Rob Carver, research scientist at Google, explores the company’s Scalable Ensemble Envelope Diffusion Sampler (SEEDS) model and how powerful generative AI could revolutionize the speed and statistical accuracy of weather forecasts.
What are the key features of SEEDS? What sets it apart from other AI weather forecasting technology?
SEEDS is part of Google’s larger effort to invest in weather and climate research to ensure that we can reliably predict and respond to our changing climate and weather. The model uses generative AI to increase the number of ‘ensemble’ forecasts – a group of forecasts that together show the possible range of weather events that could occur in the next three to 10 days. By increasing the range of possible outcomes in a forecast, SEEDS technology would enable a forecaster to better assess the likelihood of future weather events, including rare or severe weather events (for example, a prolonged heat wave), with a higher degree of accuracy. Thus, SEEDs could further accelerate and improve the statistical accuracy of weather forecasts.
How exactly does the technology improve the accuracy of weather forecasts?
Currently, the best practice to create weather forecasts with probabilities is to run a computer weather model on a supercomputer 30-50 times using slightly different initial conditions. We call this ensemble weather prediction. If it rains in only 50% of the different forecasts, then the probability of rain is 50%. It takes a lot of computer power to produce all these forecasts. In comparison, generative AI is able to generate hundreds or thousands of forecasts in record time. As a result, it can produce more accurate forecasts than the best traditional methods – which is particularly important for the challenging yet important forecasting of extreme weather.
Can you tell me more about the solution’s use of diffusion models?
SEEDS was the first time we used a diffusion model (generative AI) to transform a single, deterministic forecast into an ensemble that characterizes the range of possible outcomes. If you input just two possible weather forecasts into SEEDS, it will return a far larger set of weather forecasts that could occur at that time. That used to take hours and maybe days to do, and we’ve cut that time down by 90%. Additionally, because we base the chance of a weather event (like rain or snow) on how many of our forecasts show that it’ll happen, this huge number of probable forecasts means far greater accuracy in the chance of that event happening.
What are the advantages of using AI in weather forecasting?
In a broad sense, generative AI models are AI models that predict the most likely response to a given input. They can let us, for example, give a single text prompt, e.g. ‘a picture of a cat driving a car’ and generate lots of images that fit the prompt. Similarly, the prompt for a model like SEEDS is a single prediction of the future weather. Generative AI can then generate many more possible forecasts for that same time period, giving us higher statistical confidence in the kind of weather we can expect in the real world.
How will this technology affect the future of meteorology?
This technology can increase the speed and accuracy of weather forecasting. SEEDS uses only 10% of the computing power of the traditional supercomputer approach to get the same number of forecasts.
For more of the top insights into the future of meteorological AI, read Meteorological Technology International’s exclusive feature How will generative artificial intelligence transform the meteorological sector? here.
