nvidia/interp-dit-crps-era5

Model Overview

Description

interp-dit-crps-era5 is a global temporal interpolation model for weather forecasts. It takes the coarse output of a base prognostic model (typically 6-hourly frames) and fills in the frames between them, producing trajectories at higher temporal resolution for up to 73 ERA5 variables on the 0.25° global grid.

The model is endpoint-pinned. For a coarse interval between two consecutive base frames, the output at relative position tau in [0, 1] is out(tau) = (1 - tau) * x0 + tau * xT + sin(pi * tau) * f(x0, xT, cond, z): a linear blend of the two endpoint states plus a learned correction. The sin(pi * tau) envelope vanishes at both ends, so the model generates only the interior of the interval, and the frames at the endpoints are the base model's own frames. It is also probabilistic: trained with a Continuous Ranked Probability Score (CRPS)-based objective, it can produce an ensemble by sampling multiple latent-noise inputs.

The output cadence is chosen at inference. num_interp_steps sub-steps divide each interval, so a 6 h interval yields 1 h output at 6 sub-steps, 15 min at 24 and 10 min at 36. The model was trained across interval widths of 3–10 h, which are supplied to the network as a conditioning input, so 6 h is a common default rather than the only supported gap. A regional sub-domain mode restricts inference to a bounding box, enabling fine-cadence regional inference.

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.

Deployment Geography:

Global

Use Case:

Refining weather forecasts and reanalysis for research and applications that benefit from higher temporal resolution.

Release Date

Hugging Face [Insert MM/DD/YYYY] via https://huggingface.co/nvidia/interp-dit-crps-era5

Reference(s)

• ERA5 reanalysis (ECMWF)
• IFS Documentation, Part IV: Physical Processes (ECMWF)
• Balsamo et al. (2009), A Revised Hydrology for the ECMWF Model (HTESSEL)
• Peebles and Xie (2023), Scalable Diffusion Models with Transformers (DiT)

Codebase

• Earth2Studio
  
• PhysicsNeMo
  

Model Architecture

Architecture Type: Transformer
Network Architecture: ViT-style, patch-based 2D Transformer implemented with PhysicsNeMo DiT, using neighborhood attention and axial 2D RoPE. The model performs one-shot prediction rather than diffusion sampling.

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

Number of model parameters: 9.8 × 10^7

Input(s):

Input Type(s): Tensor (weather state on a global latitude/longitude grid) + DateTime
Input Format(s): PyTorch Tensor / NumPy array

Input Parameters:

• Six-Dimensional (6D) (batch, time, lead time, variable, latitude, longitude)
  
• DateTime (1D)
  

Other Properties Related to Input:

• The model is driven by a base prognostic model, whose consecutive coarse frames supply the two endpoints of each interval. The valid time of each generated frame is used to compute the solar zenith angle for conditioning.
  
• Input grid: 0.25° regular latitude/longitude, 720 × 1440, global, latitude descending from 90° to −89.75°. A 721-row source grid is accepted and cropped by dropping the −90° row.
  
• Input state variables (73): u10m, v10m, u100m, v100m, t2m, sp, msl, tcwv, and u, v, z, t, q at the 50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925 and 1000 hPa pressure levels.
  
• Four variables are optional. sp, u100m, v100m and tcwv were trained with channel dropout, so a base model supplying only the other 69 can drive the model by declaring the missing ones; they are then filled internally with their climatological mean and omitted from the output.
  
• Internally, the network consumes 239 channels: two 73-variable endpoint states, 87 conditioning channels (invariants, interval width and presence mask), and 6 latent-noise channels.
  
• For variable-name conventions, see the Earth2Studio CDS Lexicon.

Output(s)

Output Type(s): Tensor (interpolated weather states) with coordinate metadata
Output Format(s): PyTorch Tensor and Earth2Studio CoordSystem

Output Parameters:

• Six-Dimensional (6D) (batch, time, lead time, variable, latitude, longitude)
  

Other Properties Related to Output:

• Output grid: the 0.25° global model grid, or the configured regional sub-domain.
  
• Output cadence: interval width divided by num_interp_steps, which must divide the interval width in whole minutes.
  
• Each output frame includes its time and lead time in the coordinate metadata.
  
• Output variables: 69 to 73, depending on how many of the four optional variables the base model supplies.
  
• Frames at interval endpoints are the base model's own frames rather than model-generated fields.
  
• The batch dimension can carry an ensemble. For fixed base-model endpoints, interpolator-added spread comes from latent noise, vanishes at interval endpoints, and is redrawn for each interval.
  

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): PyTorch
Supported Hardware Microarchitecture Compatibility:

• NVIDIA Ampere
  
• NVIDIA Blackwell
  
• NVIDIA Hopper
  
• NVIDIA Lovelace
  

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: ERA5 (ECMWF)

Data Modality: Gridded geophysical (numerical reanalysis)

Data Size: Approximately 102 TB (92 TiB) uncompressed, covering the 1980-2016 training period and the 2017 validation year.

Data Collection Method by dataset

• Automatic/Sensors (assimilated observations combined with a numerical weather model)
  

Labeling Method by dataset

• Automatic/Sensors (each training example pairs two ERA5 endpoint frames with an ERA5 frame at the interior target hour)
  

Properties:

Hourly ERA5 reanalysis on the 0.25° global grid, 73 variables, for the period 1980 to 2016. Training examples use intervals of 3–10 h with hourly ERA5 interior targets.

Accumulated and time-averaged quantities, including precipitation and radiative fluxes, are excluded.

The model also uses time-invariant conditioning fields derived from ERA5 categorical surface fields using published ECMWF IFS/HTESSEL parameter tables.

Testing Dataset

Link: Not Applicable

Data Collection Method by dataset

• Not Applicable
  

Labeling Method by dataset

• Not Applicable
  

Properties:

No separate testing dataset was used. ERA5 2017 was held out for validation.

Evaluation Dataset

Link: 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:

Evaluation uses ERA5 years after 2017, disjoint from the data used for training and validation.

Inference:

Acceleration Engine: PyTorch
Test Hardware:

• NVIDIA Hopper (H100)
  
• NVIDIA Ada 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.