nvidia/healda-v2

Earth-2 Checkpoints: HealDA v2

Description

HealDA v2 is a global machine-learning data-assimilation model that maps a 48-hour sequence of satellite and conventional observations to a 104-channel, 0.25° atmospheric analysis on a regular latitude–longitude grid. The analysis can serve as an initial condition for off-the-shelf machine-learning weather forecast models.

Unlike HealDA v1, which processes one analysis time independently, HealDA v2 uses causal video data assimilation across eight consecutive 6-hourly analysis times. Each frame uses observations within ±3 hours of its analysis time, so the complete sequence covers 48 hours from the earliest to the latest observation time. Temporal attention allows later analyses to use information from earlier frames without using future model states. The public inference API returns only the final analysis frame.

This checkpoint predicts 104 atmospheric, ocean-surface, cloud, and land-surface fields. It was trained with source-level NNJA satellite and conventional observations, including microwave and infrared sounders, GNSS radio occultation, satellite-derived winds, and in-situ observations.

For training recipes see HealDA, for inference see NVIDIA Earth2Studio.

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 the LICENSE.OpenMDW-1.1.md file distributed with the model for the governing terms.

Additional Information: Apache License, Version 2.0.

Deployment Geography

Global

Use Case

Industry, academic, and government research teams interested in global data assimilation, observing-system studies, and initialization of medium-range machine-learning weather forecasts. Model analyses are statistical estimates and should not be the sole basis for safety-critical decisions.

Release Date

Hugging Face 10/01/2026 via nvidia/healda-v2

Reference(s)

Papers:

• HealDA: Highlighting the importance of initial errors in end-to-end AI weather forecasts

Code:

• PhysicsNeMo
• Earth2Studio

Model Architecture

Architecture Type: Custom observation encoder + HEALPix Diffusion Transformer (DiT) backbone + mesh-to-grid decoder, used as a deterministic regression model.

Network Architecture: Sensor-specific point-cloud embedders encode irregular observations. Observation-to-pixel cross-attention places those features into an internal HEALPix latent representation. A DiT backbone applies spatial self-attention and causal temporal attention across eight analysis frames. A GraphCast-style mesh-to-grid decoder maps the latent representation directly to the native 0.25° regular latitude–longitude output grid.

This model was developed based on: HealDA v1 and the Earth2Studio / PhysicsNeMo HEALPix transformer stack.

Number of model parameters: 917.42 million (~0.92 billion).

Input(s)

Input Type(s):

• Tensor (satellite and conventional observations represented as point-cloud data)
• Tensor (eight ERA5-derived static conditioning channels and geographic features)
• Tensor (day of year)
• Tensor (second of day)

Input Format(s): PyTorch tensors

Input Parameters:

• Observations: variable-length point clouds organized into eight 6-hourly frames
• Static conditioning: eight global 0.25° ERA5-derived channels on a 721 × 1440 latitude–longitude grid and their internal HEALPix representation
• Day of year: two-dimensional tensor (batch, time)
• Second of day: two-dimensional tensor (batch, time)
• Sequence length: eight analysis frames

Other Properties Related to Input:

• Observation context: ±3 hours around each frame. Eight frames separated by six hours provide a 48-hour span from the earliest to latest observation time.
• Microwave sounders: ATMS, AMSU-A, AMSU-B, and MHS.
• Infrared sounders: AIRS, IASI, and CrIS.
• Conventional observations: surface stations, aircraft, radiosondes, and buoys, including pressure, temperature, humidity, and wind measurements.
• GNSS radio occultation: bending-angle observations.
• Satellite-derived winds: atmospheric motion vectors and ASCAT scatterometer winds.
• Time-invariant conditioning uses eight channels from ERA5 invariant datasets: orography (orog), land-sea mask (lsm), standard deviation of subgrid orography (sdor), subgrid-orography slope (slor), subgrid-orography anisotropy (isor), standard deviation of filtered subgrid orography (sdfor), and the sine and cosine of subgrid-orography orientation (anor_sin, anor_cos). Time-varying conditioning includes calendar and solar-zenith features.

Output(s)

Output Type(s): Tensor (104-channel global atmospheric, ocean-surface, cloud, and land-surface analysis)

Output Format(s): PyTorch tensor

Output Parameters: The public inference API returns the final analysis on a 721 × 1440 latitude–longitude grid. The tensor has shape equivalent to (batch, 104, latitude=721, longitude=1440); interfaces that retain a time dimension expose it with length one.

