Brightband AI Weather Data

Brightband is a Public Benefit Corporation with a mission to provide AI Weather forecasting tools to all. As part of our commitment to accelerating innovation in AI Weather, Brightband builds AI-ready observational datasets to power a new generation of machine learning-based weather prediction tools. Let’s not lose important weather data. 💦

NNJA-AI 🥷

Brightband and others are excited about forecasting using AI from weather observations, rather than forecasting from traditionally-produced analysis like ERA5. To enable innovation in AI forecasting from observations we have produced a free open dataset that we hope will be broadly useful, as ERA5 has been for AI forecasting from analysis.

Under Brightband’s Cooperative Research and Development Agreement with NOAA, we’ve built an AI Weather Observations dataset - NNJA-AI, which is an AI-Ready, Cloud Optimized mirror of the NOAA NASA Joint Archive (NNJA) of Observations for Earth System Reanalysis. The first science-ready version of the NNJA-AI dataset was released in the Fall of 2025 with support from the NOAA Open Data Dissemination program, and improvements and software integrations are planned for 2026.

NNJA-AI contains the complete, re-processed record of over a dozen sensors onboard a mixture of geostationary and low-earth orbit satellites as well as conventional data from surface stations and radiosondes, from 2021-2024.

In 2026 we are working on:

• extending NNJA-AI with data from a variety of additional sensing systems, including aircraft, drones, radio occultation, and more…
• preparing direct integrations with AI modeling frameworks such as ECMWF’s anemoi, to support R&D across the weather modeling community
• expanding NNJA-AI with recent observations data, to continually update within a few days of real-time ⏳

NNJA-AI data is freely available under a permissive CC-BY 4.0 license, find it at: https://www.brightband.com/data/nnja-ai/

MLWP Forecast Datasets 🤖

There is an expanding model zoo of Machine Learning Weather Prediction (MLWP) or AI Weather models from different research groups, all slightly different. We love to operationalize AI weather models to make them useful, so Brightband runs them for you, and shares the forecasts, so that you don’t have to.

We are releasing an Archive for 2021-2024 of GraphCast, Panguweather, and AIFS-Single (all initialized from IFS HRES). The forecast output is available through Arraylake on the new Earthmover Data Marketplace and will initially contain the necessary fields to reproduce our ExtremeWeatherBench benchmark results; the complete 280 TB archive is also available on the Google Cloud Platform in ARCO format (Zarr+Icechunk).

Brightband runs Aurora, GraphCast, Pangu-Weather, NVIDIA Earth2 Medium-Range/FourCastNet-v3, AIFS-Single and AIFS-ENS, and will continue to add models, as well as include them in the archive.

We are also supporting the launch of the Earthmover Data Marketplace with some other data sets:

• An extremely low-latency, continuously updating archive of fields from the ECMWF IFS HRES 15-day forecast, re-processed to cloud-optimized format.
• Continuously updating analysis fields from both the IFS HRES and IFS Ensemble forecasts curated for running AI weather models, in cloud-optimized format.
• More coming 🔜…

Let us know if you have an AI Weather data request 🤔