# NVIDIA STRATA

> Source: https://aiwiki.ai/wiki/nvidia_strata
> Updated: 2026-11-04
> Categories: AI Models, AI Research, AI for Science, NVIDIA
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "NVIDIA STRATA." aiwiki.ai, 4 Nov 2026. https://aiwiki.ai/wiki/nvidia_strata
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**NVIDIA STRATA** is a neural-network emulator of the Simple Cloud-Resolving E3SM Atmosphere Model, or SCREAM. Its name expands to *Storm-resolving Tile-based autoRegressive Atmosphere Transformer Architecture*. Researchers from [NVIDIA](https://aiwiki.ai/wiki/nvidia) and U.S. national laboratories introduced it in the June 2026 paper *Scaling Storm-Resolving Atmospheric AI Simulation to the Entire Planet*. It learns short atmospheric updates from numerical simulations and applies them across a global grid.[1]

The published experiments concern emulation of SCREAM's dynamics at kilometer scale. Their reference is a physics-model simulation, rather than observations of the weather. The authors report stable 24-hour global rollouts while also documenting growing large-scale errors and precipitation bias.[1]

## Model and data

| Property | Configuration reported in the paper |
| --- | --- |
| Production model parameters | 331.7 million |
| Horizontal output spacing | Approximately 4.9 km |
| Global horizontal grid | Six cubed-sphere faces, each 2,048 x 2,048 cells |
| Vertical representation | 24 selected levels, from the surface to approximately 100 hPa |
| Time step | 10 minutes |
| Training data | 17 days selected from three SCREAM simulations |
| Evaluated global rollout duration | 24 hours |

These specifications describe the research configuration. SCREAM's effective dynamics grid is about 3.3 km, while the physics output on which STRATA trains uses the 4.9-km grid. The emulator also subsamples SCREAM's original 128-level atmosphere.[1]

The model updates potential temperature, horizontal wind, vertical pressure velocity, geopotential height, specific humidity and near-surface temperature. Diagnostic outputs include surface pressure and precipitation. Additional inputs supply solar illumination, location, surface geopotential, topography and land fraction.[1]

The training periods contain nine days from one October simulation and four days each from August and January simulations, after excluding initial spin-up. The paper acknowledges an error in the October dataset: atmospheric initial conditions came from October 1994, while prescribed sea-surface temperatures came from October 2020. The authors retained the simulation because the same SCREAM solver generated its state transitions. This remains a disclosed property of the experimental data.[1]

## Tile-based prediction

STRATA learns from local spatial tiles instead of processing the entire atmosphere in every training sample. On the model's 10-minute time scale, the method assumes that the relevant interactions are predominantly local. This produces many spatial training examples from a limited number of global snapshots.[1]

During a global rollout, the system divides each cube face into overlapping tiles, predicts their updates, and blends the overlaps into a new global state. Tiles near face boundaries receive geometrically consistent neighboring information through Duo-Grid halo padding. The same local model can therefore run on different tiles and distribute work across multiple GPUs.[1]

## Architecture and training

The backbone adapts a Diffusion Transformer architecture for deterministic atmospheric prediction. It does not use iterative diffusion sampling to generate each atmospheric update. Its three-dimensional patch embedding keeps the vertical dimension as a spatial axis, while local neighborhood attention exchanges information within nearby horizontal and vertical cells. Column-wise attention separately represents interactions along atmospheric columns.[1]

Stereographic Rotary Position Embedding, or StereoRoPE, supplies position information in a tile-centered geographic frame. The implementation also rotates wind vectors into that frame. This avoids treating the same grid-index offset as the same physical direction everywhere on the sphere.[1]

A second, pixel-space decoder reduces the repeated patch-boundary patterns that can accumulate during autoregressive prediction. The production architecture uses a 24-layer backbone with an embedding width of 1,024 and a four-layer decoder with a width of 128.[1]

Training first minimizes a normalized single-step prediction loss. Subsequent fine-tuning rolls the model forward for up to four steps and evaluates its final prediction. The authors report that this step reduces instabilities during longer rollouts. The production training procedure begins with 64 x 64 tiles and later uses 128 x 128 tiles.[1]

## Reported evaluations

The study evaluates six held-out initialization times. Its tests examine rainfall structure, error growth and precipitation distributions, alongside visual examples of different atmospheric regimes. A persistence reference holds the initial rainfall field constant; it is not a competitive operational forecasting system.[1]

