AI-generated editorial illustration; not a documentary photograph.
NVIDIA has published a technical tutorial dated September 21, 2026, showing developers how to incorporate recent sensor and satellite observations into AI weather pipelines built with Earth2Studio. The guide covers Score-Based Data Assimilation for regional diffusion models and HealDA for constructing a global atmospheric state from varied observations.
Regional forecasts with SDA
Score-Based Data Assimilation, or SDA, steers a diffusion model during its denoising process so that predictions remain consistent with supplied measurements. It does not require retraining the underlying model. Developers define an observation operator that connects model output to measurements such as temperature or wind speed; proxy observations, including wind-turbine power output, can also be supported.
The resulting forecast remains probabilistic. NVIDIA says uncertainty is lower near observation sites and grows farther away, where the model relies more heavily on its other inputs and simulations. Results depend on the quantity, distribution and accuracy of observations, as well as the suitability of the observation operator.
CorrDiff and StormCast examples
The tutorial applies SDA to two pretrained regional models. Earth2Studio's CorrDiff model converts 0.25-degree weather fields over Europe into 2.2-km predictions, while StormCast is initialized with HRRR data and forecasts the contiguous United States at 3-km resolution. The examples use GHCN hourly observations to update wind, temperature or related fields between scheduled analyses.
Developers can also adapt their own diffusion models through PhysicsNeMo utilities. The demonstrated approach wraps a pretrained denoising model with data-consistency guidance, using masks and corresponding observed values. More advanced pipelines can provide a PyTorch model that derives simulated observations from multiple grid points or observed impacts.
Global assimilation with HealDA
For global workflows, HealDA maps remote-sensing and in-situ observations collected within a time window onto a gridded atmospheric state. Its observation encoder combines each value with metadata such as location and time before a vision transformer processes the aggregated tokens. NVIDIA says the technique can estimate global atmospheric conditions in seconds.
Earth2Studio includes a pretrained HealDA model for research use. It accepts sources including microwave sounders, radio occultation, surface stations, aircraft and buoys, producing output on a 1-degree HEALPix grid. Proprietary satellite inputs may require adapting the encoder.
Requirements and limits
Following the tutorial requires Python knowledge, an Earth2Studio development environment and an NVIDIA RTX PRO or data center GPU; NVIDIA estimates about 30 minutes for the exercises. Earth2Studio exposes a common interface for supported radar, satellite, conventional observation and forecast datasets, but developers must still select appropriate measurements, model domains and assimilation settings for their application.
Source: Turn Your Latest Observations Into Timely Weather Decisions With NVIDIA Earth-2
SIGNAL TITANS
Leave a Reply