Google Announces New AI Weather Forecasting Models
Google is investing in weather and climate research, announcing advancements in AI-driven forecasting. The company has unveiled MetNet-3, a high-resolution model capable of generating forecasts up to 24-hours into the future. Additionally, Google has announced GraphCast, a weather model designed to predict weather patterns up to 10 days ahead.
These developments represent significant steps in leveraging artificial intelligence to improve the accuracy and lead time of weather predictions, a crucial factor for societal preparedness and decision-making in the face of changing climate conditions.
SEEDS: A Generative AI Approach to Ensemble Forecasting
Google Research has also announced the Scalable Ensemble Envelope Diffusion Sampler (SEEDS), a generative AI model published in Science Advances. This model efficiently generates weather forecast ensembles at scale, offering a cost-effective alternative to traditional physics-based forecasting models. SEEDS utilizes denoising diffusion probabilistic models, a generative AI technology that has seen recent advances in media generation.
SEEDS can generate large ensembles conditioned on as few as one or two forecasts from an operational numerical weather prediction system. The model has demonstrated a throughput of 256 ensemble members (at 2° resolution) per 3 minutes on Google Cloud TPUv3-32 instances. This efficiency allows for the generation of an order-of-magnitude more samples to fill in weather pattern distributions, providing more detailed and statistically robust forecasts.
Addressing Weather Uncertainty with SEEDS
The inherent stochastic nature of weather, famously illustrated by Ed Lorenz in his 1972 talk, “Does the Flap of a Butterfly’s Wings in Brazil Set Off a Tornado in Texas?”, necessitates probabilistic forecasts. Traditional methods generate ensembles of forecasts, but the computational cost often limits them to approximately ~10–50 ensemble members. This is insufficient for accurately assessing rare and extreme weather events, which can require much larger ensembles, such as a 10,000-member ensemble.
SEEDS addresses this challenge by producing plausible weather forecasts that match or exceed physics-based ensembles in skill metrics. It assigns more accurate likelihoods to tail-end forecast distributions, such as ±2σ and ±3σ weather events. For instance, in the context of the 2022/07/14 European extreme heat event, SEEDS ensembles provided better statistical coverage, extrapolating from two seeding forecasts to create an envelope of possible weather states with significantly improved coverage compared to traditional methods. The operational U.S. forecast system’s 31 members, for example, did not predict near-surface temperatures as warm as observed seven days prior, highlighting the limitations of smaller ensembles.
The Future of Weather Forecasting with AI
The computational cost of SEEDS is negligible compared to the hours required by supercomputers for traditional forecasts. SEEDS can generate forecasts using only 2 seeding forecasts from the operational system, which typically generates 31 members. This hybrid approach, where a few physics-based trajectories seed a diffusion model, offers an alternative to the current paradigm. Resources saved by SEEDS could be reallocated to increasing the resolution of physics-based models or issuing forecasts more frequently.
Ed Lorenz, in a 1963 paper, examined the feasibility of “very-long-range weather prediction”. Today, technologies like SEEDS, which leverages generative AI, are accelerating progress in operational numerical weather prediction. This advancement is crucial as accurate and timely forecasts will only increase in importance with ongoing climate change. Google believes SEEDS is a precursor to AI’s broader impact on weather and climate science, including applications in climate risk assessment where generating large ensembles of climate projections is vital for accurately quantifying uncertainty.
✨ Intelligent Curation Note
This article was processed by AI Universe’s Intelligent Curation system. We’ve decoded complex technical jargon and distilled dense data into this high-impact briefing.
Estimated time saved: ~7 minutes of reading.
Tools We Use for Working with AI:
