TL;DR

Nvidia has released Earth-2, a suite of AI weather forecasting models that the company claims beats Google DeepMind’s GenCast on more than 70 variables. The tools promise to democratise advanced weather prediction for smaller nations and organisations.

Transformer Architecture Takes on Weather

Announced at the American Meteorological Society meeting in Houston, Nvidia’s Earth-2 suite represents what the company calls “a return to simplicity” in weather forecasting AI.

“We’re moving away from hand-tailored niche AI architectures and leaning into the future of simple, scalable, transformer architectures,” said Mike Pritchard, director of climate simulation at Nvidia.

The suite includes three new models: Earth-2 Medium Range for forecasts up to 15 days out, Nowcasting for zero-to-six-hour predictions of hazardous weather, and Global Data Assimilation for creating continuous weather snapshots. These join existing tools CorrDiff and FourCastNet 3.

Reducing the Computing Burden

Traditional weather forecasting requires enormous computing resources. According to Nvidia, data assimilation alone—the process of creating starting-point snapshots for predictions—consumes roughly 50% of total supercomputing loads in conventional weather forecasting.

The new Global Data Assimilation model can complete this work in minutes on GPUs rather than hours on supercomputers. This efficiency gain could make advanced forecasting accessible to organisations that cannot afford costly supercomputer time.

Meteorologists in Israel and Taiwan are already using Earth-2 CorrDiff, whilst The Weather Company and Total Energies are evaluating Nowcasting.

Looking Forward

The democratisation of weather forecasting has significant implications. “Weather is a national security issue, and sovereignty and weather are inseparable,” Pritchard noted. For smaller nations and developing countries, access to accurate local forecasting has historically been limited by cost. If Nvidia’s claims hold up, Earth-2 could change that equation.