NVIDIA has unveiled cBottle, short for Climate in a Bottle, which it claims is the world´s first generative foundation model designed to simulate the Earth’s climate at kilometer-scale resolution. Integrated into the NVIDIA Earth-2 platform, cBottle brings an unprecedented level of granularity and speed to climate modeling, giving scientists and researchers a powerful tool for predicting and mitigating the effects of climate change. This model is capable of generating realistic atmospheric states based on inputs such as time of day, season, and sea surface temperatures, thus providing deep insights into complex climate systems.
Earth-2 leverages a sophisticated suite of software, tools, and NVIDIA GPUs to accelerate the development and visualization of climate digital twins—interactive, high-fidelity simulations of Earth’s weather systems. By using cBottle, climate predictions can now be produced thousands of times faster and more efficiently than those built on traditional computation-heavy numerical methods, all while maintaining scientific accuracy. Institutions like the Max-Planck-Institute for Meteorology and the Allen Institute for AI are working with cBottle to compress massive volumes of Earth observation data, delivering interactive querying and visualization capabilities for ultra-high-resolution climate simulations. The model’s field test at the World Climate Research Programme Global KM-Scale Hackathon highlighted its capability to enhance both data access and analysis across international research centers.
At its core, cBottle dramatically reduces computational and data storage burdens by using advanced Artificial Intelligence to compress petabytes of simulation data up to 3,000-fold per weather sample. Trained on high-resolution physical simulations and decades´ worth of observed atmospheric data, cBottle can fill in missing climate observations, correct model biases, super-resolve coarse data, and synthesize new scenarios, often needing only four weeks of training data from kilometer-scale climate simulations. This efficiency paves the way for wider accessibility and broader adoption of high-resolution climate modeling.
NVIDIA Earth-2 and cBottle have been adopted by leaders in climate research to push the boundaries of Earth system modeling. The Max-Planck-Institute has used Earth-2 for groundbreaking, full Earth system simulations at kilometer-scale details, while the Allen Institute for AI collaborates with NVIDIA to accelerate and democratize access to sophisticated, high-resolution climate forecasting. By empowering scientists, policymakers, and developers with digital twins and scalable, fast-simulating models, cBottle offers a path toward more precise weather predictions and a deeper understanding of local extreme weather events. The foundation model is already available for early access, with code and research preprints offered publicly on GitHub and arXiv for climate researchers seeking to extend or fine-tune its capabilities.
