As winter storms swept across much of the United States, traditional weather forecasts struggled to cope with the huge differences in snowfall. At this critical moment, NVIDIA officially launched its new Earth-2 weather forecasting model suite on Monday at the American Meteorological Society meeting in Houston, aiming to redefine the accuracy and efficiency of global weather forecasting through artificial intelligence technology.

Cars stopped by the roadside in the snow

Core Breakthrough: Performance Exceeds Google, Architecture Returns to Simplicity

NVIDIA's core model, Earth-2 Medium Range (Medium-range Forecast Model), caused a big reaction in the industry. According to official data from NVIDIA, the model outperforms Google DeepMind's GenCast model, released in December 2024, on more than 70 meteorological variables.

Differing from traditional models that rely on complex manual physical simulations, Earth-2 uses a new architecture based on Atlas. Mike Pritchard, director of NVIDIA's climate simulation, said this marks a shift in meteorology science toward "returning to simplicity," by discarding niche AI architectures and moving toward more scalable Transformer architectures.

The Three Pillars of the Earth-2 Suite

Aside from medium-range forecasting, NVIDIA also announced specialized tools for different scenarios, forming a complete meteorological AI ecosystem:

  • Nowcasting Model: Focuses on short-term forecasts for the next 0-6 hours. The model is trained directly using global geostationary satellite data, rather than being limited to specific regional physical models, allowing it to better capture the impact of storms and dangerous weather.

  • Global Data Assimilation Model: This model integrates multi-source data such as weather stations and weather balloons to provide an initial snapshot for predictions. Its biggest breakthrough is efficiency—tasks that previously required supercomputers to take hours and occupied 50% of computing resources can now be completed in just minutes using GPU.

  • High-resolution and variable modeling: The suite also includes CorrDiff (for generating rapid high-resolution forecasts) and FourCastNet3 (for single-variable modeling such as temperature, wind, and humidity).

Meteorological Sovereignty and Democratization: Making Super Forecasts Affordable

Pritchard pointed out that high-quality weather forecasts were once a "privilege" of wealthy countries and large corporations because traditional forecasting required paying extremely expensive supercomputer rental fees.

"Weather is a matter of national security, and sovereignty is inseparable from weather," Pritchard emphasized. Earth-2's high performance has lowered the barrier, enabling developing countries and small institutions to have their own accurate forecasting systems.

Currently, the relevant technologies of Earth-2 have been put into practice. The meteorological departments of Israel and Taiwan have started using CorrDiff; while The Weather Company (parent company of Weather Channel) and Total Energies are evaluating the practical effects of Nowcasting