The Sun will rise tomorrow, and you don’t have to bet your bottom dollar to be sure of it. Google’s DeepMind team released its latest weather forecasting model this week, outperforming the best traditional weather forecasting model in all tests put before it.
The generative AI model is called GenCast, and it is a diffusion model like those popular AI tools including Midjourney, DALL·E 3, and Stable Diffusion. Based on the team’s tests, GenCast is better at predicting severe weather, tropical storm movement, and wind strength for powerful land sweeps. A panel discussion on GenCast’s performance was published this week at The environment.
Where GenCast departs from other diffusion models is that it (apparently) focuses on climate, and “adapts to Earth’s spherical geometry,” as explained by several of the paper’s authors in a 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 recent weather, which the model then uses to generate probability distributions of future climates.
Conventional weather forecasting models such as ENS, the leading model from the European Center for Medium-Range Weather Forecasts, make their predictions by solving physics equations.
“One limitation of these traditional models is that the equations they solve are only estimates of atmospheric dynamics,” said Ilan Price, senior research scientist at Google DeepMind and lead author of the team’s latest results, in an email to Gizmodo.
The first GenCast seed was planted in 2022, but the model published this week includes structural changes and improved distribution settings that have made the model better trained to predict Earth’s weather, including extreme weather events, up to 15 days out.
“GenCast is not limited to learning dynamics/patterns that are well known and can be written down in an equation,” added Price. “Instead it has the potential to learn complex relationships and dynamics directly from the data, and this allows GenCast to outperform traditional models.”
Google has been working on weather forecasting for a while, and in recent years has made several significant strides in accurate forecasting using AI methods.
Last year, DeepMind scientists—some of whom co-authored the new paper—released GraphCast, a machine learning-based method that outperformed current models for medium-range weather forecasting on 90% of the targets used in the experiment. Just five months ago, a team consisting mostly of DeepMind researchers published NeuralGCM, a hybrid weather forecasting model that combines physics-based forecasting with machine learning tools. That team found that “deep learning is ultimately consistent with routine tasks [models] and can improve large-scale physical simulations that are important for understanding and predicting the Earth system.”
The solution found by GenCast is about six times better than NeuralGCM, but that was to be expected. “NeuralGCM is designed as a general-purpose atmospheric model primarily to support climate modeling, while GenCast’s high resolution is often expected in medium-range forecasting models of performance, which is GenCast’s target use case,” added Price. “That’s why we’ve also emphasized a wide range of tests that are critical to the use of medium-range forecasts, such as extreme weather forecasting.”
In recent work, the team trained GenCast on historical weather data for 2018, then tested the model’s ability to predict weather patterns for 2019. GenCast outperformed ENS with 97.2% of targets using different weather variables, with different lead times before a weather event. ; for lead times greater than 36 hours, GenCast was more accurate than ENS on 99.8% of targets.
The team also tested GenCast’s ability to predict the track of a tropical storm—specifically Typhoon Hagibis, the costliest typhoon of 2019, which hit Japan that October. GenCast’s forecasts were highly imprecise with seven days of lead time, but more accurate at shorter lead times. As extreme weather creates heavy, torrential rain, and hurricanes break records for how fast they grow and how early in the season they form, accurate forecasting of hurricane paths will be critical to reducing their financial and human costs.
But it doesn’t end there. In a proof-of-principle test described in the study, the DeepMind team found that GenCast was more accurate than ENS in predicting the total wind energy produced by groups of more than 5,000 wind farms in the Global Power Plant Database. GenCast’s forecasts were about 20% better than ENS’ at lead times of two days or less, and maintained statistically significant improvements up to a week. In other words, the model doesn’t just have value in disaster mitigation—it can inform where and how we deploy energy infrastructure.
What does all this mean to you, you who just don’t know the weather? However, the DeepMind team has made the GenCast code open source and the models available for non-commercial use, so you can build the tool around if you want to. The team is also working to release an archive of historical and current weather forecasts.
“This will enable the broader research and climate community to collaborate, evaluate, develop, and build on our work, accelerating progress in this field,” Price said. “We have advanced versions of GenCast to be able to take operational input, so the model can begin to be put into operation.”
There’s currently no timeline for when GenCast and the other models will go live, though DeepMind’s blog noted that the models are “starting to power user experience in Google Search and Maps.”
Whether you’re here for weather or AI applications, there’s a lot to love about GenCast and DeepMind’s extensive suite of predictive models. The accuracy of such tools will be key in predicting severe weather events with enough time to protect those at risk, be it floods in Appalachia or hurricanes in Florida.
Source link
