New GEM-3 weather model uses timestep-conditioned transformers

Researchers have developed GEM-3, a new probabilistic global weather model that uses timestep-conditioned transformers. This model allows for flexible timestep configuration at inference time, balancing the trade-off between resolving atmospheric dynamics and reducing error accumulation for different forecast horizons. GEM-3, with approximately 134 million parameters, builds upon its predecessor GEM-2 and incorporates architectural advancements, achieving near state-of-the-art medium-range probabilistic skill and stable extended-range forecasts. AI

IMPACT This model's flexible timestep configuration could improve the accuracy and usability of weather forecasts across various time scales.

RANK_REASON The cluster describes a new research paper detailing a novel machine learning model for weather forecasting. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• GEM-2

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →