Tomorrow the sun will rise and you will no longer have to bet your last dollar to be sure. Google’s DeepMind team released its latest weather forecast model this week, which outperforms a leading traditional weather forecast model in the vast majority of tests submitted.
The generative AI model is called GenCast and is a diffusion model like those underlying popular AI tools including Midjourney, DALL·E 3 and Stable Diffusion. Based on the team’s testing, GenCast is better at predictions extreme weather conditionsthe movement of tropical storms and the force of wind gusts over Earth’s powerful land expanses. The team’s discussion on GenCast’s performance was published this week in Nature.
Where GenCast differs from other diffusion models is that it is (obviously) climate-focused and “adapted to the Earth’s spherical geometry,” as a couple of the paper’s co-authors described in a paper DeepMind blog post.
Instead of a written prompt like “paint a picture of a dachshund in the style of Salvador Dalí,” GenCast’s input is the most recent weather state, which the model then uses to generate a probability distribution of future weather scenarios.
Traditional weather forecasting models such as ENSthe leading model of the European Center for Medium-Range Weather Forecasts, makes its forecasts by solving physical equations.
“A limitation of these traditional models is that the equations they solve are only approximations of atmospheric dynamics,” said Ilan Price, a senior researcher at Google DeepMind and lead author of the team’s latest findings, in an email to Gizmodo.
The first seeds of GenCast were planted in 2022, but the model released this week includes architectural changes and an improved scattering configuration that made the model better suited to predicting weather on Earth, including extreme weather events, out to 15 days .
“GenCast is not limited to learning dynamics/models that are known exactly and can be written into an equation,” Price added. “Instead it has the opportunity to learn more complex relationships and dynamics directly from the data, and this allows GenCast to outperform traditional models.”
Google has been involved in weather forecasting for a long time, and in recent years it has made a couple of substantial steps towards more accurate predictions using artificial intelligence methods.
Last year, DeepMind scientists, some of them co-authors of the new paper,released GraphCasta machine learning-based method that outperformed current medium-range weather forecasting models on 90% of targets used in tests. Just five months ago, a team made up mostly of DeepMind researchers published NeuralGCM, a hybrid weather forecasting model that combines a traditional physics-based weather predictor with machine learning components. That team found that “end-to-end deep learning is compatible with tasks performed by conventional (models) and can improve large-scale physics simulations that are essential for understanding and predicting the Earth system.”
The resolution achieved by GenCast is about six times that of NeuralGCM, but that was expected. “NeuralGCM is designed as a general-purpose atmospheric model primarily to support climate modeling, while GenCast’s higher resolution is often intended for medium-range operational forecast models, which is GenCast’s specific target use case,” he added Price. “This is also why we have emphasized a broad range of assessments that are crucial use cases for medium-term operational forecasting, such as predicting extreme weather.”
In recent work, the team trained GenCast on historical weather data through 2018, then tested the model’s ability to predict weather patterns in 2019. GenCast outperformed ENS on 97.2% of targets using different weather variables, with variable delivery times before the weather event. ; with lead times exceeding 36 hours, GenCast was more accurate than ENS on 99.8% of targets.
The team also tested GenCast’s ability to predict the path of a tropical cyclone, specifically Typhoon Hagibis, the costliest tropical cyclone of 2019, which hit Japan in October. GenCast’s forecasts were highly uncertain with a seven-day lead time, but became more accurate with shorter lead times. Like extreme weather conditions generates wetter and more abundant rainfallAND hurricanes break records However quickly they intensify and how early in the season they form, accurately predicting storm paths will be crucial to mitigating their fiscal and human costs.
But that’s not all. In a proof-of-principle experiment described in the research, the DeepMind team found that GenCast was more accurate than ENS at predicting the total wind power generated by clusters of more than 5,000 wind farms in the Global Power Plant Database. GenCast’s forecasts were about 20% better than ENS’s with lead times of two days or less and maintained statistically significant improvements for up to a week. In other words, the model not only has value in mitigating disaster, but could also provide guidance on where and how to deploy energy infrastructure.
What does all this mean to you, oh distracted climate admirer? Well, the DeepMind team has made the GenCast code open source and the models available for non-commercial use, so you can move along if you’re curious. The team is also working on publishing an archive of historical and current weather forecasts.
“This will enable the broader meteorological and research community to engage with, test, execute and develop our work, accelerating further progress in the field,” Price said. “We have tuned versions of GenCast to be able to take operational input and then the model could start to be incorporated into the operational context.”
There’s no timeline yet on when GenCast and other models will go live, though DeepMind’s blog noted that the models are “starting to power the user experience on Google Search and Maps.”
Whether you’re here for the weather or AI applications, there’s a lot to like about GenCast and the broader suite of DeepMind forecasting models. The accuracy of such instruments will be critical predict extreme weather events with sufficient lead time to protect those in danger, be it floods in Appalachia OR tornado in Florida.