Other Properties Related to Output:

• Output grid: global regular latitude–longitude grid at 0.25° resolution, represented by 721 latitude points and 1,440 longitude points.
• Pressure levels: 1000, 925, 850, 700, 600, 500, 400, 300, 250, 200, 150, 100, and 50 hPa.
• Three-dimensional fields on all 13 pressure levels: zonal wind (U), meridional wind (V), temperature (T), geopotential (Z), specific humidity (Q), and vertical velocity (W).
• Additional 10 hPa fields: U, V, T, and Z. The checkpoint does not predict Q or W at 10 hPa.
• Surface and column fields: total-column water vapour (tcwv), 2 m temperature (tas), 10 m wind (uas, vas), 100 m wind (100u, 100v), mean sea-level pressure (pres_msl), sea-surface temperature (sst), sea-ice concentration (sic), surface pressure (sp), 2 m dewpoint (d2m), skin temperature (skt), and total-column water (tcw).
• Cloud fields: total, low, medium, and high cloud cover (tcc, lcc, mcc, hcc).
• Land fields: soil temperature in layers 1 and 2 (stl1, stl2), volumetric soil water in layers 1 and 2 (swvl1, swvl2), and snow depth (sd).

Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA hardware and CUDA software libraries, the model achieves faster training and inference than CPU-only execution.

Software Integration

Runtime Engine(s): PyTorch / PhysicsNeMo

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: HealDA v2, 104-channel, 0.25° checkpoint.

Training, Testing, and Evaluation Datasets

Training Dataset

Data Modality:

• Other: numerical geophysical point observations and gridded reanalysis fields

Training Data Size:

• 35,068 nominal six-hourly ERA5 target times across 2000–2021 and 2023–2024, before availability filtering, paired with a variable number of NNJA point observations

Link: ERA5

Data Collection Method by dataset:

• Automatic/Sensors (observations combined with numerical-model reanalysis)

Labeling Method by dataset:

• Automatic/Sensors

Properties:

Six-hourly ERA5 reanalysis fields on the native 0.25° regular latitude–longitude grid are used as supervised targets. Training uses 2000–2021 and 2023–2024. Calendar year 2022 is held out for validation, and 2025 onward is reserved for out-of-sample testing.

ERA5 supplies all gridded atmospheric targets and all static conditioning fields. NNJA observations are used only as observation inputs to the data-assimilation model.

Per-channel normalization is computed for the 0.25° grid. Area-weighted loss and metrics account for the unequal cell areas of a regular latitude–longitude grid. Domain masks prevent fields such as sea-surface temperature, sea-ice concentration, and soil moisture from being scored outside their valid physical domains.

License: Creative Commons Attribution 4.0 International (CC BY 4.0).

Attribution: Modified from ECMWF Reanalysis data (ERA5), © European Centre for Medium-Range Weather Forecasts (ECMWF), distributed through the NSF NCAR ERA5 mirror.

Source: NOAA–NASA Joint Archive (NNJA) of Observations for Earth System Reanalysis, distributed through the public s3://noaa-reanalyses-pds/observations/reanalysis/ archive.

License: Creative Commons Attribution 4.0 International (CC BY 4.0).

Attribution: NOAA NASA Joint Archive (NNJA), https://www.psl.noaa.gov/data/nnja_obs/.

Additional acknowledgement for NNJA amv/merged: EUMETSAT (2021), Atmospheric Motion Vectors Climate Data Record Release 2 — MFG and MSG — 0 degree, European Organisation for the Exploitation of Meteorological Satellites, doi:10.15770/EUM_SEC_CLM_0020.

Data Collection Method by dataset:

• Automatic/Sensors

Labeling Method by dataset:

• Automatic/Sensors

Properties:

Satellite observations include microwave and infrared sounders. Conventional observations include PrepBUFR surface, aircraft, radiosonde, and buoy reports, GNSS radio occultation, and dedicated satellite-wind products. The observing system changes over time as instruments enter and leave service. Observation inputs follow the same training, validation, and out-of-sample test periods as the corresponding ERA5 targets. NNJA is an experimental research dataset under active development, and archived content may change as quality-control updates are applied.

Data Processing Description

ERA5 fields are sampled at six-hour intervals on the native 0.25° grid. Observations are quality controlled, normalized by channel (and by pressure level for conventional observations), spatially thinned, and encoded as irregular point clouds. Each training example contains eight sequential analysis frames with causal temporal attention.

Testing Dataset

Links: ERA5 and the corresponding held-out NNJA observations.

Data Collection Method by dataset:

• Automatic/Sensors

Labeling Method by dataset:

• Automatic/Sensors

Properties:

Data from 2025 onward is reserved for out-of-sample testing and is not used for optimization or model selection. Test coverage extends through the latest available matched ERA5 and NNJA observations.

Evaluation Dataset

Links: ERA5 and the corresponding held-out NNJA observations.

Benchmark Score: Not reported.

Data Collection Method by dataset:

• Automatic/Sensors

Labeling Method by dataset:

• Automatic/Sensors

Properties:

Validation uses calendar year 2022. Analysis error is evaluated against ERA5 on the native 0.25° grid using latitude-area weighting. Downstream forecast comparisons use matched forecast models, initialization dates, grids, regridding methods, and verification targets.

Inference

Acceleration Engine: PyTorch / PhysicsNeMo

Test Hardware:

• NVIDIA Hopper (H100)
• NVIDIA Blackwell (B300 / GB300)

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.