In an era where artificial intelligence is sweeping across industries, NVIDIA has once again proven that AI can not only generate text or images, but also predict the future of the planet.
At the American Meteorological Society (AMS) annual meeting held in Houston, NVIDIA officially announced that it would...Earth-2 digital twin platformThe transformation into a fully open-source family of models is not just about releasing a few model weights, but about providing a complete toolchain from data processing and model training to inference. The goal is to enable meteorological units, research institutions, and even startups around the world to build "sovereign AI" meteorological systems on their own infrastructure.
Why does weather forecasting need an AI revolution?
Traditional weather forecasting relies heavily on massive supercomputers to calculate physical equations, which is not only time-consuming but also extremely costly. Mike Pritchard, Director of Climate Simulation Research at NVIDIA, points out that as extreme weather events intensify, the stakes in weather forecasting are becoming increasingly high. The intervention of AI could bring about 100 times, or even 1000 times, faster forecasting speeds and improved energy efficiency.
However, many past AI weather models were closed systems, making it difficult for developers to fine-tune the details. NVIDIA's Earth-2 open-source project includes three groundbreaking new models that cover the entire process from data assimilation and global forecasting to regional severe weather warnings.
Earth-2 adds three new core models
• Earth-2 Medium Range (Mid-Range Forecast):Employing a new architecture called "Atlas," this model can provide global weather forecasts up to 15 days in advance. In standard tests, its prediction accuracy for more than 70 weather variables has surpassed that of the current leading open-source model, GenCast. This means that developers can achieve accuracy comparable to or even exceeding that of traditional numerical models with less computing power.
• Earth-2 Nowcasting:This is, in my opinion, the most interesting application. Based on the "StormScope" architecture, it uses generative AI technology to learn storm dynamics directly from satellite cloud images. Unlike traditional physical models that require long computation times, Earth-2 Nowcasting can generate high-resolution lightning cell predictions at the kilometer level for the next 0 to 6 hours within minutes. This provides invaluable real-time information for disaster response, airport scheduling, and energy management.
• Earth-2 Global Data Assimilation:Before making forecasts, it is essential to understand the current state of the Earth. Traditional data assimilation is extremely computationally intensive, consuming approximately 30% of a supercomputer's load. NVIDIA's new architecture, HealDA, can generate global initial conditions on GPUs in seconds, rather than the traditional hours, thus solving one of the biggest bottlenecks in the weather forecasting process.
Taiwan's Central Weather Administration (CWA) has implemented the application.
It is worth noting that NVIDIA specifically mentioned [the topic] during the presentation.Taiwan Ministry of Transportation and Communications Central Meteorological Administration (Central Weather Administration, CWA).
According to NVIDIA, Taiwan's Central Weather Bureau used the CorrDiff model (a technique that uses diffusion models for downscaling) from the Earth-2 family to generate more accurate typhoon landfall forecasts. Given that hundreds of typhoons have struck Taiwan over the past two decades, using AI for high-resolution disaster impact assessments is extremely important for resource allocation and disaster preparedness.
Interview: Generative AI and the Future of "Mini-Earth"
In a conversation with Mike Pritchard, Director of Climate Simulation Research at NVIDIA, I also asked questions about several key technical details:
Q: How frequently is Earth-2's data updated? Is it able to respond in real time?
Mike Pritchard:This depends on the model type. For medium-range forecasts, we predict physical state variables (such as temperature and wind speed), which usually requires conversion from raw observation data, resulting in a certain delay. However, our new global data assimilation model (HealDA) reduces this process from hours to seconds.
As for nowcasting, because it learns directly from observational data (such as geostationary satellite images), it reacts extremely quickly and can complete the prediction within minutes of the observational data arriving, making it very suitable for warnings of very short-term and drastic weather changes.
Q: What role does generative AI play here? How can the accuracy of the data be ensured?
Mike Pritchard:Generative AI is at the heart of Earth-2 Nowcasting (StormScope) and CorrDiff. Traditional models become blurry when resolution is insufficient, while generative AI excels at "super-resolution," capable of restoring coarse signals to fine textures (such as the structure of a typhoon eyewall or localized torrential rain). By learning from vast amounts of physical data and satellite imagery, AI can fill in microscopic details that are difficult to simulate using traditional computation, while adhering to physical properties.
Q: Can Earth-2 be used as a "mini-Earth" for games or environmental simulations?
Mike Pritchard:Absolutely. In the past, this level of Earth atmospheric simulation could only run on national supercomputers and was mainly used for scientific research. But with Earth-2 lowering the computational barrier and opening up its tools, this simulation can be used for educational visualization and even integrated into game engines to create dynamic, realistic, and interactive weather environments, which was unimaginable in the past.
Analysis: The Era of "Sovereignty" in Weather Forecasting
Meteorological data often involves national security and localization needs (for example, Taiwan needs to focus on typhoons and plum rains, while Israel focuses on desert climates). By providing open tools like Earth2Studio and the PhysicsNeMo framework, NVIDIA is not only selling hardware, but also building a meteorological AI ecosystem centered on NVIDIA GPUs.
For countries like Taiwan that are deeply affected by extreme weather, being able to fine-tune state-of-the-art open-source models by combining them with local observational data will be an important step in enhancing disaster resilience.
