NVIDIA, a leader in AI computing technology, is expected to help establish climate-related data-based adaptation strategies by providing real-time, kilometer-scale climate simulation capabilities.

At GTC Paris, NVIDIA unveiled its next-generation generative AI foundation model dedicated to climate simulation, Climate in a Bottle.

A key component of the NVIDIA Earth-2 platform, cBottle realistically reproduces Earth’s atmosphere at kilometer-scale resolution, while delivering processing speeds thousands of times faster and greater energy efficiency than existing numerical models.

cBottle automatically generates atmospheric phenomena by reflecting climate input variables such as time of day, season of the year, and sea surface temperature.

It is trained using a combination of high-resolution physics-based simulations and 50 years of real-world observations, and has the ability to supplement missing or corrupted data and transform low-resolution data into super-resolution.

It compresses petabyte-level climate data by up to 3,000 times, dramatically improving storage and analysis efficiency, and reduces data preparation costs as retraining is possible with just four weeks of data.

Major scientific research institutions are using Earth-2 and cBottle to build climate digital twins and improve simulation accuracy. />
The Max Planck Institute for Meteorology (MPI-M) has performed the world’s first kilometer-scale global modeling based on ICON Earth.

“The combination of Earth-2 and AI-accelerated computing is a turning point for climate science,” said MPI-M Director Bjorn Stevens. “It will greatly accelerate the development of data-driven adaptation strategies.”

The Global KM-Scale Hackathon, hosted by the Allen Institute for Artificial Intelligence (Ai2), involved 10 centers from 8 countries to accelerate the development of high-resolution, high-fidelity climate models.

“cBottles can dramatically improve our ability to respond to local crises by efficiently simulating extreme weather events,” said Christopher Bretherton, AI2’s senior director of climate modeling.

cBottle is available on GitHub Early Access, and the paper is available on arXiv.

Developers can leverage Earth-2’s integrated stack of AI, GPU acceleration, physics simulation, and computer graphics to build interactive digital twins and implement low-latency, high-throughput climate prediction scenarios.