nvidia/corrdiff-cosmo-era5

Model Overview

Description

corrdiff-cosmo-era5 is a downscaling model in NVIDIA's CorrDiff family. It performs spatial downscaling of global ERA5 reanalysis data to high-resolution regional reanalysis over Europe, reproducing the COSMO-REA regional reanalyses produced by the University of Bonn and the German Meteorological Service (DWD). The model takes coarse-resolution ERA5 atmospheric fields (0.25°) and predicts high-resolution surface, near-surface, and lower-atmosphere variables on the native rotated-pole COSMO grid.

Two resolutions are provided, each a self-contained model in the same package:

• REA6: downscales ERA5 to the COSMO-REA6 6 km grid (824 × 848), predicting 45 variables (surface, radiative/turbulent fluxes, and the lowest six model levels of wind, temperature, humidity, and turbulent kinetic energy, plus the planetary boundary-layer height).
• REA2: downscales ERA5 to the COSMO-REA2 2.2 km grid (780 × 724), predicting 22 variables (surface fields plus the lowest six model levels of the 3D wind components and the boundary-layer height).

Each resolution offers two inference modes from the same training pipeline:

• a deterministic (regression) mode that predicts the conditional mean field, and
• a generative (diffusion) ensemble mode that samples stochastic realizations, enabling uncertainty quantification and sharper fields.

The model allows the prediction of high-resolution regional reanalysis fields from coarse global reanalysis that would otherwise require an expensive limited-area numerical weather prediction simulation.

This model is ready for commercial or non-commercial use.

License/Terms of Use:

Governing Terms: Use of this model is governed by the Linux Foundation OpenMDW License Agreement, version 1.1. See LICENSE file for governing terms. The files rea6/invariants_rea6_ext.nc and rea2/invariants_rea2_ext.nc are licensed under CC BY 4.0, with attribution obligations listed in ATTRIBUTION.md.

Deployment Geography:

Global

Use Case:

High-resolution regional climate and weather analysis for the wind- and solar-energy sector (resource assessment, hub-height wind), climate-risk and hydrology research, renewable-energy planning, and weather and climate researchers who need kilometre-scale regional reanalysis fields derived from coarse global reanalysis at a fraction of the cost of a dynamical limited-area model. Outputs are statistical estimates and are not a substitute for validated risk models; they should not be the sole basis for safety-critical decisions (see Ethical Considerations). The model is validated only in-domain on the native COSMO-REA grids; use over extended margins or other regions is unvalidated.

Release Date

Hugging Face: 07/27/2026 via https://huggingface.co/nvidia/corrdiff-cosmo-era5

Reference(s)

• COSMO-REA6 regional reanalysis (DWD)
  
• COSMO-REA2 regional reanalysis (DWD)
  
• ERA5 reanalysis (ECMWF)
  
• GLOBE v1 elevation (NOAA NGDC)
  
• ESA WorldCover 2021 v200
  
• Diffusion-based downscaling / CorrDiff
  
• Diffusion model preconditioning (EDM)
  

Codebase

• Earth2Studio
  
• PhysicsNeMo
  

Model Architecture

Architecture Type: Transformer (Diffusion Transformer, DiT)
Network Architecture: Diffusion Transformer (DiT) with axial 2D rotary position embeddings (RoPE), localized neighborhood attention (NATTEN, kernel size 23), and a patch-size-2 tokenizer. The deterministic (mean) model is a regression DiT (98M parameters) with a convolutional detokenizer head; the generative (ensemble) model is an EDM-preconditioned diffusion DiT (174M parameters) with the standard patch detokenizer, sampled with an 18-step deterministic (Heun) sampler. Both resolutions (REA6, REA2) share this architecture. Note: although part of the CorrDiff downscaling family, this model uses a Diffusion Transformer (not a U-Net), and its generative model predicts the target field directly rather than a residual correction to a deterministic prediction as in the original CorrDiff.

This model was developed based on: the Earth2Studio / PhysicsNeMo Diffusion Transformer downscaling stack.

Number of model parameters: 9.8 × 10^7 (regression / mean mode); 1.7 × 10^8 (diffusion / ensemble mode).

Input(s):

Input Type(s): Tensor (47 ERA5 surface and pressure-level variables) + grid-tied static invariants + sine/cosine of latitude and longitude + cosine of the solar zenith angle. The static-invariant set differs by resolution: REA6 uses surface elevation, land fraction, surface roughness length, terrain slope (east/north), 50 km continentality (land fraction within 50 km), and distance to coast; REA2 uses a reduced set (surface elevation, land fraction, surface roughness length, terrain slope (east/north)), without continentality or distance to coast. The static invariants are derived from open external datasets: surface elevation and terrain slope from GLOBE v1 (NOAA NGDC, public domain); land fraction, surface roughness length, continentality, and distance to coast from ESA WorldCover 2021 v200 (CC BY 4.0).

Input Format(s): PyTorch Tensor.

