DeepMind AI predicts weather more accurately than existing forecasts

Google DeepMind says its latest weather forecasting AI can make predictions faster and more accurately than physics-based simulations.

GenCast is the latest in DeepMind’s ongoing research project to use artificial intelligence to improve weather forecasting. The model was trained on four decades of historical data from the European Center for Medium-Range Weather Forecasts (ECMWF). ERA5 archivewhich includes regular measurements of temperature, wind speed and pressure at various altitudes around the world.

Data up to 2018 was used to train the model and then data from 2019 was used to test the predictions against known weather. The company found that it exceeded ECMWF’s industry standard ENS forecast 97.4 percent of the time overall, and 99.8 percent of the time looking ahead by more than 36 hours.

Last year, DeepMind worked with ECMWF to create an AI that outperformed the “gold standard” high-resolution HRES 10-day forecast. more than 90 percent of the time. Previously, he developed “nowcasting” models He predicted a chance of rain 5 minutes to 90 minutes in a given area of ​​1 square kilometer using 5 minutes of radar data. And Google is also working on ways to work Using AI to replace small parts of deterministic models while maintaining accuracy to speed up computation.

Existing weather forecasts are based on physics simulations run on powerful supercomputers that deterministically model and extrapolate weather patterns as accurately as possible. Forecasters typically run dozens of simulations in groups called sets with slightly different inputs to better capture possible outcomes. These increasingly complex and numerous simulations are very computationally intensive and require increasingly powerful and energy-hungry machines to run.

AI can provide a cheaper solution. For example, GenCast generates predictions with a set of 50 possible futures, each just 8 minutes long on a custom-built, AI-based Google Cloud TPU v5 chip.

GenCast operates at a cell resolution of about 28 square kilometers at the equator. Since the data used in this study were collected, ECMWF’s ENS has been upgraded to 9 km resolution.

Ilan Price DeepMind argues that AI may not need to follow and can provide a way forward without gathering finer data and performing more intensive calculations. “When you have a traditional physics-based model, that’s a prerequisite for more accurate predictions because it’s a prerequisite for solving the physical equations more accurately,” says Price. “(With machine learning) (going to higher resolution is not necessarily a requirement to get more accurate simulations or predictions from your model.”

David Schultz At the University of Manchester in the UK, he says AI models can make weather forecasts more effective, but they often overshoot, and it’s important to remember that they rely on training data from traditional physics-based models.

“Will (GenCast) revolutionize numerical weather prediction? No, because you still have to run numerical weather prediction models to train the models,” Schultz says. you wouldn’t have these AI tools. That’s like saying ‘I can beat Garry Kasparov at chess, but after analyzing every move he’s ever played’.”

Sergey Frolov The US National Oceanic and Atmospheric Administration (NOAA) believes that AI will need higher resolution training data to progress. “Essentially what we’re seeing is that all of these approaches are stalling (progress) due to the fidelity of the training data,” he says. “And the training data comes from operational centers like ECMWF and NOAA. To advance this field, we need to generate more training data with higher-fidelity physics-based models.”

But for now, GenCast offers a way to make predictions at a lower computational cost and faster. Kieran Hunt At the University of Reading, UK, he says that just as a collection of physics-based predictions can produce better results than a single prediction, he believes ensembles will increase the accuracy of AI predictions.

He points to Hunt Record temperature of 40°C (104°F). 2022 was seen in the UK as an example. A week or two earlier, there were lone members of the groups announcing this, but they were considered abnormal. Then, as we got closer to the heat ball, more and more predictions came into line, and they were able to warn us that something unusual was coming.

“It gives you a bit of cover for a member showing something really extreme; it could happen, but probably not,” says Hunt. “I wouldn’t necessarily see it as a step change. The tools we’ve been using in weather forecasting for a long time are being combined with a new AI approach in a way that will work to improve the quality of AI forecasts. It will certainly be better than the first wave of AI weather forecasts.”