Global Storm AI: How STRATA Rewrites Atmospheric Modeling Rules ✛ ✦
Context
Atmospheric processes that drive extreme precipitation and cloud feedbacks remain the main source of uncertainty in climate predictions. Current numerical models of the global atmosphere, forced to parameterize convection due to coarse resolution (25–100 km), suffer from systematic errors. Global storm-resolving models (GSRMs) explicitly capture deep convection on kilometer-scale grids, but they guzzle enormous amounts of energy—about 1 MWh per model day. As John von Neumann, who laid the foundations of numerical weather prediction, pointed out, the key to accuracy lies in resolution—and today we stand on the brink of achieving that with AI.
Methods
The STRATA approach is built on a tile strategy: it trains on small spatial patches (128×128 cells, ~576 km) with 10-minute steps, amassing a wealth of local examples. This locality makes physical sense—in 10 minutes, acoustic waves travel only about 200 km. The architecture features 3D patch embeddings, local 3D attention (neighborhood attention), and a novel positional encoding method called StereoRoPE, based on stereographic projection. This allows it to incorporate geopotential and solar radiation as forcing. To suppress artifacts linked to the high information entropy of convective fields, a pixel-wise anti-aliasing decoder is used. In essence, the model learns to approximate the Navier-Stokes equations and moist air thermodynamics (including the ideal gas law) purely from data.
Results
STRATA produces stable 24-hour global forecasts with realistic kilometer-scale dynamics: tropical cyclones, orographic precipitation, island convection, and frontal systems. Quantitative metrics back up the quality—the Fractions Skill Score for precipitation handily beats persistence, and precipitation power spectra closely match the SCREAM benchmark. Energy efficiency reaches 48 model days per MWh, roughly 50 times better than SCREAM, while throughput on 512 H100 GPUs hits 741 model days per wall-clock day. Scaling analysis reveals that emulating convection demands about 10× more FLOPs per grid cell than coarse-resolution weather models, a consequence of the higher entropy in convective data.
Implications
This work marks a breakthrough in numerical modeling of the atmosphere, making global storm-resolving modeling practically accessible. It opens the door to century-scale ensemble calculations for assessing shifts in extreme precipitation frequency and the role of cloud feedbacks. Moreover, the drastic drop in energy consumption (equivalent to cutting CO₂ emissions) aligns with sustainable computing goals.
Future development
Next steps include hybridizing with a coarse global model to improve large-scale circulation, and gradually pushing resolution toward 100–200 m—the threshold needed to directly resolve low clouds and feedbacks tied to humidity and solar radiation. Such systems will offer fresh insights into climate sensitivity.
Impact
These findings will directly influence climatology, meteorology, and energy-efficient computing, spurring a shift from classic models to hybrid AI systems.
Next steps
Immediate plans involve training on longer, more diverse data samples and devising methods to correct large-scale biases, perhaps via conditional generative approaches.
Key open problems
This research is intimately tied to a central problem in climate physics—quantifying cloud feedbacks, the largest source of uncertainty in warming forecasts. It also touches on the fundamental question of extreme weather predictability in a changing climate.
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The paper introduces STRATA, an AI emulator of the global atmosphere with kilometer-scale resolution, trained on short local data patches. The model generates stable 24-hour forecasts, reproducing tropical cyclones, orographic precipitation, and frontal systems, while achieving a 50-fold gain in energy efficiency compared to the traditional SCREAM model.
Where its strength is
The key strength is the tiling strategy for training and architectural choices (local 3D attention, StereoRoPE, pixel de-aliasing decoder) that efficiently emulate high-entropy convective dynamics at significantly lower computational cost.
Where it could go next
- It might be interesting to see how STRATA's approach to emulating the global atmosphere could be adapted for modeling the atmospheres of other planets, like Mars, where data are even sparser but physical principles are similar.based on: Martian Weather, Earthly Playbook arXiv:2608.05054v1
- Related work on using diffusion models to improve precipitation in climate models is nearby; perhaps its techniques could further enhance precipitation reproduction in STRATA, especially at kilometer scales.based on: Cloud Spices: How a Neural Network Improves Climate arXiv:2510.05963
- Interpretation methods for internal representations of weather AI, discussed in a neighboring paper, could shed light on how STRATA encodes convective processes and help further optimize the architecture.based on: Peeking Inside the Mind of a Weather AI arXiv:2604.20467
Worth trying
- Test whether STRATA's tiling strategy can be used for training on data from other planets, such as Mars, where only limited observations are available.
- Try adding noise or a diffusion step to improve the spectral characteristics of precipitation, as in the work on cloud spices.
- Apply attention analysis methods to identify key patterns the model uses for convection forecasting, to simplify the architecture without losing accuracy.
Why this may matter
If STRATA can be extended to other planets or climate scenarios, it could greatly accelerate research on atmospheric dynamics under limited data and computational resources, and improve our understanding of extreme weather events on Earth.
Nearby works in our archive
- Martian Weather, Earthly Playbooksimilarity 0.66
- Pixel Rain: How to See a Downpour Where There Was Nonesimilarity 0.65
- Peeking Inside the Mind of a Weather AIsimilarity 0.63
- Weather Forecast: Sharp Focussimilarity 0.62
- Solar Flares: Forecasting with Confidencesimilarity 0.61
- Cloud Spices: How a Neural Network Improves Climatesimilarity 0.61
The closest work in the archive sits at 0.66 by meaning — the usual level for our corpus.
Original abstract
as written by the authors arXiv:2606.31248
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.
STRATA trained on just 17 days of global data, equivalent to binging one season of a TV show, but already captures planetary convective dynamics.
The vision of a fast AI emulator for global weather echoes the 'Prediction Machine' from an Asimov story, where a supercomputer divined the future — except now it's becoming real.
Key numbers
- resolution: 4.9 km
- time step: 10 minutes
- global grid: ~25 million cells
- energy efficiency: 48 model days per MWh
- speedup vs SCREAM: ~50 times
- daily throughput: 741 model days per day (512 H100)