Google Deepmind researchers have built an AI weather forecasting tool that makes faster and more accurate predictions than the best system available today.
Dubbed GenCast, the new model outperformed the ENS forecast, widely regarded as the world leader, 97% of the time for predictions up to 15 days in advance. It was tested on over 1,320 weather scenarios, including tropical cyclones and heatwaves.
“Outperforming ENS marks something of an inflection point in the advance of AI for weather prediction,” Ilan Price, a research scientist at Google DeepMind, told the Guardian. “At least in the short term, these models are going to accompany and be alongside existing, traditional approaches.”
GenCast is a diffusion machine learning model, similar to those used in generative AI for tasks like image or text creation. However, it’s uniquely adapted for weather prediction, trained on four decades of data from the European Centre for Medium-Range Weather Forecasts (ECMWF) — the agency behind ENS.
During the experiments, researchers asked GenCast to generate a forecast for 2019. They then compared the results to the actual weather during that year as well as ENS’ predictions.
What day is today? It’s CYBER MONDAY!
TNW Conference is offering an exclusive 30% discount on their startup and scaleup programs this week only. This is the best deal you’ll get before prices change in January.
GenCast creates an ensemble of 50+ different predictions, each showing a possible future scenario. This data helps authorities prepare for extreme weather events like hurricanes or wind farm operators better predict power output days in advance.
The fancy name for this technique is probabilistic ensemble forecasting. It’s already the gold standard in traditional forecast systems. However, GenCast is taking things up a notch. The system can spit out predictions in far less time: 8 minutes, compared to hours for traditional models.
That’s because models like ENS run on massive supercomputers that have to crunch through millions of equations to make a prediction. By contrast, GenCast runs on a single Google Cloud TPU, a chip designed for machine learning. That’s because the AI has been trained, it’s “learnt” the data — it doesn’t have to go through it every single time it needs to make a forecast.
GenCast improves upon Deepmind’s GraphCast model unveiled last year. Other tech firms are also developing their own AI weather forecasters. Nvidia released FourCastNet in 2022, while Huawei launched its Pangu-Weather model in 2023.
So will AI replace traditional forecasting soon? Probably not. Models like GenCast still rely on data from traditional weather systems and models to train and calibrate their predictions. However, AI can certainly enhance current methods.
“The greatest value comes from a hybrid approach, combining human assessment, traditional physics-based models and AI-based weather forecasting,” Steven Ramsdale, chief forecaster at the UK’s Met Office, told the Financial Times.
