Green Shipping

Google uses AI to fast-track predicting weather catastrophes

Google has developed a new generative AI model that it says can predict weather patterns more efficiently than conventional physics-based weather forecasting models, up to 14 days in advance.

This means that difficult to identify weather catastrophes such as hurricanes could soon be detected with greater assurance. Enabling emergency service planners and the public to enable to prepare better and respond faster to weather-related events. And for the supply chain executives, advanced planning for weather related events can not only save lives and property but enable better decision making for freight movements.

Google’s findings were published on March 29th in the journal Science Advances.

Rob Carver

In an interview with AJOT, Rob Carver, Research Scientist and co-author of the Google study said:” And I will have a numerical model of the atmospheric conditions and I will then… get the current conditions and then I'll run my numerical model up forward in time and I will get a forecast. And that was the great revolution of meteorology in the sixties and seventies making NWP possible.”


Numerical weather prediction (NWP) is a method of weather forecasting that employs a set of equations that describe the flow of fluids.

Carver said the NWP based system utilized a series of ensembles-multiple weather forecasts- that could only be run on a supercomputer which was costly and time consuming and ran a range of 30-50 ensembles.

Carver says: “Generative AI techniques work really fast, really quickly and using not so much hardware. So, that really helps us … when we go to like a thousand members (ensembles) that lets us see what's lurking as a very rare forecast. … something that can only happen less than 1% of the time. If you have 30-50 members … you're not going to have a good chance of seeing it.”

The Scalable Ensemble Envelope Diffusion Sampler (SEEDS) that Google developed which can, “right now (generate) up to a 14 day weather forecast… We haven't explored SEEDS in the terms… of basically sub seasonal, seasonal forecast. We haven't got there yet.”

Carver added: “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 ("a picture of a cat driving a car") and generate lots of images that fit the prompt.

Similarly, the "prompt" for SEEDS is a single prediction of the future weather. SEEDS 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."

Resham Parikh, the Public Relations Manager, Emerging Technology and Exploratory Research at Google, told AJOT that Google is working on a number of different research projects in improving weather predictions, “SEEDS is one really, really cool project across Google research. There's lots of… people working on weather research… Rob was able to discuss this particular project… there are teams working on … seasonal forecasting… There are… teams actually working on that. That's a completely different technology. And we… haven't released any of those findings yet… we haven't published them yet…”

Parikh explained: “Our weather forecasting research extends beyond the SEEDS project. Right now, we're exploring a lot of different methodologies for improving shorter and longer term weather forecasting, including sub-seasonal forecasting, both of which would be helpful for navigating our changing climate.”

Google Deep Mind AI

Google Deep Mind has developed AI that can generate forecasts far faster than the industry gold-standard HRES, which is produced by the European Centre for Medium-Range Weather Forecasts, Parikh said adding: “DeepMind was purchased by Google years ago and then integrated into Google … And we've been integrating these teams more and more over time. And actually, a lot of Google researchers now work in DeepMind and a lot of Google researchers on the Google research side collaborate very, very closely with DeepMind.”

The Google Deep Mind GraphCast model has generated a more accurate 10 day forecast than the system used by the European Centre for Medium Range weather forecasting and makes predictions in minutes rather than hours and can run on a desk top computer. GraphCast is based on 38 years’ worth of global weather readings.

Stas Margaronis
Stas Margaronis


Contact Author

© Copyright 1999–2024 American Journal of Transportation. All Rights Reserved