STRATA maintains organized rainfall better than that persistence reference. The authors also report qualitatively similar humidity and precipitation structures in three-hour regional rollouts on two grids absent from training: latitude-longitude and rotated stereographic grids. The fields do not agree exactly, and the comparison covers a particular region and horizon.[1]

### Throughput and energy accounting

The following values reproduce the paper's comparison. Throughput depends on the amount and type of hardware allocated to a run.[1]

| System | Hardware allocation | Simulated days per wall-clock day | Simulated days per MWh |
| --- | --- | ---: | ---: |
| STRATA | 16 H100 GPUs, two DGX H100 systems | 23.6 | 48.1 |
| STRATA | 512 H100 GPUs, 64 DGX H100 systems | 741 | 47.3 |
| SCREAM reference | Frontier supercomputer run | 460 | 0.78 |

The energy calculation assumes full rated system power: 20.4 kW and 652.8 kW for the two STRATA configurations and 24,607 kW for Frontier. These are the authors' accounting assumptions, rather than directly measured electricity consumption for every run. The SCREAM and STRATA rows use different hardware allocations; they should not be read as an equal-hardware speed comparison.[1]

## Code, data and checkpoint availability

The public `NVlabs/STRATA` repository supplies SCREAM-specific data preparation, geometry, training configurations and rollout tools. The model architecture resides in NVIDIA PhysicsNeMo under `physicsnemo.experimental.models.strata`. The surrounding package retains its earlier `screamcast` name, including its imports and `SCREAM_*` environment variables.[2][3]

The maintainers recommend an NGC PyTorch container with compatible CUDA, NVIDIA DALI and NATTEN dependencies. A plain Python installation resolves its own compatible packages and does not reproduce that pinned container environment. Training also needs the simulation stores and auxiliary grid files.[2]

The current release documents three public Hugging Face Zarr stores: `nvidia/STRATA-SCREAM-sdecadal`, `nvidia/STRATA-SCREAM-sdy1` and `nvidia/STRATA-SCREAM-sdy2`. The first dataset card identifies the E3SM project as the data owner and describes the release as a reformatted version of publicly available SCREAM simulation output.[2][4]

Code and data availability should be distinguished from checkpoint availability. On November 4, 2026, the README still described the STRATA v1 checkpoint bundle as private, with a public replacement planned. Its example forecast command therefore did not establish that every reader could download the required pretrained weights.[2]

The STRATA code uses Apache License 2.0. The SCREAM dataset card declares Creative Commons Attribution 4.0. External software and other downloaded components retain their applicable terms.[2][4][5]

## Limits of the results

The paper identifies excessive error growth in large-scale circulation over 12-24 hours and global mean rainfall that can be about 20% too low. Its comparison with the coarse-resolution ACE2 atmosphere emulator shows that STRATA does not match that system's slower large-scale error growth.[1]

The authors propose coupling the local model to a coarse global emulator as future work. Their 24-hour tests do not establish reliable century-scale climate simulation, operational weather forecasting or general accuracy under changing climate conditions. STRATA is an [AI for Science](https://aiwiki.ai/wiki/ai_for_science) research system for a specified simulation problem. Its reported accuracy measures agreement with its SCREAM reference.[1]

## References

1. Hu, Zeyuan, et al. [Scaling Storm-Resolving Atmospheric AI Simulation to the Entire Planet](https://arxiv.org/html/2606.31248v1). arXiv:2606.31248v1, June 30, 2026. Methods, evaluation and Appendices B, C and G.
2. NVIDIA. [STRATA repository and README](https://github.com/NVlabs/STRATA). Code, environment, data and checkpoint availability. Accessed November 4, 2026.
3. NVIDIA. [STRATA model implementation in PhysicsNeMo](https://github.com/NVIDIA/physicsnemo/tree/main/physicsnemo/experimental/models/strata). Accessed November 4, 2026.
4. NVIDIA and the E3SM Project. [STRATA-SCREAM-sdecadal dataset card](https://huggingface.co/datasets/nvidia/STRATA-SCREAM-sdecadal). Data provenance, format and license. Accessed November 4, 2026.
5. NVIDIA. [STRATA software license](https://github.com/NVlabs/STRATA/blob/main/LICENSE.txt). Apache License 2.0.