Input Parameters:

• Four-Dimensional (4D) (batch, variable, latitude, longitude)
  

Other Properties Related to Input:

• 0.25° regular latitude/longitude ERA5 grid covering the European domain (plus a margin around the target grid).
  
• ERA5 input variables (47): u10m, v10m, u100m, v100m, t2m, sp, msl, and u/v/z/t/q at the 300, 400, 500, 600, 700, 850, 925, and 1000 hPa pressure levels.
  
• For variable-name conventions, see the Lexicon in Earth2Studio.
  

Output(s)

Output Type(s): Tensor (high-resolution COSMO-REA variables on the native rotated-pole grid).

Output Format(s): PyTorch Tensor.

Output Parameters:

• Five-Dimensional (5D) (batch, samples, variable, latitude, longitude). The sample dimension is 1 in deterministic (mean) mode and N in ensemble (diffusion) mode.
  

Other Properties Related to Output:

• The output grids are rotated-pole: a regular latitude/longitude grid in a rotated frame centred on the domain, so spacing is near-uniform. Each cell's true latitude/longitude ships with the package (grids.nc).
  
• REA6: ~6 km rotated-pole grid (824 × 848); 45 variables: U_10M, V_10M, T_2M, PS, TOT_PRECIP, CLCT, ASWDIFD_S, ASWDIR_S, ALWU_S, ATHD_S, QV_2M, VMAX_10M, LHFL_S, SHFL_S, H_PBL, and model levels L35 to L40 of U_L*, V_L*, T_L*, TKE_L*, Q_L*.
  
• REA2: ~2.2 km rotated-pole grid (780 × 724); 22 variables: 10U, 10V, 2MT, TD_2M, TOT_PRECIP, CLCT, ASWDIFD_S, ASWDIR_S, PS, H_PBL, and model levels L45 to L50 of U3D_L*, V3D_L*.
  
• Derived outputs (optional): hub-height wind components u{H}m/v{H}m at user-chosen heights, vertically interpolated on-model from the 3D wind levels (heights outside the resolved level range are clamped to the nearest level, not extrapolated); wind speed ws{H}m is obtained by composing the standard Earth2Studio DerivedWS diagnostic. This derived output was not independently validated; it inherits the model-level wind skill (see the U_L*/V_L* / U3D_L*/V3D_L* metrics) plus interpolation error.
  
• For variable-name conventions, see the Lexicon in Earth2Studio.
  

Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA’s hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.

Software Integration

Runtime Engine(s): Not Applicable
Supported Hardware Microarchitecture Compatibility:

• NVIDIA Ampere
  
• NVIDIA Blackwell
  
• NVIDIA Hopper
  

Supported Operating System(s):

• Linux
  

The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.

Model Version(s)

Model version: v1

Training, Testing, and Evaluation Datasets:

Training Dataset

Link: COSMO-REA6 / COSMO-REA2 (DWD), ERA5 (ECMWF), GLOBE v1, ESA WorldCover 2021 v200

Data Collection Method by dataset

• Automatic/Sensors (assimilated observations + numerical model reanalysis; satellite-derived land cover and elevation for the static invariants)
  

Labeling Method by dataset

• Automatic/Sensors (ERA5 input paired with the corresponding COSMO-REA target)
  

Properties:

ERA5 (coarse input) paired with COSMO-REA6 (6 km) / COSMO-REA2 (2.2 km) regional reanalysis (high-resolution target) over the European domain. ERA5 is a global hourly reanalysis blending historical observations with a modern weather model. COSMO-REA6 and COSMO-REA2 are DWD regional reanalyses produced with the COSMO limited-area model. Training time period: 2009 to 2014 and 2017 to 2018 (validation: 2015). The model additionally consumes time-invariant static fields derived from open external datasets: surface elevation and terrain slope from GLOBE v1 (NOAA NGDC); land fraction, surface roughness, continentality, and distance-to-coast from ESA WorldCover 2021 v200.

Testing Dataset

Link: COSMO-REA (DWD) / ERA5 (ECMWF)

Data Collection Method by dataset

• Automatic/Sensors
  

Labeling Method by dataset

• Automatic/Sensors
  

Properties:

Held-out ERA5 / COSMO-REA period not used in training. Test time period: 2016.

Evaluation Dataset

Link: COSMO-REA (DWD) / ERA5 (ECMWF)

Benchmark Score: Quantitative per-variable validation scores will be published separately as Earth2Studio model scorecards.

Data Collection Method by dataset

• Automatic/Sensors
  

Labeling Method by dataset

• Automatic/Sensors
  

Properties:

Held-out in-domain ERA5 / COSMO-REA evaluation set on the native COSMO-REA grid. Evaluation time period: 2016.

Inference:

Acceleration Engine: PyTorch
Test Hardware:

• NVIDIA Ampere (A100)
  
• NVIDIA Hopper (H100)
  
• NVIDIA Lovelace (L40S)
  

Ethical Considerations:

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse. For more detailed information on ethical considerations for this model, please see the Model Card++ Bias, Explainability, Safety & Security, and Privacy Subcards.

Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here.