%, and total error grows too fast at 12-24 h on synoptic scales relative to ACE2.
Limitations
- Large-scale circulation biases: error growth too fast at 12-24 h; does not match ACE2 at synoptic scales.
- Global mean rainfall biased low by up to 20%; suppressed spectral power at 100-2000 km scales.
- Trained on only 17 days of data from three SCREAM simulations, one with erroneous SST prescription.
Experimental setup
Datasets: SCREAM sdecadal simulation SCREAM DYAMOND1 simulation SCREAM DYAMOND2 simulation
Benchmarks: SCREAM 24-hour rollouts (held-out sdecadal initializations)
Metrics: energy efficiency (SD/MWh) simulated days per wall-clock day fractions skill score RMSE growth of 850-hPa temperature and zonal wind precipitation distribution and spectra
Baselines: SCREAM ACE2 PixelDiT-M-ps4 DiT-M-ps4
Builds on: Diffusion Transformer (DiT) PixelDiT NATTEN RoPE DUO-grid halo padding
Closest papers by meaning
- Precipitation Downscaling with Spatiotemporal Video Diffusion (NeurIPS 2024 · similarity 0.86)
- Cloud4D: Estimating Cloud Properties at a High Spatial and Temporal Resolution (NeurIPS 2025 · similarity 0.85)
- ChaosBench: A Multi-Channel, Physics-Based Benchmark for Subseasonal-to-Seasonal Climate Prediction (NeurIPS 2024 · similarity 0.85)
- Satellite Observations Guided Diffusion Model for Accurate Meteorological States at Arbitrary Resolution (CVPR 2025 · similarity 0.85)
- Adaptive Flow Matching for Resolving Small-Scale Physics (ICML 2025 · similarity 0.85)
- Probablistic Emulation of a Global Climate Model with Spherical DYffusion (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.